Next Article in Journal
Study on a Fault Mitigation Scheme for Rub-Impact of an Aero-Engine Based on NiTi Wires
Next Article in Special Issue
An Ensemble Learning Model for COVID-19 Detection from Blood Test Samples
Previous Article in Journal
Novel Cooperative Automatic Modulation Classification Using Vectorized Soft Decision Fusion for Wireless Sensor Networks
Previous Article in Special Issue
Examination of Potential of Thermopile-Based Contactless Respiratory Gating
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Systematic Review

Internet of Things Technologies and Machine Learning Methods for Parkinson’s Disease Diagnosis, Monitoring and Management: A Systematic Review

by
Konstantina-Maria Giannakopoulou
1,2,
Ioanna Roussaki
1,2,* and
Konstantinos Demestichas
1,2
1
School of Electrical and Computer Engineering, National Technical University of Athens, 15773 Athens, Greece
2
Institute of Communication and Computer Systems, 10682 Athens, Greece
*
Author to whom correspondence should be addressed.
Sensors 2022, 22(5), 1799; https://doi.org/10.3390/s22051799
Submission received: 28 December 2021 / Revised: 17 February 2022 / Accepted: 21 February 2022 / Published: 24 February 2022
(This article belongs to the Special Issue AI and IoT Enabled Solutions for Healthcare)

Abstract

:
Parkinson’s disease is a chronic neurodegenerative disease that affects a large portion of the population, especially the elderly. It manifests with motor, cognitive and other types of symptoms, decreasing significantly the patients’ quality of life. The recent advances in the Internet of Things and Artificial Intelligence fields, including the subdomains of machine learning and deep learning, can support Parkinson’s disease patients, their caregivers and clinicians at every stage of the disease, maximizing the treatment effectiveness and minimizing the respective healthcare costs at the same time. In this review, the considered studies propose machine learning models, trained on data acquired via smart devices, wearable or non-wearable sensors and other Internet of Things technologies, to provide predictions or estimations regarding Parkinson’s disease aspects. Seven hundred and seventy studies have been retrieved from three dominant academic literature databases. Finally, one hundred and twelve of them have been selected in a systematic way and have been considered in the state-of-the-art systematic review presented in this paper. These studies propose various methods, applied on various sensory data to address different Parkinson’s disease-related problems. The most widely deployed sensors, the most commonly addressed problems and the best performing algorithms are highlighted. Finally, some challenges are summarized along with some future considerations and opportunities that arise.

1. Introduction

1.1. Parkinson’s Disease

Parkinson’s disease (PD) is the second most common neurodegenerative disease [1] and is responsible for a considerable amount of disability-adjusted life years and deaths globally [2], leading to an extremely high demand for respective health resources. In recent decades, we have witnessed a dramatic rise in the amount of people suffering from it worldwide, which is correlated with the ageing of the global population, as well as with other potential factors, such as air pollution and smoking [1,2]. PD is related to the prominent loss of dopaminergic neurons, the development of Lewy bodies and neuroinflammation [3]. It is attributed to a complex combination of genetic and environmental factors [1,3]. However, the exact cause of PD remains unknown.
PD is manifested with mainly motor symptoms, such as resting tremor, muscular rigidity, bradykinesia, or even akinesia, postural and gait impairment, but is also related to non-motor characteristics, such as sleep dysfunction, autonomic dysfunction, including orthostatic and postprandial hypotension, fatigue, pain, hyposmia, bladder and gastrointestinal disturbances, cognitive deficits, depression, mood disorders, dementia and hallucinations [1,3,4,5]. There are also indications that phonation and speech disorders are common early signs among PD patients [6,7]. These clinical symptoms may be manifested differently from patient to patient, as some of them may be absent, while others may be quite severe. Similarly, the progression of the disease varies across PD patients. This phenotypic variability has led to the definition of several PD subtypes. One of the most widely accepted classifications corresponds to tremor-dominant PD and non-tremor-dominant PD [8]. Additionally, the manifestations of the disease demonstrate fluctuations in the same patient, in concordance with the ON and OFF states, which are also related to the impact of the levodopa treatment [9]. This is one widely proposed treatment for the management and alleviation of PD symptoms, without however curing the disease.
Νon-motor symptoms may be present earlier than the motor ones. However, clinical diagnosis is performed based on the latter, including the presence of bradykinesia, tremor or rigidity [5], as the former are often not easily detected and are usually attributed to other health factors and to senescence. Furthermore, there is no specific test for PD diagnosis. Conventionally, the patient’s medical history, signs, symptoms and medical examinations are taken into consideration by the doctor, including some simple physical examinations and exercises and some mental tasks, to finally diagnose PD. Brain images, such as magnetic resonance images (MRIs), computed tomography (CT) scans, positron emission tomography (PET) scans and radiographs, or other lab tests may also be exploited to exclude other medical conditions [10].
The assessment of the PD symptoms is usually accompanied by some scores in common PD-related motor or non-motor rating scales, such as the Movement Disorders Society Sponsored Revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS), the Hoehn and Yahr (H&Y) Staging Scale, the Schwab and England Activities of Daily Living (ADL) Scale, as well as some self-administered questionnaires, such as the Self-Assessment Parkinson’s Disease Disabilities Scale (SPDDS) [11]. The aforementioned clinical tests and the respective rating scales are also used to estimate the progression of the disease during regular follow-up appointments with the neurologist.
The lack of well-established PD biomarkers and the similarity of manifestations among different neurological disorders complicate and delay successful diagnosis. Furthermore, the patient may suffer from several fluctuations between two consecutive doctor’s appointments, which makes diagnosis and monitoring based on clinical examinations more difficult, especially when patients fail to precisely report their symptoms. Consequently, the patient’s clinical picture is subsampled, and useful information may vanish. Additionally, both inter-subject and intra-subject variability concerning PD clinical symptoms highlight the urgent need for personalized treatment suggestions and complicate the relevant medical decision-making processes. Provided that the diagnosis is performed early enough and the clinicians have adequate available information when monitoring a PD patient, the impact of the tailor-made treatments is expected to be greater.

1.2. Atrificial Intelligence and Internet of Things for Parkinson’s Disease Diagnosis and Management

Artificial intelligence (AI) is the field of computer science which deals with intelligence attributed to machines. According to Russel and Norvig [12], intelligent agents perceive their environment and make decisions in order to maximize their chance to achieve their goals. Conventionally, AI is rule-based and makes use of human experts’ knowledge. Machine learning (ML) is a more flexible and data-driven subfield of AI, according to which a computational machine may improve its performance in a specific task through acquiring experience [13], imitating the way that humans learn. Furthermore, deep learning (DL) is a subfield of ML which is less dependent on human intervention and is based on artificial neural networks (ANNs) that consist of multiple neurons stacked in many different layers and connected with each other, inspired by the structure of the human brain [14]. ML and DL algorithms and techniques can be applied to data collected from sensors, embedded on wearable devices or other everyday objects. Nowadays, these data are abundant and easy to acquire thanks to the Internet of things (IoT) paradigm, which enables various physical objects with embedded sensors and unique identifiers to communicate with each other and transfer data over a network [15].
AΙ and ΙοΤ technologies can contribute to both the early diagnosis of PD and PD patients’ monitoring, aiming to enable improved personalized treatments and assess the proposed ones [16], complementing the already established conventional methods. Moreover, by training ML algorithms on sensory data collected from PD patients, evaluation of suitability of current prescribed medication can be facilitated, as well as the optimization of surgical treatments, the prediction of the course of the disease and the prevention of undesirable consequences even in real time [16]. These interventions pave the way for precision medicine in PD, among other chronic diseases, and support transition from clinic-centric to patient-centric healthcare approaches. Finally, the opportunities that arise from remote monitoring with the help of wireless sensors, other wearables and telemedicine tools may cover the respective urgent need in the COVID-19 pandemic era and help PD patients to continue their care and treatment without putting themselves at risk [17], overcoming at the same time the spatial and temporal constraints in place.
More specifically, ML and DL techniques have been applied to brain images and other conventional clinical examinations to implement computer-aided diagnosis (CAD) tools [18]. These approaches are of utmost importance, as they may extract features that are not easily detectable by experts. However, the respective published studies will not be discussed in the current literature review, which is dedicated to IoT interventions combined with AI methods. Sensors are a valuable source of information that can feed ML and DL algorithms with rich data, collected during or between doctor’s appointments, in laboratories, hospitals or in free-living environments, at a reasonable cost. All these wearable or non-wearable sensors are networked and can transmit data from mobiles, tablets or other smart devices to remote databases by exploiting the IoT infrastructure available [19].
The vast majority of the deployed sensors include inertial sensors or inertial measurement units (IMUs), such as accelerometers, gyroscopes and magnetometers [20], as well as pressure sensors and ground reaction force (GRF) plates measuring features correspondent to PD motor symptoms [21,22]. Moreover, according to Monje et al. [21], there are several other technologies deployed at various technology readiness levels. For example, firstly, image and depth sensors, as well as sound or audio sensors have been also used to capture motor and non-motor PD manifestations. Other deployed sensors that capture various bio-signals include electroencephalography (EEG), electrocardiography (ECG) and electromyography (EMG) sensors, pain measurement devices, portable sleep measurement devices and polysomnography (PSG) sensors, eye tracking systems, heart rate and temperature sensors, among others. Finally, these data sources and the respective ML and DL models trained are sometimes integrated in mobile applications [23,24], conversational agents or chatbots [25] and serious games [26] and can be combined with controllers [27] to improve the patient’s quality of life.

1.3. Aim of the Current Systematic Review

The purpose of the review presented in this paper is to investigate the use of ML and DL models trained on sensory data, in support of PD patients, their caregivers and the clinicians in every phase of the disease. Therefore, it summarizes the main findings of various research articles that propose PD predictive and estimation models based on novel IoT technologies and optimal sensor deployments. The aim is to shed light on more novel approaches and present new possibilities that have not been discussed to a sufficient extent in the previously published reviews, providing neurologists with useful insights that could potentially revolutionize PD diagnosis and care. For example, there are some review studies in the current literature that summarize different types of sensors and commercially available devices, which are deployed to monitor PD patients with respect to various aspects and motor symptoms, without however emphasizing predictive ML models [20,21,22,28,29,30,31]. There are also others that focus only on one problem, such as rigidity quantification [32], impaired gait analysis [33,34,35], freezing of gait episodes and potential falls detection [36,37,38], as well as computer-aided PD diagnosis [39,40]. The current systematic review intends to fill all these gaps, by presenting a more holistic approach, examining both the sensorial and the algorithmic side of strictly IoT- and ML-based approaches.
Nevertheless, there are some other preceding reviews that address this topic, following a similar approach [41,42,43,44,45,46,47]. However, the approach adopted herein is more comprehensive than most of these systematic reviews, as 112 studies are analyzed and considered, compared to 10–48 studies considered in previous approaches. Only Rovini et al. [41] examines 136 papers, but this systematic review was published in 2017 so naturally does not consider all approaches published in the last four years. Moreover, the review presented in this paper sheds light on additional questions, such as which the most commonly addressed problems regarding PD are, which the most deployed sensors are, and which the most used and best-performing ML algorithms are. Finally, the review of this paper highlights not only the added value of these interventions, but also the challenges and the potential pitfalls that they involve, as well as some gaps and open issues, proposing some future directions for improvement.

2. Methods

The current literature review is conducted in a systematic way that is largely consistent with the latest PRISMA guidelines [48]. As already stated, the purpose of this review is to summarize the advances in AI approaches based on data collected through sensors and IoT technologies that support PD patients and neurologists in every phase of the disease. On this basis, the authors made the following decisions regarding the identification of search strategy and data sources, the criteria for the final selection of the studies and the selection and data extraction processes.

2.1. Literature Search Strategy

Firstly, the most suitable databases and registers of the respective literature were selected. The authors opted for three popular academic citation databases, which guarantee the high quality and impact of the considered papers. The first one is PubMed, one of the most popular databases of biomedical literature. The second one is IEEE Xplore, which is a well-known database of technical literature and the third one is Scopus, which indexes academic articles from a wide range of research fields.
Then, the most suitable keywords were selected, to assure the inclusion of studies that would answer the posed research questions. Indicative such terms are “Parkinson’s disease”, “artificial intelligence”, “machine learning”, “deep learning”, “sensor”, “internet of tings” and some specific types of sensors, such as “accelerometer”, “gyroscope”, etc. All these terms were combined with logical expressions. Additionally, some other keywords were added, such as “remote”, “portable”, etc., to ensure that clinical equipment which cannot be deployed for remote diagnosis or monitoring would be excluded. The selection of the final technical keywords was based on methods and devices discussed in previous review articles, after several adjustments to ensure the inclusion of all the technologies of our interest. Finally, the considered studies should be articles published in journals or in conference proceedings in the last decade (from January of 2012 till August of 2021) and written in English. During the last decade, the exploitation of AI algorithms and IoT tools in healthcare was hyped. This period is also long enough to support a comprehensive review approach with concrete conclusions about the evolution of the respective methods proposed regarding PD over the years.
After this fine-tuning, the final search query that was submitted to the Scopus database is the following: TITLE-ABS-KEY (“Parkinson’s disease” AND (“machine learning” OR “deep learning” OR “artificial intelligence”) AND (sensor OR device OR “internet of things” OR “accelerometer” OR “gyroscope” OR imu OR “inertial measurement unit” OR “force platform” OR “force plate” OR video OR camera OR smartphone OR smartwatch OR ((electromyography OR electroencephalography OR polysomnography OR electrocardiography) AND (portable OR wearable OR home OR remote)))) AND PUBYEAR > 2011 AND LANGUAGE (English) AND DOCTYPE (ar OR cp). Equivalent, semantically and syntactically, queries were submitted to the other two databases as well.

2.2. Eligibility Criteria

Some inclusion and exclusion criteria were defined to ensure that the research questions are answered. These eligibility criteria can be summarized as follows:
  • Firstly, the considered studies should be strictly articles or conference papers, published between January 2012 and August 2021, as already stated.
  • The considered studies should be about any aspect or phase of PD. The hypotheses should be tested on adult human subjects, under strict ethical guidelines. Some of the subjects should be PD patients but healthy subjects and other neurological disorders patients may also be included for, e.g., differential diagnosis. However, articles that present technologies which can be leveraged for PD diagnosis or management but are tested only on mixed or healthy populations are excluded. Furthermore, there should be a concise description of the datasets used and the respective signals should be real and not simulated.
  • Moreover, the signals used should derive directly from wearable or non-wearable sensors, smart devices, ambient or other technologies that are related to IoT. Studies that exploit data from other sources, such as conventional medical equipment, interviews or medical reports that cannot be adopted for remote monitoring or diagnosis are excluded.
  • At the same time, at least one specific AI algorithm should be proposed to solve the corresponding PD-related problem. The respective ML models should be trained over datasets that are in accordance with the previous guidelines and their performance should be measured with specific evaluation metrics. Studies that present only a statistical analysis and not any AI methods are excluded. Finally, the provided conclusion should be consistent with the initial research goal.
In addition to the previous guidelines and restrictions, since multiple studies may repeatedly address the same problems, proposing the same ML methods based on the same input data, some studies were clustered in groups and only some discrete representatives of each group were considered to avoid redundancy. More specifically:
  • If there are both a conference paper and an extended research article for the same study, the latter is preferred.
  • In the case that the same or similar PD-related problems are addressed by one research group multiple times, utilizing similar methods or same datasets (e.g., proposing one or two more algorithms each time), only one is preserved and the rest are excluded. The most recent or comprehensive and detailed one was preferred.
  • Finally, even if the research group is different when there are two or more almost identical approaches, only one is presented and the rest are excluded. By the term identical approaches, the paper refers to the exact same PD-related task, same ML/DL algorithms and the exact same open dataset or similar train and test population.

2.3. Selection and Data Collection Processes

One investigator (K.M.G.) submitted the search queries to the three selected databases. The studies were retrieved automatically with the help of Zotero (https://www.zotero.org/, accessed on 17 February 2022). The same investigator proceeded to the removal of duplicate records, again automatically by the Zotero tool. Then, the investigator (K.M.G.) screened the title and the abstract of the remaining studies to remove further articles, according to the inclusion and exclusion criteria set, without clustering similar approaches at this initial stage. Moreover, the investigator (K.M.G.) identified and excluded a few studies, for which the full text could not be retrieved. At this point, two investigators (K.M.G. and I.R.) screened independently the full text of the remaining studies. Each investigator concluded independently to a set of studies to be considered in the current literature review, according to the inclusion and exclusion criteria set. Then, the two investigators (K.M.G. and I.R.) discussed their conclusions. The agreement rate was high and after solving all the divergences, they reached a consensus about a final set of studies to be considered. The two investigators also agreed upon the information that should be collected from each report and the first evaluator (K.M.G.) finally performed the data collection process from each study individually, according to what has been discussed.

3. Results

3.1. Search Results

The steps of the PRISMA-based methodology discussed in Section 2, along with the results obtained at the end of each phase are illustrated in the flow diagram depicted in Figure 1. Initially, 770 articles were retrieved, after submitting the queries to the three databases (142 articles from PubMed, 192 articles from IEEE Xplore and 436 articles from Scopus). Next, 269 duplicates were removed, resulting in 501 studies to consider. Then, other 199 reports were excluded, after screening the title and the abstract of the remaining studies, according to the inclusion and exclusion criteria. Moreover, the full text of three studies could not be retrieved via the web, resulting in 299 reports to be assessed for eligibility. Eventually, 112 studies were selected to be considered in the literature review presented in this paper, after screening the full text of each study according to the inclusion and exclusion criteria set.
The distribution of the initially retrieved studies per academic citation database is depicted in the pie chart of Figure 2. The vast majority of the studies which were finally selected to be analyzed in the current review, are listed in more than one of the considered databases. Thus, the respective pie chart for the finally considered studies is not presented, as it provides no significant additional information.
Moreover, it is substantiated that the results obtained are up to date. More specifically, the distribution of the publication years of the considered papers is depicted in the column chart of Figure 2. As one can easily observe, less than 20% of the considered studies are published before 2018, while the rest are published in the last 4 years. It is worth mentioning that more than 30% of the studies considered are published in 2020. Additionally, the respective rate for 2021 is also expected to be higher, as this literature research was conducted in the August of 2021 and the remaining 4 months of the year are not considered. To conclude, the results reached by the present study may serve as a complement of the preceding comprehensive systematic review conducted by Rovini et al. [41] in 2017, providing additional state-of-the-art information and insights from the last 4 years.

3.2. Results of Individual Studies and Sensor-Wise Synthesis

The selected articles are divided into six major categories according to the types of sensors and IoT technologies deployed. These technologies span across inertial sensors, pressure or force sensors, image and depth sensors, audio or voice sensors, other wearable biosensors or other types of sensors in general that are not considered in the previous categories and the combination of different types of the aforementioned sensors. For each category, the studies that address the same or similar PD-related problems are grouped and the main points of the respective approaches are briefly presented. Furthermore, the performance of the various AI algorithms deployed are compared with each other and an attempt is made to identify the methods that outperform the rest with respect to a specific PD-related task based on one specific type of sensory input data each time.
Finally, the key parts of each study discussed, including the input data, the population of the dataset, the problem addressed, the deployed algorithms and their performance on the basis of specific evaluation metrics, are also summarized in six comprehensive tables in Appendix A of this paper, which are again divided according to the six different categories of sensors introduced in the current section.

3.2.1. Inertial Sensors

The vast majority of the considered articles are based on wearable inertial sensors or IMUs, such as accelerometers, gyroscopes, magnetometers and other angular velocity sensors, which measure several kinematic parameters. The studies considered in this category provide evidence that these measurements may shed light on various aspects of PD, mostly related to motor symptoms, supporting PD diagnosis, differential diagnosis between various neurological disorders, identification of PD different subtypes, detection of PD symptoms and estimation or prediction of their severity, detection or prediction of FoG episodes, detection of symptoms fluctuations, segmentation of gait, rehabilitation assessment, as well as estimation and prediction of patients’ response to the prescribed medication.
There are 52 studies in this category that make use of inertial sensors solely, out of which 21 studies correspond to lower-limb analysis, 21 studies refer to upper-limb analysis and 10 studies correspond to whole-body analysis. Firstly, acceleration and angular velocity signals collected with IMUs attached to the waist or to the feet at several positions can provide useful insights regarding impaired gait ability or postural instability due to PD. Similarly, acceleration and angular velocity signals from hand-mounted inertial sensors can shed light on PD-related tremor parameters, bradykinesia and dyskinesia symptoms. IMUs can also be attached to other body parts, such as the sternum to extract more motor-related features, which are possibly correlated with the course of PD. Finally, these sensors are sometimes embedded in smartphones, smartwatches or other everyday objects. More details about the respective review findings for this type of sensorial data exploited are provided in Table A1 in Appendix A of this paper.
The first problem that will be discussed is PD diagnosis or equivalently the classification between PD patients and healthy controls, which is frequently addressed based on inertial signals. To this end, gait parameters have been extracted either manually via feature engineering techniques [49,50,51,52,53] or automatically via deep convolutional neural networks (CNNs) [54], to feed several classification algorithms. The deployed algorithms include support vector machines (SVMs), decision trees (DTs), random (RFs), bagged, boosted and fine trees, k-nearest neighbors (kNN), logistic regression (LR), linear discriminant analysis (LDA) and naïve Bayes (NB) classifiers, as well as multi-layered perceptrons (MLPs) or other neural networks (NNs). It appears that when IMUs are attached to the feet, researchers achieve slightly higher performance than when sensors are attached to the waist or to the lower spine. For example, in [49,51,53], PD diagnosis is performed with 90–99.33% accuracy by a DT, an MLP and an RF classifier, respectively, trained on feet signals, while 84.5–85.51% accuracy is obtained when kNN algorithms are trained over waist signals [50,52]. Moreover, the performance does not seem to improve significantly when deep CNNs are deployed in [54], leading to 0.87 area under the receiver operating characteristic curve (AUROC).
Similarly, features extracted from upper-limb motor analysis can be leveraged for PD diagnosis. In concordance with the previous case, various ML classifiers, including NB, LR, kNN, SVM, adaptive boosting (AdaBoost), DT, RF, gaussian mixture model (GMM) and deep neural networks (DNN) classifiers, have been proposed in the literature to address this problem. To elaborate, tremor measurements have been obtained with the help of inertial sensors embedded in smartphones [55], Wii Remote [56] and spoon handles [57] to differentiate PD patients from healthy controls by exploiting conventional ML algorithms. RF and SVM models perform very well, achieving 0.94 and 0.98 AUROC, respectively. Similarly, signals from wrist-worn inertial sensors have been exploited for this cause. In [58], a moderate classification accuracy of 89% is achieved with the help of SVMs, while in [59], gait patterns are extracted from wrist-acceleration signals and detect PD patients with GMMs, achieving AUROC = 0.85. Finger-mounted sensors have also been used for recording motor data during finger-tapping or other hand movements for PD diagnosis. For example, Park et al. [60] exploit finger-mounted IMUs to train DNNs and achieve AUROC = 0.888–0.95. As one can easily observe, similar performance is obtained for all these approaches. In [59], the GMMs, which contain an unsupervised learning phase, perform slightly worse than the rest methods, while the DNN in [60] does not outperform significantly the conventional ML models.
It can be concluded that there is no significant difference between lower-limb and upper-limb analysis, regarding the performance of the proposed models for PD diagnosis based on IMUs. Moreover, it should be highlighted that sensor networks combining multiple inertial sensors’ signals from various body parts placements have been also deployed for PD detection. However, their exploitation has not significantly improved the classification performance of the previously discussed studies. Some indicative approaches are [61,62,63,64]. In [62], an SVM classifier achieves 95% accuracy, while in [63], an RF classifier achieves 86–94.6% accuracy. The highest performance (96% accuracy) is obtained in [64], with the help of a majority voting scheme over several conventional ML algorithms, while the lowest performance (68.64–73.81% accuracy) is obtained in [61] with an MLP, trained with features extracted by a convolutional autoencoder (AE), which was pre-trained on healthy subjects’ data.
Besides binary classification between PD patients and healthy people, differential diagnosis between several neurological disorders with similar manifestations can be achieved with the help of wearable inertial sensors and ML algorithms. For example, in [65,66], signals from hand-mounted inertial sensors are used to train ML models, such as RFs and SVMs, to detect PD among other neurological disorders and achieve a moderate accuracy of 72–79%. Moreover, in [67], PD patients are efficiently differentiated from essential tremor (ET) patients by an SVM trained on smartphone angular velocity signals, with 77.8% accuracy. A bit higher classification performance (89% accuracy) is achieved by Moon et al. [68] when addressing the same problem with NNs trained over fused signals from sternum-, lumbar-, wrist- and foot-worn inertial sensors. Similarly, De Vos et al. [69] managed to differentiate PD patients from progressive supranuclear palsy (PSP) patients by training an RF with signals from multiple IMUs placed at various body places, achieving 88% accuracy. Finally, gait features can classify different forms of parkinsonism with 73.33% accuracy, by deploying deep belief networks (DBNs), as indicated in [51]. As one can easily observe, SVMs, RFs and NNs are usually chosen as the best option to address differential diagnosis. Finally, the models detect PD among other diseases with significantly higher accuracy when multiple IMU signals are combined from various body placements, as seen in [68,69].
Inertial sensors have been also used for the quantification of PD-related symptoms and the estimation of the respective scores in well-known scales, such as the UPDRS and the H&Y scale. In [52], walking features are extracted from lower spine signals with a hidden Markov model (HMM) and train a kNN model to finally estimate 36-Item Short Form Survey (SF-36) scores with 27.81 mean absolute error (MAE). Similarly, gait features extracted from smartphone inertial signals can classify MDS-UPDRS severity levels with AUROC ranging between 0.92 and 0.97, by training ANNs and RF or bagged trees models [53,70]. It seems that the last two classification tasks are easier to be performed successfully than the regression one, presented in [52]. Nevertheless, it needs to be taken into account that the output scales are different in each case and that the used signals in [53,70] derive directly from the feet and not the lower spine.
Furthermore, signals from hand-mounted inertial sensors have been exploited to extract features either manually and train conventional ML models [71,72,73,74] or automatically and train DL models, such as CNNs [72], to estimate UPDRS scores. All these approaches exploit wrist acceleration or angular velocity signals, except for the study [73], where the authors make use of finger-mounted sensors. In [71], 91% accuracy is achieved with an MLP; in [72], 85% accuracy is achieved with a CNN; in [73], 96–97.33% accuracy is achieved with an RF; while in [74], 0.79–0.88 AUROC is achieved with an LR model. Unexpectedly, the DL approach [72] performs a bit worse than other conventional ML approaches, but the different placement of the sensors may have influenced the results. Finally, H&Y scale scores have also been estimated with SVM models based on tremor measurements from inertial sensors embedded in everyday objects, achieving accuracy up to 77% and correlation coefficient up to 0.97 [56,57]. Consequently, in the considered studies, PD patients are classified according to their H&Y scores with slightly lower accuracy than when they are classified according to their UPDRS scores.
Moreover, sensor networks that capture both upper-limb and lower-limb motion characteristics have been deployed for PD severity estimation as well. For example, in [75], a random undersampling boosting (RUSBoost) algorithm based on full-body inertial signals estimates H&Y scale scores, without however outperforming the previous approaches, reaching up to 78% accuracy and 0.87 AUROC. Nevertheless, in another approach [64], SVM models achieve 87.75–94.5% accuracy, which corresponds to the highest performance obtained for H&Y scale. Additionally, in [76,77] wrist- and foot-worn IMU signals are combined to quantify PD severity. In the first case, CNNs and long short-term memory (LSTM) networks are combined to estimate UPDRS scores, while in the second, the adaptive neuro-fuzzy inference system (ANFIS) is proposed. Correlation coefficient reaches up to 0.79 and 0.814, respectively. Similarly, failures of PD patients to follow the MDS-UPDRS III movement protocol have been identified with various ML algorithms and the respective accuracy reaches up to 78% [78]. In conclusion, it seems that the performance of PD severity estimation models is independent of the placement of the sensors, as upper-limbs, lower-limbs and combinational approaches demonstrate similar performance regarding this task.
Additionally, sometimes the researchers address the detection of PD-related symptoms and do not proceed to the quantification of their severity. This is the case in the studies that are presented below, where hand-mounted IMUs are used to train ML models to detect tremor and bradykinesia. Firstly, in [79] bradykinesia is detected in PD patients with 90.9% accuracy by training DNNs over hand accelerations. Similarly, Shawen et al. [80] detect both bradykinesia and tremor with the help of an RF classifier, achieving AUROC up to 0.77 and 0.79, respectively. Tremor detection, accompanied by the estimation of the respective duration, are also performed by training an MLP classifier with automatically extracted features from a CNN, obtaining AUROC = 0.887 [81]. The tremor onset can be identified with an MLP trained over manually extracted features with 92.9% accuracy, as well [82]. Finally, Channa et al. [83] discriminate PD patients with tremor from PD patients with bradykinesia and from healthy controls, with the help of a kNN model, achieving 91.7% accuracy. In conclusion, it is observed that the RF classifier in [80] performs poorer than the MLP, DNN and kNN models in the rest approaches.
On the other hand, lower-limb analysis with inertial sensors can support the detection of freezing of gait (FoG) episodes. In [84], accelerometer and gyroscope signals feed an AdaBoost classifier and detect FoG with sensitivity ranging between 81.7% and 86%. For the same cause, DNNs have been deployed, such as CNNs [85,86], LSTMs [87] or their combination [88]. The highest accuracy (91.9%) is obtained with the combination of a CNN and an attention-enhanced LSTM [88], while CNNs and LSTMs alone achieve 89% and 83.38% accuracy, respectively [85,86,87]. Furthermore, ML algorithms trained with inertial sensors signals may predict FoG episodes in a short time of period, or equivalently may classify gait sequence segments in walking, pre-FoG, FoG or post-FoG/walking phases. This problem is addressed in several studies [89,90,91,92,93,94] that mainly use the open Daphnet dataset (https://archive.ics.uci.edu/ml/datasets/Daphnet+Freezing+of+Gait, accessed on 17 February 2022). Many algorithms have been deployed to address this task, including LSTM, SVM, kNN, MLP, extreme gradient boosting (XGBoost), RF, gradient boosting machine (GBM), LR and LDA models among others. In [89], 85–95% accuracy is obtained with an LSTM NN; in [90,91,94], 77–86.1% accuracy is obtained with SVMs; in [93], 83% accuracy is obtained with LDA; while in [92], 98.92% is obtained with a kNN classifier. Generally, in this case, it can be concluded that DL approaches tend to perform better than the ones that make use of conventional ML models. Finally, gait features extracted from feet-worn IMUs can also be leveraged for strides detection or gait segmentation of PD patients with the help of hierarchical HMMs [95]. The respective f1-score ranges between 95.9% and 100%.
Besides PD symptoms monitoring, inertial sensors can be used for the monitoring of treatment-related fluctuations and the detection of treatment-induced adverse symptoms. In [96,97], on and off medication states are detected based on gyroscopic signals from an ankle-mounted sensor and acceleration signals from a knee-worn sensor, respectively. In the first case, an LSTM NN is deployed, achieving 73–77% accuracy, while in the second, an RF classifier is trained, and accuracy reaches up to 96.72%. In addition, in [98], a third class is added which corresponds to the levodopa-induced dyskinesia. A CNN is trained with acceleration signals collected from a wrist-worn sensor in a free-living environment and discriminates the three classes with 65.4% accuracy. The levodopa-induced dyskinesia can also be detected, as a binary classification problem, by training SVMs with wrist or hip acceleration and angular velocity signals, obtaining 65–82% accuracy [99]. In this case, as one can observe, DL approaches demonstrate moderate performance and do not outperform the conventional ML ones. As a matter of fact, higher classification accuracy is obtained with RF and SVM models than with CNN or LSTM networks.
Moreover, the response to the levodopa treatment can be estimated based on accelerometer or gyroscope data from wrist-worn sensors. In [74], a classification approach is presented which makes use of an LR model, with AUROC ranging between 0.82 and 0.92. On the other hand, in [58], a support vector regressor (SVR) estimates the exact values of the treatment response index with 0.69 root mean square error (RMSE). Finally, Watts et al. [100] first cluster with k-means PD patients according to their response to prescribed medication and then they classify them to these groups based on their motor characteristics. In this approach, 86.9% accuracy is obtained with an RF classifier. It is worth mentioning that none of the retrieved studies exploits DNNs to address the treatment response estimation problem.

3.2.2. Pressure and Force Sensors

Another popular type of sensors for PD diagnosis and monitoring is the one that measures applied force or pressure. The results of the current literature research indicate that these sensors have been mainly used for PD diagnosis, differential diagnosis and PD motor symptoms quantification. A group of seven relative studies that train ML models over the signals produced by these sensors to address PD-related problems is presented in Table A2 in Appendix A of this paper. As one can observe, only lower-limb analyses are considered. Even though pressure sensors can be embedded in everyday objects for upper-limb motion measurements as well, to the best of the authors knowledge, there has been no study where these sensors have been used stand-alone to collect the necessary data. Therefore, the respective studies that combine such sensors with others for upper-limb observations are presented in Section 3.2.6, which is dedicated to the combinational analysis of multiple types of sensors.
In concordance with the inertial sensors studies, PD diagnosis and severity estimation are commonly addressed problems and are discussed in five out of the seven studies considered [101,102,103,104,105]. In all these approaches, an open dataset [106] that merges three gait datasets, with vertical GRF signals from eight in-sole sensors in each foot, is exploited. More specifically, PD diagnosis is performed by training a feed-forward DNN [102], a 1D-CNN [103] and a dual-modal CNN accompanied by an attention-enhanced LSTM [104], with 99.29–99.52%, 98.7% and 99.07–99.31% accuracy, respectively. In all these cases, DL approaches achieve almost excellent classification accuracy, outperforming other conventional ML algorithms tested.
At the same time, these studies estimate the severity of PD-affected gait, with respect to the H&Y scale, achieving 98.57–99.1% [102], 85.3% [103] and 98.03–99.01% accuracy [104], respectively. In this case, the feed-forward DNN and the dual-modal CNN with the attention-enhanced LSTM clearly outperform the CNN approach, maintaining again an exceptionally high accuracy, greater than 98%. PD patients have been also classified according to their H&Y scores with the help of a time delay NN trained with Q-backpropagation [101] and an SVM classifier [105], obtaining 90.91–92.19% and 98.8% accuracy, respectively. Consequently, the SVM classifier demonstrates similar performance with the two previous DL approaches.
Furthermore, GRF signals have fed ML models for differential diagnosis of several neurological disorders. Papavasileiou et al. [107] discriminate PD patients from post-stroke patients and from healthy controls with a multiplicative multi-task feature learning model, achieving 0.88–0.994 AUROC. Additionally, in [108], kNN, SVM and RF classifiers manage to discriminate PD patients from Huntington’s disease (HD) and amyotrophic lateral sclerosis (ALS) patients, as well as from healthy controls with 81.25–90.91% accuracy. The fact that the classification of PD patients and patients with other neurological disorders is a rather harder problem than the binary classification between PD patients and healthy controls is reflected in the obtained accuracies.

3.2.3. Image and Depth Sensors

Useful information regarding the manifestations and the course of PD can also be acquired with the help of 2D or 3D video recordings or even images, in more rare occasions, after they are analyzed with computer vision techniques and ML methods. More specifically, this type of input data has been exploited for PD diagnosis, severity estimation, FoG and other symptom detection, medication state detection, segmentation of exercises and rehabilitation assessment. These insights have been extracted from 15 related studies presented hereafter, out of which 8 studies capture movements of the whole body while walking or exercising, 1 combines videos from several body parts, 4 focus on hand movements and 2 capture facial expressions. The respective findings are thoroughly presented in Table A3 in Appendix A of this paper.
Firstly, postural and kinematic features are extracted from videos that are recorded while PD patients and healthy controls are walking, to discriminate them. In [109,110], ANNs are trained with signals collected via the Microsoft Kinect (MS Kinect) sensors to diagnose PD. In [109], the combination of a CNN with an LSTM outperforms the stand-alone CNN and LSTM classifiers, reaching up to 83.1% accuracy, while in [110], the ANN classifier outperforms the SVM model, with 89.4% accuracy. In the last study, besides gait data, features extracted from finger- and foot-tapping activities have been evaluated, without however improving the classification performance. In the same vein, in [111], raw video-recorded gait sequences feed a CNN and discriminate PD patients from healthy controls with 88–90% accuracy. Finally, no studies that propose conventional ML algorithms as the best option to detect PD based on gait features extracted from videos were identified among the studies considered in the current systematic review.
Moreover, kinematic parameters from hand motor tasks can be extracted with the help of Leap Motion sensors and feed conventional ML algorithms to diagnose PD. For example, in [112], these signals train a bagged tree classifier, achieving 98.62% accuracy, while in [113], an SVM classifier demonstrates similar performance (98.4% accuracy) regarding PD diagnosis. Other tested algorithms that did not outperform the previous span across kNN, DT, LDA and LR models. Additionally, the authors in [114], train, again, an SVM model with automatically extracted features via CNNs and AEs from video recordings of hand movements to detect PD, obtaining 91.8% accuracy. They further classify PD patients with and without medication, obtaining a justified poorer performance (73.5% accuracy).
Furthermore, facial video recordings and images extracted from their frames discriminate PD patients from healthy controls as well. Firstly, in [115], an SVM classifier detects PD with f1-score = 99%, outperforming other conventional ML models and DL approaches, including recurrent neural networks (RNNs) and LSTMs. Additionally, in [116], only static information is provided to train DTs and RFs among other ML models, decreasing the obtained classification performance to 60.7–85.92% accuracy. It is obvious that data extracted from videos and not just static images can be leveraged to address PD-related problems more efficiently, as expected. Finally, in most cases hand features collected with the Leap Motion sensors diagnosed PD with higher accuracy than gait and facial features.
In addition, image and depth sensors can support the quantification of PD symptoms. In [110], PD severity is classified as mild or moderate with ANN or SVM models trained on gait, finger- or foot-tapping features extracted from MS Kinect. The highest classification accuracy is achieved with ANNs over gait data and reaches up to 95%. Similarly, a combination of two parallel LSTMs has classified UPDRS scores with 77.7% accuracy, based on gait-related MS Kinect signals [117]. On the other hand, Li et al. [118] estimate the exact values of UPDRS-III scores with an RF regressor based on 2D-videos, which recorded leg movements and toe tapping activities. The obtained RMSE is 7.765. In the same study, levodopa-induced dyskinesia severity in the Unified Dyskinesia Rating Scale (UDysRS-III) has been also estimated, this time based on video-recorded drinking and communication activities with RMSE = 2.906. Finally, Liu et al. [119] train SVMs with features extracted from video recordings of hand movements to classify bradykinesia-related MDS-UPDRS scores, achieving 89.7% accuracy. Regarding the task of severity estimation, in contrast to the diagnosis task, upper-limb analysis does not seem to be more insightful than lower-limb analysis.
In the end, some other PD-related problems that can be addressed with image and depth sensors are presented in this paragraph. Firstly, Timed Up and Go (TUG) tests have been video-recorded to shed light on several PD manifestations. In [120], the subtasks of a TUG test are automatically segmented with the help of an LSTM, obtaining 93.1% accuracy. Furthermore, in [121] an RNN with gated recurrent units (GRUs) is deployed to detect FoG episodes based on video-recorded TUG tests. The respective classification performance reaches up to 82.5% accuracy. Another task addressed based on signals acquired with MS Kinect sensors is the estimation of the adherence to medication [122]. DTs, RFs and SVMs achieve that with accuracy greater than 67.7%, when personalized models are built. Finally, in [123] a virtual physical therapist is developed. In this case, the subtasks of the performed exercises are automatically segmented with HMMs, the patients’ performance is estimated with an SVM classifier, which achieves 86.3–94.2% accuracy, and finally, the next steps regarding the next exercise to perform are recommended again automatically with RFs.

3.2.4. Voice and Audio Sensors

Speech impairment is common among PD patients and could serve as an indicator of the presence and the degree of severity of the disease. Therefore, speech recordings, which can be usually collected with the help of smartphones or other portable devices, have been exploited for PD diagnosis, differential diagnosis and severity estimation, according to the respective eight studies considered in the current literature review. These studies and their main findings are presented hereafter and are also summarized in Table A4 in Appendix A of this paper.
In four of these studies, the detection of PD is discussed. Firstly, the authors in [124,125] leverage the vocal measurements gathered in the mPower study [126] to discriminate PD patients from healthy controls. In the first case, features are automatically extracted with the help of a CNN and the classification is performed with 90.45% accuracy, while in the second case features are manually extracted in time, frequency and cepstral domain and train a XGBoost classifier, after least absolute shrinkage and selection operator (LASSO) was applied, achieving 95.78% accuracy. Similarly, Zhang et al. [127] deploy stacked AEs to extract features from smartphone speech recordings. These features are forwarded to several ML algorithms, including SVM, kNN, NB, DT and LDA models. The kNN algorithm outperforms the others in the PD detection task with accuracy ranging between 94% and 98%. In the same vein, in [128] various types of features, such as phonation, unvoiced and speech are extracted from signals collected with smartphones and acoustic cardioid sensors. Again, the kNN model outperforms the others when it is trained with phonation features, achieving 92.94–94.55%. Consequently, one can observe that simple classifiers, such as the kNN and boosting algorithms, trained with vocal features usually outperform other more complex approaches in detecting PD. Finally, differential diagnosis between PD patients, rapid eye movement sleep behavior disorder (RBD) patients and healthy controls is performed with an RF classifier based on smartphone speech recordings [129]. Each health state is detected with a relatively low sensitivity ranging between 59.4% and 74.9% for each pairwise classification.
In the last study, the exact values of the scores in several PD-related scales are also estimated with an RF model trained over smartphone speech recordings. Low MAE is obtained for each one of the included clinical scales, which span across MDS-UPDRS, Montreal Cognitive Assessment (MoCA), Epworth Sleepiness Scale (ESS), Beck Depression Inventory (BDI) and Visual Analog Scale (VAS). Moreover, UPDRS scores have been estimated with approximately 2–3 MAE by a positive transfer learning model which makes use of speech recordings captured by a portable device [130]. The same data source is exploited in [131] for UPDRS score predictions in approximately 6 months. The XGBoost regressor achieves 5.09 MAE. Similarly, scores in the motor subscale of the UPDRS (mUPDRS) are estimated with ridge regression in [132] and are predicted in approximately 6 months with XGBoost regression in [131]. The obtained MAE are 5.5 and 6.45, respectively. As expected, the prediction of the course of the disease is harder than the current estimation of its severity. This is reflected in the slightly higher error metrics in the first case. However, in both cases, very low errors are obtained, in most cases with the help of ensemble ML techniques, such as bagging and boosting.

3.2.5. Other Types of Sensors

Besides the aforementioned types of sensors, there are other smart devices that can collect signals to support PD patients’ diagnosis and monitoring. These signals may be related to muscles, brain or heart function, to body temperature or other motor-related measurements, such as displacement and keyboards dynamics. In this respect, eight related studies have been considered in this review, the details of which are presented in Table A5 in Appendix A of this paper. These studies are further presented below and address problems such as PD diagnosis, symptom detection and estimation of their severity levels, medication response estimation and emotion recognition.
Firstly, several devices that are widely used as conventional medical equipment can be adopted in free-living environments in portable versions as well. For example, EEG signals produced by portable headsets have trained feed-forward NNs to detect PD with 96.5% accuracy [133]. Surface EMG (sEMG) signals collected from wrist-bands have trained RF regression models that estimate MDS-UPDRS III scores with 0.739 correlation coefficient [134]. Additionally, Capecci et al. [135] have classified emotions as positive or negative with the help of SVMs trained over temperature, heart rate and galvanic response signals, which are acquired via a smartwatch. The respective classification accuracy ranges between 88.6% and 91.3%.
Furthermore, echo state networks (ESNs) have been trained over finger position sequences that were collected with finger-mounted electromagnetic sensors while performing finger-tapping tests to diagnose PD [136]. The respective AUROC reaches up to 0.802. Similarly, Picardi et al. [137] gather flexion measurements from smart gloves and orientation data from a wrist-worn tracking system and train SVM, ANN and genetic programming algorithms to classify PD patients and healthy controls with several levels of cognitive impairment. All the deployed algorithms demonstrate similar performance, with AUROC ranging between 0.72 and 0.99.
Useful features that shed light on several PD motor aspects can also be extracted from writing, drawing or typing activities on computer keyboards or tablet devices. In [138], spatiotemporal features are extracted while PD patients are drawing spirals on a touchscreen. These features train an MLP model that differentiates PD patients with bradykinesia from PD patients with dyskinesia, reaching up to 84% accuracy. Keystroke logs have also been exploited for PD diagnosis [139] and estimation of the medication response [140]. In both cases, DNNs have been deployed. In the first case, an LSTM with fuzzy recurrent plots outperforms CNNs, achieving 65.14% accuracy, while in the second case, an RNN achieves 76.5% classification accuracy and 0.75 AUROC. In the last study, the authors further predict the response to the prescribed medication in the future with 0.69–0.75 AUROC. As expected, the prediction task is harder than the estimation one, obtaining a lower AUROC.

3.2.6. Combination of the Previous Sensors

There are also 22 studies that combine two or more types of sensors to address PD-related problems. Some commonly encountered combinations span across inertial and pressure sensors, inertial or pressure sensors with other biosensors, e.g., sEMG, heart rate and temperature portable devices, inertial or pressure sensors with speech or video recordings, among others. Some of the problems discussed in these studies are PD diagnosis, differential diagnosis, symptom detection and estimation of their severity, as well as FoG episodes and fall risk detection. More details on these approaches are presented in Table A6 in Appendix A of this paper.
More specifically, in some studies, position, inertial and pressure signals are fused to diagnose PD, detect its manifestations and quantify their severity. For example, Aharanson et al. [141] develop a smart support walker that contains two encoders on the wheels, two force sensors underneath the hand grips and a tri-axial accelerometer. They reduce the dimensionality of the derived signals with principal component analysis (PCA) and feed an LDA model to detect PD with 91–96% sensitivity. Inertial and in-sole pressure sensors have been also combined for FoG detection and prediction, by identifying the pre-FoG state [142]. Selected features in time and frequency domain train a RUSBoost classifier and detect or predict FoG events with 81.4–98.5% and 61.9–78% sensitivity, respectively. The results confirm that FoG prediction is a harder task than FoG detection. Moreover, Wu et al. [143] classify PD patients according to their UPDRS-III scores based on displacement and acceleration signals from hand movements. NNs outperform conventional ML models, such as SVMs and RFs, obtaining 91.18–95.30% accuracy.
Furthermore, there are three studies that combine acceleration and EMG signals to support PD patients at various aspects of the disease. Firstly, Cole et al. [144] leverage these signals to detect tremor and dyskinesia with the help of a HMM and a DNN, respectively, and estimate their severity levels with a Bayesian classifier. The severity of tremor is estimated with slightly higher sensitivity (95.2–97.2%) than the severity of dyskinesia (91.9–95%). Secondly, Hossen et al. [145] perform differential diagnosis between PD and ET patients with an MLP classifier, achieving 91.6% accuracy. Finally, Tahafchi et al. [146] detect FoG episodes with the help of a fully connected NN, achieving 0.906–0.963 AUROC. The results indicate that portable EMG devices can support FoG detection as effectively as pressure sensors when they are combined with IMUs.
Similarly, in [147], force, inertial and mechanomyography (MMG) signals train three distinct classifiers, kNN, MLP and AdaBoost, to diagnose PD and quantify its symptoms. More specifically, PD patients are discriminated from healthy controls with 96.6% average accuracy, PD patients with positive UPDRS scores are discriminated from those with zero due to DBS with 89% average accuracy and UPDRS scores levels are classified with 85.4% accuracy. Moreover, skin-mounted sensors with acceleration, gyroscope and thermometer measurements collect signals during TUG tests, which train CNNs for multi-source multitask learning to detect fall risk and estimate the severity of PD [148]. In the first case, a 94% f1-score is obtained, while in the second, 0.06 RMSE is achieved. In all these studies that make use of inertial or pressure sensors combined with other biosensors, PD severity is successfully estimated, with high accuracy in the classification case and low error in the regression.
Moreover, sometimes, signals from inertial sensors and speech recordings are combined to diagnose PD and quantify its severity. For example, an ensemble learner consisted of kNN, SVM and NB models detect PD with accuracy 99.8% when it is trained over phonation recordings and acceleration signals collected with a smartphone [149]. In the same study, MDS-UPDRS levels are classified with a kNN model, achieving 90.5–98.5% accuracy. Similarly, in [150], accelerometer, gyroscope and magnetometer data are combined with speech recordings acquired from a headset and train an extreme learning machine (ELM) that classifies PD patients at several severity levels and healthy controls with accuracy ranging between 92.45% and 95.93%. There is no significant difference between the obtained accuracies, regarding the severity classification of these two approaches. On the other hand, Papadopoulos et al. [151] combine accelerometer with typing dynamics data collected from a smartphone to detect tremor or fine-motor impairment and finally diagnose PD by training NNs, achieving 0.834–0.868 AUROC.
Moreover, inertial or pressure measurements are usually used together with video recordings or images to feed ML algorithms and diagnose PD. For example, in [152], acceleration signals and silhouette images are collected during cooking. Convolutional variational autoencoders (VAEs) are deployed to extract features that will feed NNs to finally differentiate PD patients from healthy controls with 66% f1-score. Similarly, Wahid et al. [153] diagnose PD with the help of an RF classifier trained over spatio-temporal gait features extracted from videos and GRF signals. The respective AUROC reaches up to 0.96. Albani et al. [154] exploit videos that record upper-limb movements for pretraining and then lower-limb motor signals from IMUs to train a kNN that detects PD with 91.5–98.6% accuracy and estimates its severity according to the UPDRS with 60.7–79.1% accuracy. It is observed that RF and kNN models detect PD more accurately than NNs, which are deployed in [152]. Visual and audio features can also be extracted from short interview audio-video clips to train an NN that will estimate facial expressivity of PD patients [155]. A low f1-score (55%) is obtained in the classification approach, while MAE = 0.48 is obtained in the regression approach.
Additionally, there are a few studies that take advantage of smart pens equipped with various types of sensors. All of them differentiate PD patients from healthy controls by training either conventional ML algorithms or DNNs. In [156], an RF classifier is proposed which is trained with the most important principal components of pressure, tilt and acceleration signals. In this approach, PD is detected with 88.8–89.4% accuracy. Similarly, Gallicchio et al. [157] diagnose PD with an ensemble model of deep ESNs trained with pen position, pressure and grip angle signals, achieving 89.33% accuracy. The aforementioned signals along with recordings from a microphone embedded in the smart pen have also been fed to either a pre-trained CNN or an optimum path forest (OPF) classifier and PD is successfully detected with 77.92–87.14% accuracy [158]. To conclude, all these approaches demonstrate similar performance in diagnosing PD. Only the OPF model seems to perform slightly poorer than the rest methods. Finally, data acquired from a smart pen can be combined with shoe-mounted IMU signals and train an AdaBoost classifier that manages to detect PD with sensitivity up to 100% [159].
There are also two studies [160,161] that use multiple types of data acquired with the help of a mobile application in the mPower study [126]. These data include touchscreen logs from a tapping activity, accelerometer signals from a walking activity, performance in a memory game and vocal measurements. In both studies, ensemble methods that combine DL methods, such as CNNs and RNNs, and conventional ML approaches, such as LR and RF, have been deployed to detect PD with f1-score 82% and 87.1%, respectively. Finally, in [162], several data have been collected from ambient sensors in a smart house environment, including infrared motion sensors on the ceiling, light sensors, magnetic door sensors, temperature sensors and vibration sensors on selected items, along with signals from two wearable inertial sensors. All these data after dimensionality reduction with either k-means or random resampling were used to train an AdaBoost classifier that diagnoses PD with 80–86% accuracy.

4. Collective Comparison and Overall Insights across the Studied Approaches

The literature review presented herein considers 112 research studies that leverage IoT technologies and ML methods to support neurologists at various decision-making processes involved in every stage of PD and thus potentially advancing PD patients’ care and quality of life. The considered approaches are grouped per type of sensor used for data collection. For each individual study, the problem addressed, the dataset used for training and evaluation, the input data, the processing techniques and the AI algorithms deployed, as well as the evaluation metrics and the final performance of the proposed models are briefly presented. Different approaches within each category of type of sensors are compared with each other, in terms of performance, to finally infer which ML algorithms are most suitable to address the discussed problems with respect to the specific type of input data, each time. After presenting the results, additional pieces of evidence should be provided related to the obtained average performances in the scope of both intra- and inter-categories of deployed sensors, to identify ill-addressed problems and further infer which type of input data may be more insightful to address each of the discussed PD-related problems.

4.1. Sensors and Devices Deployed

The types of sensors that are deployed for data collection and their popularity among the studies considered are depicted in Figure 3. As Figure 3 indicates, inertial sensors or IMUs, including accelerometers, gyroscopes and magnetometers are by far the most popular data source among the considered studies. Combination of different types of sensors follow in popularity, where again most of the times one from the multiple data sources used corresponds to inertial sensors. Moreover, data collected with image or depth sensors are another popular type of input data. On the other hand, speech recordings collected with audio or voice sensors and force signals collected with pressure sensors are more rarely encountered in the literature. Finally, it is worth mentioning that despite EEG, EMG and other medical signals are commonly used in the clinical practice for PD diagnosis, severity estimation and course prediction, there are very few studies considered that make use of portable biosensors, which monitor brain, heart or muscles functionality. These types of sensors fall in the less popular category, named as other, which includes besides portable EEG, sEMG and ECG devices, thermometers, galvanic response sensors, as well as flexion sensors, position sensors, encoders, tracking devices, time loggers, ambient sensors, including light, temperature, vibration sensors, etc.
Additionally, the systematic review presented in this paper is focused on IoT approaches, which means that all the aforementioned sensors must be embedded in devices that can transfer the collected patients’ data to other computational machines or remote databases via Wi-Fi, Bluetooth or any other communication protocol. To this end, the discussed types of sensors may be embedded in smartwatches, smart bracelets, shoes’ insoles, other wearable devices, smartphones, video cameras, other portable devices and other everyday objects in general.
As Figure 4 indicates, the most widely encountered type of device in the literature corresponds to other wearable sensors or sensors that are mounted somehow to the patients’ corpus, e.g., via flexible bands, such as Shimmer3 or Physilog sensors. Mostly IMUs are embedded to this type of devices, but there are some other types of sensors and one reference to a pressure sensor that were embedded to these devices among the studies considered. The second most popular type of devices deployed is the smartphones. The smartphones are leveraged for their in-built sensors, which include but are not limited to IMUs, microphones, cameras and others. Stand-alone video cameras follow in popularity, including the Microsoft Kinect tool and the Leap Motion sensor, which are mainly used for video recordings but were also used once for speech recordings as well. Moreover, in 18 studies everyday objects that do not fall into other categories were deployed. For the specific studies considered in this review, these objects span across smart pens, keyboards, Wii controllers, support walkers and spoon handles. Various types of sensors, including inertial, pressure, audio and other are embedded to these objects. On the other hand, insole sensors are deployed solely for collecting gait-related pressure signals from the respective eight studies presented in Section 3.2.2. Similarly, smartwatches and smart bracelets are exploited almost exclusively for collecting inertial signals. Finally, in a few studies some other portable devices are used for speech recordings and other measurements.

4.2. Meta-Analysis over Sensor-Specific Results Obtained

For each category of sensors, the problems addressed and the respective average obtained accuracies are presented. Each subsection dedicated to one type of sensor is accompanied by one column chart which summarizes which problems are addressed based on the respective type of sensors and one boxplot that illustrates the performance range achieved for each problem addressed. Boxplots are selected for the latter cause as they offer an excellent illustration of the dispersion of the data. More specifically, the mean of the values is depicted by an ‘x’ and the median by a horizontal line which divides the data points in a top and a bottom half. The medians of these two halves form a box which corresponds to the interquartile range. The values outside this range are depicted by two whiskers (vertical lines), and there may also be some outliers, as distinct data points that exceed 1.5 times or more the interquartile range.

4.2.1. Inertial Sensors

Based on the results presented in Section 3.2.1, it is evident that inertial sensors have already been widely used in the literature to address various PD-related problems with the help of ML and DL algorithms. Diagnosis and differential diagnosis are two commonly addressed problems of this category. Diagnostic models demonstrate very high performance, with approximately 88.5% average accuracy and 89.5% median accuracy, while the performance of differential diagnosis models is a bit poorer, with accuracy ranging between 72% and 89%. However, if one takes into account that it is much more difficult to differentiate patients with similar manifestations, the obtained performance can be considered very satisfactory.
Other popular problems addressed based on the usage of inertial sensors are the detection and the quantification of PD symptoms. Tremor, bradykinesia and levodopa-induced dyskinesia detection corresponds to a classification problem, for which approximately 84.5% average accuracy is obtained from the considered studies. FoG episodes seem a little easier to be detected, as the average accuracy for this classification task is approximately 90%. Regarding the severity estimation of PD symptoms, the classification approach is presented in more studies than the regression and the respective average accuracy (approximately 83.5%) is a bit lower than the accuracies obtained in the two previous problems, which can be attributed to the increase in targets-classes. Finally, the detection of medication-related fluctuations and the estimation of the patients’ response to medication are some problems that are addressed more rarely based on data collected with inertial sensors. The respective average accuracy is again very satisfactory but remains the lowest achieved (77%) regarding the detection of fluctuations among all the other problems. Finally, identification of PD subtypes, gait segmentation and rehabilitation assessment are some problems that are addressed only once among the studies considered based on inertial signals.
The occurrences of all the problems tackled with inertial data are summarized in the column chart of Figure 5, which confirms that PD diagnosis, symptom severity estimation and FoG detection are the most widely addressed problems based on inertial data. A wide variety of ML and DL algorithms have been proposed for all these cases. Moreover, the obtained range of performances for each category of tasks is illustrated in the boxplot of Figure 5, based on the accuracy metric, which is the simplest and most widely used, among the considered papers, evaluation metric. Studies that do not apply this metric in their evaluation stage are not considered in the boxplot of Figure 5. Consequently, the construction of this boxplot is based on 45 discrete models out of the 52 inertial sensors-based studies, which are considered in this review and for each one of them only the best performing proposed method is taken into consideration.

4.2.2. Pressure and Force Sensors

As the column chart of Figure 6 indicates, very few studies that address solely PD diagnosis, estimation of the respective severity and differential diagnosis based on pressure sensors have been identified across the considered studies. No force sensors-based studies that address the problems of FoG detection, PD symptoms and their fluctuations detection, as well as medication response estimation were identified. Moreover, as the boxplot of Figure 6 indicates, when pressure or force sensors are deployed, PD diagnosis is performed with approximately 95% average accuracy and 99% median accuracy, severity estimation is performed again with approximately 95% average accuracy and 98% median accuracy, and differential diagnosis is performed with slightly lower accuracy (90%). Again, only the studies that make use of the accuracy metric for the evaluation of the proposed models are taken into consideration for the construction of this boxplot. This time, only one out of the seven studies discussed in this category is excluded.

4.2.3. Image and Depth Sensors

As the column chart in Figure 7 indicates, PD diagnosis is the most commonly addressed problem among the considered studies, based on input data collected from image and depth sensors, while estimation of PD symptom severity follows. In more rare cases, this data source has been used to detect medication state, segment exercises and motor tests and assess rehabilitation, as well as to detect FoG episodes and other PD-related symptoms. It is also worth mentioning that no studies addressing the problem of differential diagnosis based on this type of sensors were identified. Consequently, a reasonable question raised is whether video recordings can differentiate patients with similar manifestations. Moreover, although computer vision has been widely used for mood estimation of healthy people, no studies were identified to address this problem based on PD patients’ data.
Finally, the obtained accuracies for each category of addressed PD-related problems based on video recordings or imagery input data are depicted in the boxplot of Figure 7. As 2 out of the 15 studies discussed do not measure the performance of the proposed methods with the accuracy metric, they are excluded from this boxplot. By taking into consideration the remaining studies, the following observations can be made. The best performance (96.5% average accuracy) is achieved for the segmentation of exercises and motor clinical tests in subtasks. Rehabilitation assessment tasks are also effectively performed based on video recordings, with 89.5% average accuracy. PD diagnosis and severity estimation follow with average accuracy slightly greater than 87%. Finally, ML algorithms perform a bit poorer in detecting FoG episodes (82.5% accuracy), dyskinesia (71.4% accuracy), as well as in estimating medication adherence (71% average accuracy), when images and videos are used as inputs.

4.2.4. Voice and Audio Sensors

Based on the results presented in Section 3.2.4, speech signals have been exploited for PD diagnosis in four studies, for quantification of symptom severity in four studies and for differential diagnosis in only one study. No studies that address the problems of fluctuations detection and medication response estimation were identified. Every study in this category that addresses the symptoms quantification problem proposes a regression and not a classification model. Consequently, their performance cannot be compared with the performance obtained in the diagnosis task, in terms of the obtained accuracy. Moreover, the accuracy metric is not used in the evaluation of the differential diagnosis which is performed in [129]. Therefore, only the accuracies of ML algorithms regarding PD diagnosis from four studies could be compared with each other. In all these cases, the obtained performance is very high, with accuracy greater than 90% (approximately 94% average accuracy). Due to the lack of more accuracy data points, the presentation of the respective boxplot diagram, in alignment with the previous subsections is considered to be unnecessary.

4.2.5. Other Types of Sensors

Moreover, various PD-related problems, such as diagnosis, detection of symptoms and estimation of their severity, estimation of patients’ response to the prescribed medication and emotion recognition, are addressed based on input data provided by a wide variety of other sensors. Due to this variability, it is considered that there is no reason to proceed to a grouping of approaches or to a thorough comparison between the different considered studies. The only common addressed problem is the diagnosis of PD and the highest accuracy for this task is obtained when EEG signals are used. Approaches based on finger-mounted position sensors and hand-mounted flexion sensors follow. Finally, the lowest accuracy is obtained when keystroke log data are deployed.

4.2.6. Combination of the Previous Sensors

As the column chart of Figure 8 indicates, the combination of more than one different type of sensor has been exploited mainly for PD diagnosis and symptom severity estimation, and more rarely for FoG and other PD symptom detection, fall risk identification and differential diagnosis. Moreover, no studies that address the problems of fluctuations detection and medication response estimation were identified. Additionally, the obtained accuracies for each category of addressed PD-related problems based on fusion of data obtained by combinations of various types of sensors are depicted in the boxplot of Figure 8. The main classification tasks, which are evaluated based on the accuracy metric, are PD diagnosis, differential diagnosis, severity estimation and cognitive impairment detection. Therefore, 13 out of the 22 total studies, which address these problems, are considered in this boxplot. In all these cases, features extracted from multimodal sensory data lead to very satisfactory results. PD is detected by an average accuracy of approximately 89%. On the other hand, PD symptoms are quantified with approximately 88% average accuracy and 91% median accuracy. Consequently, despite severity estimation is a hard problem, it can be addressed with the help of various sensory data to a sufficient extent, with an approximately equal performance to the one achieved in the diagnosis case. Finally, differential diagnosis demonstrates 91.6% accuracy, cognitive impairment detection demonstrates 86% accuracy, and the remaining tasks are evaluated based on other metrics and therefore are not considered in the boxplot of Figure 8.

4.3. Overall Addressed Problems Related to Parkinon’s Disease and Obtained Accuracies

4.3.1. Frequency of Problems Addressed

As Figure 9 indicates, diagnosis is the most widely addressed problem among the considered studies and corresponds to a binary classification problem between PD patients and healthy controls. The quantification of PD symptom severity follows, which can be both a classification and a regression task, most of the time with respect to specific scales, such as the UPDRS and the H&Y scale. However, most of these approaches estimate the current patients’ situation and do not make any predictions in the future, which would be more fruitful for personalized treatment suggestions. The detection of FoG episodes and other PD symptoms, including tremor, bradykinesia, dyskinesia, cognitive impairment and deficient facial expressivity, are also quite popular topics in the current literature search. These are binary classification problems between the presence and the absence of certain manifestations. Differential diagnosis follows, which corresponds to either a binary or a multi-class classification problem between PD patients and patients with other neurological disorders.
Despite fluctuations detection, medication state detection, estimation of patients’ response to medication and rehabilitation assessment are problems of utmost importance to achieve precise medicine, they are encountered more rarely in the literature. Most of the times, these approaches correspond to classification problems, but they can also be seen as regression problems, in the terms of, e.g., specific treatment indices estimation. Additionally, gait and exercises segmentation, which can be addressed by both supervised and unsupervised learning techniques, is presented in few studies. Finally, some other classification problems, which are discussed only once across the studies considered in this paper, include fall risk detection, emotion recognition and PD subtypes identification. All these findings are summarized in the bar chart of Figure 9.

4.3.2. Overall Performance Comparisons per Task

Firstly, diagnosis, which is the most common addressed problem, is performed based on input data produced from all the types of sensors discussed. In every case, classification results are very satisfactory and average accuracy is greater than 87%. The highest average accuracy is obtained for studies based on pressure signals (approximately 97% accuracy) and then studies based on voice recordings. Other types of sensors and combinational analyses follow. Finally, approaches based on inertial, image and depth sensors perform slightly worse, with approximately 87% average accuracy.
Additionally, severity levels are again more accurately estimated by pressure signals (95% average accuracy). Video recordings and combinations of different types of sensors follow, while approaches based on inertial signals achieve 85% average accuracy. Moreover, despite severity estimation is performed based on data produced from all the types of sensors discussed, speech recordings and other types of sensors are solely deployed for PD symptom severity regression and not classification. Therefore, the performance of the respective models cannot be compared with the others, in terms of accuracy. Generally, the classification of PD patients to severity levels-categories, with respect to various motor and non-motor symptoms, is a more difficult problem than PD detection. This is reflected in the slightly lower accuracies obtained in the first case, in almost every category of sensory data and it can be attributed to both the increase in similarity among the classes-targets and the increase in the number of the classes.
The highest obtained accuracies from all the proposed methods per category of sensors for these two popular tasks (PD diagnosis and severity estimation) are depicted in Figure 10. These diagrams confirm that inertial sensors are more widely exploited in the literature compared to other types of sensors for both diagnosis and severity estimation. Moreover, one can easily observe that independently of the type of sensors leveraged to collect input data, there are some ML or DL models that reach almost excellent classification accuracy (greater than or equal to 95%) with respect to both problems. The respective average accuracies range between 87% and 97%, demonstrating the predictive power of the hidden information of all the types of sensory data considered in the current literature review.
Another interesting question is whether the fusion of different types of data improved the respective performance of models trained on single data sources. Combinational analyses are mainly proposed for PD diagnosis and severity estimation. In both cases, it is easily observed that the average accuracy increased slightly when different data sources were combined to address these problems. However, this is not a statistically significant increase. Finally, as already mentioned, pressure sensors have led to higher average accuracies in diagnosis and severity estimation tasks than inertial sensors. However, this may be misleading, as there are many more inertial sensors-based studies which make use of many different datasets, while most pressure sensors-based studies leverage the same dataset. Therefore, the results obtained by the inertial sensors-based studies may be more stable and trustworthy.
Continuing with the remaining less frequently addressed problems, FoG episodes and other PD symptoms are detected almost exclusively by leveraging inertial sensors. In the first case, 90% average accuracy is obtained with inertial sensors, while there is also one video-based approach which detects FoG with slightly lower accuracy. In the second case, PD symptoms are detected with 85% average accuracy with inertial signals, while there is also a video-based approach which achieves slightly lower accuracy and an approach based on other types of sensors which demonstrates similar average accuracy. Finally, the exploitation of more than one types of sensors has been explored in both cases. However, the respective results are not evaluated with the accuracy metric, thus their performance is not compared with the inertial-based and video-based approaches.
Moreover, differential diagnosis is performed with every category of sensors presented, except for video recordings and other types of sensors. Most studies addressing this problem are based on inertial or pressure sensors. In concordance with the binary classification problem of diagnosis, the highest average accuracy (approximately 88%) is obtained when force sensors are deployed. Generally, classification performance for differential diagnosis is significantly lower than what is obtained for discrimination of PD patients and healthy controls. This can be attributed to the similarity among manifestations of different neurological disorders, as well as to the increase in the number of targets-classes, in some cases.
Finally, some other medication-related problems, such as the estimation of patients’ response to medication, the detection of symptoms fluctuations and whether a patient is on or off medication, are addressed with the help of inertial, image, depth and other sensors. No studies based on pressure or voice sensors were identified for this cause. The performance is slightly higher (88% average accuracy) for medication response estimation, when inertial sensors are used. Regarding fluctuations and medication detection, all the types of the deployed sensors demonstrate similar performance, with average accuracy ranging between 77% and 81%. Consequently, it can be concluded that this is a slightly harder problem than the ones discussed so far. In contrast, the segmentation of walking tests and other physical exercises to subtasks and phases, as well as the assessment of patients’ rehabilitation have been performed with exceptionally high average accuracy (96% and 90%, respectively), by video recordings. There is also one inertial sensors-based study which address the problem of gait segmentation, but the accuracy metric is not applied in the evaluation process.

4.4. Ranking of Machine Learning Algorithms

All the considered approaches take advantage of AI, ML and DL algorithms, to predict, estimate or infer the desired results in each respective task. In Figure 11, all algorithms encountered in this systematic literature review are gathered. These algorithms correspond to classification, regression, clustering or simply automatic feature extraction and modelling techniques. The considered approaches span across probabilistic methods, such as NB, probabilistic neural networks (PNNs), ANFIS, HMMs, Bayesian neural networks (BNNs) and DBNs; simple conventional ML algorithms, such as LR, LDA, quadratic discriminant analysis (QDA), kNN, OPF, DTs and SVMs; ensemble learning methods, such as RFs, AdaBoost, RUSBoost, XGBoost, GBMs, bagged trees, boosted trees and extra trees, among others; and neural networks, such as MLPs, CNNs, LSTMs, RNNs, ESNs, self-organizing maps (SOMs) and AEs, among others. Many of the previous algorithms are used for both classification and regression tasks. Some other proposed models are suitable only for regression, including linear, ridge, LASSO and stepwise regression. Additionally, some clustering techniques used include k-means, k-medoid algorithms and GMMs. Finally, a genetic programming approach has also been discussed in one study.
As the blue bar indicates in Figure 11, the three most commonly evaluated models are SVMs, RFs and the kNN algorithm. CNNs follow, along with DTs, NB, LR, LDA models, MLPs, LSTMs, other NNs and other ensemble methods. The rest algorithms are encountered more rarely in the literature to address PD-related problems. The fact that these algorithms were discussed in many of the considered approaches does not necessarily mean that they were considered in the final proposed solutions, as the best performing models. The most commonly proposed algorithms (orange bar in Figure 11) are CNNs, SVMs and RFs, while the kNN algorithm, other NNs, MLPs, LSTMs and other ensemble methods follow. Indeed, although some simple models, such as DTs, NB, LR and LDA, are widely tested, they are rarely selected as the best performing solution when other, possibly more complex, algorithms are also tested in the same study. Regarding the respective ratio, CNNs, other NNs and ensemble methods demonstrate far better efficiency, as they are selected as the best performing solution in 85%, 80% and 75%, respectively, of the occasions that they are evaluated in. LSTMs, MLPs, AdaBooost and RFs follow, while SVMs and kNN demonstrate lower efficiency ratio. The rest methods either demonstrate even lower efficiency or are not taken into consideration in the ratio-based sorting, as they have been tested very few (equal or less than 5) times in the literature and the obtained results would be misleading and would lack credibility. Finally, it is worth mentioning that more and more complex DL and ensemble learning approaches are mainly proposed in the last few years.

4.5. Subjects Enrolled

Another aspect which affects largely the obtained performance of the proposed ML models is the number and the variability of the subjects enrolled in each study. In the current review, most of the considered studies enroll a relatively small group of PD patients, healthy controls and patients with other neurological disorders. As the first column chart of Figure 12 indicates, the vast majority of the considered approaches enroll 10–100 subjects. Few studies enroll 100–1000 subjects and less than or equal to 10 subjects and only four studies enroll more than 1000 subjects. Moreover, in 47 approaches solely PD patients are enrolled, while in the rest of the approaches both PD patients and healthy controls or patients with other neurological disorders are enrolled to enable a differentiation between these groups. As the second column chart of Figure 12 indicates, in most cases, the classes of PD patients and other subjects are balanced, with PD patients rate ranging between 40% and 60%. When datasets are unbalanced, PD patients outnumber other groups in slightly more cases. The opposite is usually observed when large cohorts are exploited with more than 100 subjects enrolled. It is also worth mentioning that in six approaches, the problem addressed refers to a multiclass classification, where PD patients’ rate is usually much lower than 50% of the enrolled subjects, as expected. To conclude, despite the relatively small population size in most cases and the possible imbalanced classes in some cases, ML and DL models generalize reasonably well.

5. Discussion and Conclusions

The systematic review presented in this paper considers 112 studies, published in the last decade, which propose ML models trained with data collected via sensors and IoT technologies to address various PD-related problems. PD is a neurodegenerative disease that affects many, mainly elderly, people worldwide. The studies considered in this review provide evidence that useful knowledge can be extracted from sensory data with the help of AI methods, regarding motor and non-motor manifestations of PD, enabling doctors to make evidence-based decisions. This could potentially support personalized treatment suggestions and ensure high quality PD patients’ care even remotely. Thus, easing the burden of the healthcare system and moving from a clinic-centric to a patient-centric approach could be achieved. Therefore, it is crucial to summarize some early achievements obtained in this domain and identify remaining challenges and steps to be made in this direction.
Based on the results obtained by the presented review, it is concluded that inertial sensors have already been studied in depth for PD diagnosis and monitoring, while other less frequently used include image/depth, voice and pressure sensors, among others, as well as their combination. Moreover, there are indications that portable medical devices monitoring brain, heart and muscle functionality or other bio-signals, such as body temperature, have not yet been studied to a sufficient extent. Similarly, ambient sensors were encountered extremely rarely in the literature. It is very important that more experiments, exploiting other types of sensors rather than IMUs should be performed to address various PD-related problems.
Furthermore, the deployed sensors in the majority of the considered studies were not embedded in everyday objects and portable devices that patients are familiar with. Consequently, the need to explore the deployment of more everyday devices and wearables, such as smartphones and smartwatches among others is highlighted, to ensure unobtrusiveness and minimize the risk of PD patients’ stigmatization during the experiments. The wider deployment of these sensors is also expected to facilitate the execution of experiments in free-living environments with the help of IoT infrastructures. The vast majority of the considered studies were performed in laboratories or clinical environments and PD patients performed instructed movements. Only a handful of approaches examined the predictive power of sensorial data under free-living conditions. However, moving in this direction is of utmost importance, as sensors and IoT technologies need to be evaluated under real-world conditions, to finally achieve PD diagnosis and care revolution.
Additionally, to get more accurate measurements, the optimized position of each sensor needs to be identified. It is inferred from the considered studies that more accurate gait features were extracted when inertial sensors were mounted to the feet than when mounted to the waist. Similarly, the optimized arrangement of multiple sensors needs to be determined to maximize knowledge extraction and deploy methods for efficient multimodal data combination. Furthermore, sampling frequency should be fine-tuned in order to avoid scaling issues and minimize power consumption [163], while at the same time ensuring there is no significant information loss. Another crucial issue is the fast data transmission to remote databases and cloud servers, which may need to be performed even in real time, in some applications. The commercialization of 5G networks and the development of 6G networks prototypes is expected to solve this problem [164].
Moreover, PD-related problems addressed by the studies considered in this review include diagnosis, differential diagnosis, severity estimation, detection of FoG and other PD manifestations, fluctuations detection and medication response estimation, among others. PD diagnosis and estimation of symptom severity levels are the most in-depth studied problems, addressed with every type of sensor discussed. However, in the vast majority of the considered studies the current PD patients’ state was estimated and only in a handful of them, the course of the disease was predicted in the future. These predictions could actually revolutionize PD care by offering neurologists useful insights in advance and enabling them to act proactively, by, e.g., increasing the proposed levodopa daily dosage. Therefore, researchers should focus more on conducting such experiments based on various types of sensors. Additionally, other tasks such as fluctuations detection, medication response estimation and PD subtypes identification, which are of utmost importance to shed light on inter- and intra-patient variability and hence proposing personalized treatments, are addressed extremely rarely based almost exclusively on inertial signals. Consequently, it is highlighted that more experiments should be conducted in this direction. Finally, across the studies considered, specific popular tasks were not addressed at all by specific types of sensors, such as differential diagnosis based on video recordings. In this respect, it would be interesting to explore various types of sensors in new PD-related problems.
In principle, it can be concluded that very good results have already been achieved for all the discussed PD-related problems with the proposed AI methods based on each type of the deployed sensors. Some of these tasks can be addressed more easily and satisfactorily than others. For example, higher accuracy is obtained for diagnosis rather than differential diagnosis, prediction of symptom severity and fluctuation detection. Similarly, some types of sensors tend to lead to more accurate estimations, with respect to specific tasks, but that is also dependent on the specific datasets used, especially when few studies are considered in some categories of sensors. It is worth mentioning that in most cases, the combination of multiple types of sensors did not lead to a statistically significant increase in the obtained performance.
In any case, the most common best performing ML models correspond to CNNs, LSTMs, MLPs, other NNs, as well as ensemble learning techniques, such as boosting and RFs. It is clear that more complex approaches based on ANNs and metaclassifiers usually outperformed the simpler ones. The simpler approaches that follow are SVMs and the kNN algorithm. Moreover, the novel technique of transfer learning, which was discussed in a handful of the studies considered could be further tested to investigate whether it could support DL approaches and ensemble methods to obtain higher accuracies in PD-related problems, while at the same time it could potentially decrease training time. Additionally, in the vast majority of the studies considered, supervised approaches were discussed and only extremely few of them corresponded to semi-supervised approaches. Since a shift to collect patients’ data in free-living environments is highly suggested and that would indispensably lead to a large volume of unlabeled data, it is of utmost importance to follow more semi-supervised and self-supervised approaches that enable the exploitation of unlabeled data as well.
Moreover, to enable the exploitation of DNNs and recent advances in the big data era to address PD-related problems, larger cohorts of PD patients should be enrolled. In this respect, researchers should be encouraged to share the acquired data with the research community, to potentially enable the creation of data lakes with respect to PD. Furthermore, it would be useful to build both generalized and personalized models to support optimized predictions and estimations for new users-patients and patients who have already provided abundant data as well. In this direction, special care should be taken to maintain long-term patients’ databases and to leverage the digitalization of medical records. Finally, during every stage of the respective pipeline, very strict security guidelines should be followed to ensure patients’ data privacy and security. The advances in distributed systems and blockchains could support this cause [165].
To conclude, there is evidence that ML models and IoT technologies can revolutionize the way that PD and other chronic diseases are diagnosed and treated. As the promising results of the last decade indicate, the adoption of smart technologies in clinical practice can support clinicians in several decision-making processes, potentially reducing the current extremely high healthcare costs and counterpoising the consequences of the medical resources shortage. Moreover, the continuous remote monitoring of PD patients with wearable or non-wearable sensors could potentially provide much more useful information and shed light on PD aspects that otherwise may not be perceived through follow-up appointments. To achieve that, more sensors could be deployed in larger cohorts to feed novel ML and DL models with data and achieve more accurate predictions and estimations in currently ill-addressed PD-related problems, overcoming the respective challenges. In this way, AI and IoT interventions may finally support precise medicine and help clinicians to propose personalized treatment schemes and potentially maximize patients’ responses to them.

Author Contributions

Conceptualization, K.-M.G. and I.R.; methodology, K.-M.G.; validation, K.-M.G., I.R. and K.D.; formal analysis, K.-M.G.; investigation, K.-M.G. and I.R.; data curation, K.-M.G.; writing—original draft preparation, K.-M.G.; writing—review and editing, I.R.; visualization, K.-M.G. and I.R.; supervision, I.R.; project administration, I.R.; funding acquisition, K.D. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the European Union through the Horizon 2020 Research and Innovation Programme, in the context of the ALAMEDA (Bridging the Early Diagnosis and Treatment Gap of Brain Diseases via Smart, Connected, Proactive and Evidence-based Technological Interventions) project under grant agreement No GA 101017558.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

Not applicable.

Conflicts of Interest

The authors declare no conflict of interest.

Abbreviations

The following abbreviations are used in this manuscript:
2DTwo Dimensions
3DThree Dimensions
AdaBoostAdaptive Boosting
ACCAccuracy
ADLActivity of Daily Living
AEAutoencoder
AIArtificial Intelligence
ALSAmyotrophic Lateral Sclerosis
ANFISAdaptive Neuro-Fuzzy Inference System
ANNArtificial neural Network
ANOVAAnalysis of Variance
ARAutoregressive
ARMAAutoregressive Moving Average
AUAction Unit
AUCArea Under the Curve
AUPRArea under the Precision-Recall Curve
AUROCArea Under the Receiver-Operating Characteristic Curve
AvgAverage
BDIBeck Depression Inventory
BEBackward Elimination
CADComputer-Aided Diagnosis
CARTClassification and Regression Tree
CFSCorrelation-based Feature Selection
CNNConvolutional Neural Network
CTComputed Tomography
CWTContinuous Wavelet Transform
DADiscriminant Analysis
DBNDeep Belief Network
DCALSTMDual-modal Convolutional Neural Network + Attention Enhanced Long Short-Term Memory
DCNNDual-modal Convolutional Neural Network
DDOther Movement Disorder
DESNDeep Echo State Network
DFTDiscrete Fourier Transform
DLDeep Learning
DNNDeep Neural Network
DTDecision Tree
DTWDynamic Time Warping
DWTDiscrete Wavelet Transform
ECGElectrocardiography
EEGElectroencephalography
EEREqual Error Rate
ELMExtreme Learning Machine
EMGElectromyography
ESNEcho State Network
ESSEpworth Sleepiness Scale
ETEssential Tremor
EWPTEmpirical Wavelet Packet Transform
EWTEmpirical Wavelet Transform
FAFactor Analysis
FFTFast Fourier Transform
FIRFinite Impulse Response
FLDAFisher Linear Discriminant Analysis
FNRFalse Negative Rate
FoGFreezing of Gait
FPRFalse Positive Rate
FRPFuzzy Recurrent Plot
GBMGradient Boosting Machine
GDGradient Descent
GDMGradient Descent with Momentum
GMGeometric Mean
GMMGaussian Mixture Model
GPGaussian Process
GRFGround Reaction Force
GRUGated Recurrent Unit
GSGraph Sequence
H&YHoehn and Yahr
HBNNHierarchical Bayesian Neural Network
HDHuntington’s Disease
hHMMHierarchical Hidden Markov Model
HMCIHealthy People with Mild Cognitive Impairment
HMMHidden Markov Model
HOAHealthy Older Adult
HOGHistogram of Oriented Gradients
IBkInstance-Based with parameter k
ICCIntraclass Correlation Coefficient
IIRInfinite Impulse Response
IMUInertial Measurement Unit
IoTInternet of Things
IPDIdiopathic Parkinson’s Disease
JMIMJoint Mutual Information Maximization
KELMKernel Extreme Learning Machine
KFDKernel Fisher Discriminant Analysis
kNNk-Nearest Neighbors
LASSOLeast Absolute Shrinkage and Selection Operator
LDALinear Discriminant Analysis
LIDLevodopa-Induced Dyskinesia
LOSOLeave One Subject Out
LRLogistic Regression
LSTMLong Short-Term Memory
MAEMean Absolute Error
MCCMatthew’s Correlation Coefficient
MDS-UPDRSMovement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale
MFCCMel-Frequency Cepstral Coefficient
MLMachine Learning
MLPMultilayer Perceptron
MMGMechanomyography
MMLMulti-source Multi-task Learning
MMTFLMultiplicative Multi-Task Feature Learning
MoCAMontreal Cognitive Assessment
MRIMagnetic Resonance Imaging
mRMRMaximum Relevance Minimum Redundancy
MS KinectMicrosoft Kinect
mUPDRSMotor subscale of the Unified Parkinson’s Disease Rating Scale
NBNaïve Bayes
NNNeural Network
OPFOptimum Path Forest
PCAPrincipal Component Analysis
PDParkinson’s Disease
PDDParkinson’s Disease Patients with Dementia
PDMCIParkinson’s Disease Patients with Mild Cognitive Impairment
PDNCParkinson’s Disease Patients with Normal Cognition
PETPositron Emission Tomography
PKGPersonal KinetiGraph
PNNProbabilistic Neural Network
PPVPositive Predictive Value
PRECPrecision
PRISMAPreferred Reporting Items for Systematic Reviews and Meta-Analyses
PSDPower Spectral Density
PSGPolysomnography
PSPProgressive Supranuclear Palsy
Q-BTDNNQ-Backpropagation for a Time Delay Neural Network
QDAQuadratic Discriminant Analysis
RBDRapid Eye Movement Sleep Behavior Disorder
RBFRadial Basis Function
RECRecall
Res-NetResidual Neural Network
RFRandom Forest
RGBRed Green Blue
RGB-DRed Green Blue-Depth
RMSERoot Mean Square Error
RNNRecurrent Neural Network
RUSBoostRandom UnderSampling Boosting
SBSSequential Backward Selection
SCGScaled Conjugate Gradient
SDESparse Difference Embedding
SDHSum and Difference Histogram
sEMGSurface-Electromyography
SENSSensitivity
SF-3636-Item Short Form Survey
SFSSequential Forward Selection
SOMSelf-Organizing Map
SPDDSSelf-Assessment Parkinson’s Disease Disability Scale
SPECSpecificity
STFTShort-Time Fourier Transform
STLSingle Task Learning
SVMSupport Vector Machine
SVRSupport Vector Regression
TRISTreatment Response Index
TUGTimed Up and Go
UDysRSUnified Dyskinesia Rating Scale
UPDRSUnified Parkinson’s Disease Rating Scale
VAEVariational Autoencoder
VaPVascular Parkinsonism
VASVisual Analog Scale
XGBoost/XGBExtreme Gradient Boosting

Appendix A

Table A1. Studies based on inertial sensors.
Table A1. Studies based on inertial sensors.
StudyProblemDataset PopulationInput DataAnalysis/AlgorithmsEvaluation
Rastegari et al. (2020) [49]Diagnosis10 PD patients and 10 healthy controls from another study [166]Accelerometer and gyroscope data from both ankles using SHIMMER sensorsSegmentation + bag-of-words features extraction based on sub-sequences clustering with k-medoid + SVM/DT/RF/kNNBest performing: DT with ACC = 90%, PREC = 90%, REC = 90%
Zhang et al. (2020) [54]Diagnosis656 PD patients and 2148 healthy controls from the mPower study [126]Gait features from smartphone sensors located at the pocketDeep CNNAUROC = 0.87
Juutinen et al. (2021) [50]Diagnosis29 PD patients and 29 healthy controlsAccelerometer and gyroscope signals from a waist-mounted smartphone while performing walking testsInterpolation + low-pass filtering + smoothening + segmentation to strides + feature extraction + feature selection with mRMR/SFS/SBS + classification of strides with DT/Gaussian kernel/LDA/ensemble/kNN/LR/NB/SVM/RF + majority voting for subject classificationBest performing: SFS + kNN with ACC = 84.5%, SENS = 88.5%, SPEC = 81.3%
Fernandes et al. (2018) [51]Parkinsonism diagnosis and differential diagnosis between idiopathic PD (IPD) and vascular parkinsonism (VaP)15 IPD, 15 VaP patients and 15 healthy controlsGait signals from wearable Physilog motion sensors placed on both feet and MOCA scoresNormalization + feature selection based on Kruskal–Wallis and Mann–Whitney tests with Bonferroni correction + features ranking with RFs + MLP/DBNBest performing for parkinsonism detection: MLP with ACC = 99.33%, SENS = 93.33%; for IPD-VaP differential diagnosis: DBN with ACC = 73.33%, SENS = 73.33%, SPEC = 73.33%
Cuzzolin et al. (2017) [52]Diagnosis and severity estimation in the SF-36 (0–100) scale156 PD patients and 424 healthy controlsSignals from an IMU attached to the lower spine while walkingHMM + kNNFor diagnosis: ACC = 85.51%, F1-SCORE = 81.54%; for severity estimation: MAE = 27.81
Borzì et al. (2020) [70]Severity estimation (MDS-UPDRS related to leg agility scores classification)93 PD patientsAcceleration, angular velocity and orientation data from a smartphone applicationLow-pass Butterworth filter + FFT + feature selection + DT/kNN/SVM/ANNBest performing: ANN with ACC = 77.7%, AUROC = 0.92, r = 0.92, RMSE = 0.42, ICC = 0.82
Hssayeni et al. (2018) [96]Fluctuation identification (on/off states detection)12 PD patients (1st dataset) and 7 PD patients (2nd dataset)Signals from an ankle-mounted triaxial gyroscopeFIR bandpass filter + LSTMACC = 73–77%, SENS = 63–75%, SPEC = 78–83%
Aich et al. (2020) [97]Fluctuation identification (on/off states detection)20 PD patientsStatistical and gait parameter features from 2 knee-worn accelerometersFeature selection + RF/SVM/kNN/NBBest performing: RF with ACC = 96.72%, PREC = 96.92%, REC = 97.35%, F1-SCORE = 0.97, AUROC = 0.99
Abujrida et al. (2020) [53]Diagnosis and severity estimation regarding walking balance (MDS-UPDRS-2.12), shaking/tremor (MDS-UPDRS-2.10) and FoG152 PD patients and 304 healthy controls from the mPower study [126]Signals from smartphone embedded accelerometers and gyroscopes, demographics and lifestyle dataSegmentation + smoothening + feature extraction in time, frequency (FFT + PSD), wavelet (DWT) domain + RF/bagged trees/boosted tress/fine tree/cubic SVM/weighted kNN/LR/LDABest performing for diagnosis: RF with ACC = 95%, PREC = 94%, AUROC = 0.99; for walking balance: RF with ACC = 93%, PREC = 92%, AUROC = 0.97; for tremor: bagged trees with ACC = 95%, PREC = 95%, AUROC = 0.92; for FoG: bagged trees with ACC = 98%, PREC = 96%, AUROC = 0.98
Kim et al. (2015) [84]FoG detection15 PD patientsGyroscopic and accelerometer data from smartphones placed at the waist, pocket and ankleAdaBoostSENS = 81.1–86%, SPEC = 91.5–92.5%
Ashour et al. (2020) [87]FoG detection10 PD patientsAccelerometer data from sensors placed on the ankles, knees and hipsPatient-dependent LSTM/{DWT/FFT + Patient-dependent SVM/ANN}Best performing: LSTM with ACC = 68.44–98.89%
Torvi et al. (2016) [89]FoG prediction in 1 s, 3 s and 5 s horizons10 PD patients from Daphnet datasetAccelerometer data from sensors placed on the ankles, the thighs and the trunkLSTM/RNN with or without transfer learningBest performing: LSTM + transfer learning with ACC = 85–95%
Arami et al. (2019) [90]FoG prediction (2-class classification FoG/no-FoG and 3-class classification pre-FoG/FoG/no-FoG)10 PD patients from Daphnet datasetAccelerometer data from sensors placed on the ankles, the thighs and the trunkWindowing + filtering + feature extraction + feature selection with mRMR/BE + features time series prediction with AR/ARMA + RBF-SVM/PNNFor 2-class classification: ACC = 94%, SENS = 93%, SPEC = 87%; for 3-class classification: ACC = 77%
Kleanthous et al. (2020) [91]FoG detection (FoG/walking/transition from walking to FoG classification)10 PD patients from Daphnet datasetAccelerometer data from sensors placed on the ankles, the thighs and the trunkLow pass Butterworth filter + Boruta algorithm + GBM + XGB for feature selection + XGB/RF/GBM/RBF-SVM/MLPBest performing: RBF-SVM with ACC = 79.85%, SENS = 72.34–91.49%, SPEC = 87.36–93.62%
Li et al. (2020) [88]FoG detection10 PD patients from Daphnet datasetAccelerometer data from sensors placed on the ankles, the thighs and the trunkFiltering + segmentation + data augmentation + CNN with squeeze-and-excitation blocks for feature extraction + attention-enhanced LSTMACC = 98.1–99.7%, SENS = 95.1–99.1%, SPEC = 98.8–99.8% for generalized and personalized models with 10-fold cross-validation; AUC = 0.945, ACC = 91.9%, EER = 10.6% with LOSO validation
Halder et al. (2021) [92]FoG states classification (pre-FoG, FoG, pre-post-FoG, no-FoG)10 PD patients from Daphnet datasetAccelerometer data from sensors placed on the ankles, the thighs and the trunkSecond-order Butterworth low-pass filtering + PCA + kNN/MLP/SVMBest performing: kNN with ACC = 98.92%, SENS = 94.97%, SPEC = 99.19%, F1-SCORE = 95.25%, PREC = 95.55%
Palmerini et al. (2017) [93]Pre-FoG detection (classification between gait and pre-FoG)11 PD patientsAccelerometer and gyroscope signals from sensors placed at the lower-back and at the anklesWindowing + feature extraction + LDASENS = 83%, SPEC = 67%, AUROC = 0.76
Borzì et al. (2021) [94]FoG and pre-FoG detection11 PD patients on and off therapyA single angular velocity signal from 2 shins-mounted inertial sensors while performing TUG testsNormalization + segmentation + wrapper feature selection + SVM/kNN/LDA/LRBest performing: SVM with ACC = 85.5–86.1%, SENS = 84.1–85.5%, SPEC = 85.9–86.3%, F-SCORE = 73.4–74.6%
Shi et al. (2020) [85]FoG detection63 PD patientsAccelerometer, gyroscope and magnetometer signals from IMUs placed on both ankles and the spine while performing TUG testsTime-series segmentation with overlapping windows + Morlet CWT/FFT/raw data + 1D-CNN/2D-CNN/LSTMBest performing: CWT + 2D-CNN with ACC = 89.2%, SENS = 82.1%, SPEC = 96%, GM = 88.8%
Camps et al. (2018) [86]FoG detection21 PD patientsSignals from a waist-mounted IMU with accelerometer, gyroscope and magnetometer while performing several walking tests and ADLsSpectral window stacking (with FFT) + RUSBoost/SVM-RBF/CNNBest performing: CNN with ACC = 89%, SENS = 91.9%, SPEC = 89.5% and geometrical mean of SENS-SPEC = 90.6%
Ghassemi et al. (2018) [95]Gait segmentation (strides detection)10 PD patients for the 1st experiment and 34 PD patients for the 2nd experimentAcceleration and angular velocity signals from foot-worn IMUs while walking in straight line (1st experiment) and walking, turning or performing other movements (2nd experiment)Peak detection algorithm/Euclidean DTW/Probabilistic DTW/hierarchical HMMBest performing for the 1st experiment: all with F-SCORE = 99.8–100%;
For the 2nd experiment: hHMM with PREC = 98.5%, REC = 93.5%, F-SCORE = 95.9%
Kostikis et al. (2015) [55]Diagnosis25 PD patients and 20 healthy controlsTremor measurements from accelerometer and gyroscope smartphone sensorsNB/LR/SVM/AdaBoost/C4.5/RFBest performing: RF with SENS = 82%, SPEC = 90%, AUROC = 0.94
Williamson et al. (2021) [59]Diagnosis202 PD patients and 178 healthy controls from the UK Biobank dataset (https://www.ukbiobank.ac.uk/, accessed on 17 February 2022)Acceleration signals from a wrist-worn sensorSegmentation + automatic segments labelling (gait or low movement) + segmentation into frames + features extraction + GMMSENS = 65–75%, AUROC = 0.85
Park et al. (2021) [60]Diagnosis25 PD patients and 21 healthy controlsSignals from IMUs attached to the thumb and index fingers while performing finger tapping, hand movements and rapid altering movementsLinear regression + correlation between motor parameters and UPDRS scores + DNN/LR for PD diagnosisFor motor parameters-UPDRS scores correlation: r = 0.838–0.849; Best performing for diagnosis: DNN with AUROC = 0.888–0.950
Talitckii et al. (2021) [65]Differential diagnosis between PD and other extrapyramidal disorders41 PD patients and 15 patients with other extrapyramidal disordersAccelerometer, gyroscope and magnetometer signals form a dorsal-mounted sensor while performing UPDRS-related tasksSTFT + feature extraction + linear-PCA/RBF-PCA/poly-PCA/LDA/FA + RF/SVM/LR/LightGBM/stacked ensemble modelBest performing: standard classifier with ACC = 72–85%, PREC = 72–85%, REC = 77–100%, F1-SCORE = 76–88%
Varghese et al. (2020) [66]PD or other movement disorders (DD) patients and healthy controls classification; PD patients and DD patients or healthy controls classification192 PD patients, 75 DD patients and 51 healthy controlsAccelerometer data from 2 smartwatches placed at both hands and answers from electronic questionnaires distributed via smartphonesFFT + PCA + RBF-SVM/RF/ANNBest performing for PD/DD detection: ANN with ACC = 89%, PREC = 94%, REC = 92%, F1-SCORE = 93%; for PD detection: RBF-SVM with ACC = 79%, PREC = 81%, REC = 89%, F1-SCORE = 85%
Loaiza Duque et al. (2020) [67]Healthy and trembling subjects classification; PD-ET differential diagnosis19 PD patients, 20 ET patients and 12 healthy controlsAngular velocity signals from smartphone built-in triaxial gyroscope with the help of a smartphone applicationKinematic features extraction + feature selection based on chi-square and unbiased tree + linear-SVM/LR/DA/NB/DT/ensemble subspace kNNBest performing for trembling patients detection: ensemble subspace kNN with ACC = 97.2%, SENS = 98.5%, SPEC = 93.3%; for PD-ET differential diagnosis: linear-SVM with ACC = 77.8%, SENS = 75.7%, SPEC = 80%
Channa et al. (2021) [83]Classification between PD patients with tremor and PD patients with bradykinesia and healthy controls10 PD patients with tremor, 10 PD patients with bradykinesia and 20 healthy controlsAccelerometer and gyroscope signals from a smart braceletButterworth bandpass IIR filter + FFT to extract both time and frequency domain features + feature selection based on ANOVA test + NN-SOMs clustering/kNNBest performing: kNN with ACC = 91.7%, SENS = 83–100%, SPEC = 89–100%
Li et al. (2019) [57]Diagnosis and severity (H&Y scores) estimation13 PD patients and 12 healthy controlsAcceleration and angular velocity signals from a prototype handle for spoons with embedded inertial sensorsIIR filtering + windowing + feature extraction + normalization + kNN/Adaboost/RF/linear-SVM for diagnosis; RF regression for severity estimationBest performing for diagnosis: linear-SVM with ACC = 92%, SENS = 92.31%, SPEC = 91.67%, AUROC = 0.98; for severity estimation: MAE = 0.166, r = 0.97
Koçer et al. (2016) [56]Diagnosis and severity (H&Y scores) estimation35 PD patients and 20 healthy controlsResting tremor acceleration signals from the Nintendo Wii Remote (Wiimote)Windowing + feature extraction (with FFT) + SVMFor diagnosis: ACC = 89%, PREC = 91%, REC = 94%; for severity estimation: ACC = 33–77%
Bazgir et al. (2015) [71]Severity estimation (UPDRS scores classification)52 PD patientsAccelerometer and gyroscope data from a smartphone placed at the wristFiltering + STFT + MLP trained with the back propagation algorithmACC = 91%, SENS = 89.6%, SPEC = 90.64%
Kim et al. (2018) [72]Severity estimation (UPDRS scores classification)92 PD patientsAccelerometer and gyroscope signals from a wrist-worn deviceHigh pass filter + FFT + RF/NB/linear regression/DT/MLP/SVM/CNNBest performing: CNN with ACC = 85%, kappa = 0.85, r = 0.93, RMSE = 0.35
Dai et al. (2021) [73]Severity estimation (MDS-UPDRS scores classification)42 PD patients and 30 healthy controlsAccelerometer, gyroscope and geomagnetic data from a finger-mounted sensor while measuring rest and postural tremor and during finger tappingDenoising with an IIR bandpass filter + FFT + SVM/RF/kNNBest performing: SVM with ACC = 96–97.33%, SENS = 96.36–100%, SPEC = 95–96.67%
Khodakarami et al. (2019) [74]Severity estimation (UPDRS scores classification) and prediction of response to levodopa (absolute value and percentage)151 PD patients and 174 healthy controlsSignals from the wrist-worn smartwatch of the Parkinson’s Kinectigraph systemFeature extraction + JMIM-based feature selection + (+PCA) LR/RBF-SVM/gradient boosting DTsBest performing for severity estimation: LR with AUROC = 0.79–0.88, AUPR = 0.65–0.88; for absolute levodopa response estimation AUROC = 0.92 and AUPR = 0.87; for levodopa response percentage estimation AUROC = 0.82, AUPR = 0.73
Javed et al. (2018) [58]Diagnosis and treatment response index (TRIS) estimation19 PD patients and 22 healthy controlsAccelerometer and gyroscope data from both wrists using SHIMMER3 sensors while performing hand rotation tests before and after the dose administrationPCA/stepwise regression/LASSO regression + SVM/linear regression/DT/RFBest performing for diagnosis: stepwise regression + SVM with ACC = 89%; for TRIS estimation: RMSE = 0.69, r = 0.84
Watts et al. (2021) [100]PD patients classification according to their levodopa regimens and response26 PD patientsBradykinesia and dyskinesia-related signals from Personal KinetiGraph (PKG), demographics and MDS-UPDRS-III scoresk-means for clustering based on regimen features + RF for classification based on PKG features, demographics and MDS-UPDRS-III scoresACC = 86.9%, SENS = 86.5%, SPEC = 87.7%, PPV = 95.3%, F1-SCORE = 90.7%, AUROC = 0.871
Pfister et al. (2020) [98]On/Off/Dyskinesia motor states classification30 PD patientsData from a wrist-worn accelerometer in a free-living environmentData augmentation + CNN/SVM/kNN/RF/MLPBest performing: CNN with ACC = 65.4%, Kohen’s Kappa = 0.47, SENS = 64.45–66.68%, SPEC = 66.72–89.48%, F1-SCORE = 62.4–69.01%, 1vsALL ACC = 66.7–82.56%
Eskofier et al. (2016) [79]Bradykinesia detection10 PD patientsAccelerometer data from IMUs mounted on both handsTimeseries segmentation + AdaBoost/PART/kNN/SVM/CNN-DNNBest performing: DNN
ACC = 90.9%
Shawen et al. (2020) [80]Tremor and bradykinesia detection; severity estimation (UPDRS scores classification)13 PD patientsAccelerometer and gyroscope signals from a flexible skin-mounted sensors and accelerometer signals from a wrist-worn smartwatchCubic spline interpolation + segmentation + high-pass filtering + time, frequency, entropy, correlation and derivative-based feature extraction + RFFor tremor detection: AUROC = 0.68–0.79; for bradykinesia detection: AUROC = 0.61–0.69; for tremor severity estimation: AUROC = 0.67–0.77; for bradykinesia severity estimation: AUROC = 0.59–0.66
San-Segundo et al. (2020) [81]Tremor detection; tremor duration estimation12 PD patients for laboratory set and 6 PD patients for in-the-wild setAccelerometer signals from wrist-worn sensorsDownsampling + FFT + unsupervised non-negative tremor factorization + feature extraction manually/with a CNN + RF/MLPBest performing for tremor detection: CNN + MLP with
AUC = 0.887; for tremor duration estimation: MAE = 4.1–9.1%
Ibrahim et al. (2020) [82]Tremor onset detection13 PD patientsSignals from hand-mounted IMUs while performing 6 different rest, postural and motor tasksButterworth filtering + zero-phase shifting + Hilbert-Huang Transform + MLPACC = 92.9%, PREC = 98.7%, REC = 86.7%, SPEC = 98.9%, F1-SCORE = 0.923
Som et al. (2020) [61]Diagnosis152 healthy subjects for pre-training and 18 PD patients and 16 healthy controls for the final classificationAccelerometers signals from a wrist-worn sensor while performing various ADL for the pre-training + accelerometer signals from 6 sensors located at the sternum, the lumbar, both ankles and wrists while walkingZero-centering + normalization + segmentation + feature extraction with pre-trained convolutional AE + PCA/global-average-pool layer + SVM/MLPBest performing: MLP with ACC = 68.64–73.81%, PREC = 69.27–76.53%, REC = 68.64–73.81%, F1-SCORE = 67.65–73.89%
Ricci et al. (2020) [62]Diagnosis30 newly diagnosed untreated PD patients and 30 healthy controlsAcceleration, angular velocity and orientation signals from a network of 14 IMUs distributed in the whole bodyFeature selection with ReliefF ranking and Kruskal–Wallis + NB/kNN/SVMBest performing: SVM with ACC = 95%, PREC = 95.1%, AUROC = 0.95
De Vos et al. (2020) [69]PD and PSP differential diagnosis20 PD patients and 21 PSP patientsAccelerometer, gyroscope and magnetometer signals from 6 IMUs placed on the lumbar spine, the sternum, both wrists and feetANOVA + LASSO + LR/RFBest performing: RF with ACC = 88%, SENS = 86%, SPEC = 90%
Moon et al. (2020) [68]PD and ET differential diagnosis524 PD patients and 34 ET patientsBalance and gait characteristics from 6 IMUs placed on the lumbar spine, the sternum, both wrists and feetNN/SVM/kNN/DT/RF/LRBest performing: NN with ACC = 89%, PREC = 61%, REC = 61%, F1-SCORE = 61%
Kuhner et al. (2017) [63]Diagnosis; correlation with severity metrics14 PD patients and 26 healthy controlsFusion of accelerometer, gyroscope and magnetometer signals from XSens motion capture suit while performing several motor testsRF with probability distributions for classification
PCA for correlation
For diagnosis: ACC = 86–94.6%, SENS up to 91.5% and SPEC up to 97.2%; for correlation between the 1st pc and the UPDRS scores: r = 0.79
Caramia et al. (2018) [64]Diagnosis and severity estimation (H&Y scores classification)25 PD patients and 25 healthy controlsAccelerometer, gyroscope and magnetometer signals from 8 IMUs attached to both feet dorsum, thighs, shanks and to the chest and lumbarExtraction of range of motion parameters/spatio-temporal parameters (+PCA) + LDA/NB/kNN/linear-SVM/RBF-SVM/DT + majority voting with equal weights/weights analogue to the individual accuraciesBest performing for diagnosis: majority voting with weights analogue to the individual accuracies with ACC = 96%; for severity estimation: RBF-SVM with ACC = 87.75–94.5%
Hssayeni et al. (2021) [76]Severity (UPDRS-III scores) estimation24 PD patientsAngular velocity from one wrist-mounted and one ankle-mounted inertial sensorGradient tree boosting/dual-channel LSTM with hand-crafted features and with or without transfer learning/1D-CNN-LSTM for raw signals/2D-CNN-LSTM for time-frequency data + ensemble learningBest performing: ensemble of dual-channel LSTM with hand-crafted features and transfer learning, 1D-CNN-LSTM for raw signals and 2D-CNN-LSTM for time-frequency data with r = 0.79, MAE = 5.95
Butt et al. (2020) [77]Severity (MDS-UPDRS III scores) estimation64 PD patients and 50 healthy controlsGyroscope and geomagnetic data from 4 wrist-mounted IMUs and 1 foot-mounted IMUKolmogorov–Smirnov test + Mann–Whitney U-test + normalization + CFS/PCA ranker/correlation attribute evaluation/chi-square attribute evaluation/wrapper subset evaluation + SVR/RF/LR/ANFISBest performing: CFS + ANFIS with r = 0.814, RMSE = 0.101
Mirelman et al. (2021) [75]Severity estimation (H&Y stages classification)332 PD patients and 100 healthy controlsAccelerometer and gyroscope signals from sensors placed on the ankles, the wrists and the lower back while performing various walking tests + demographic dataLow-pass Butterworth filter + feature selection based on RF permutation importance/neighborhood component analysis/mRMR + RUSBoost + DT/QDA for weak learnerSENS = 72–84%, SPEC = 69–80%, AUROC = 0.76–0.90
Stamate et al. (2018) [78]Identification of failures to follow the UPDRS-III movement protocol12 PD patientsMotor signals from smartphone sensors with the cloudUPDRS applicationFiltering + frequency transformations + extra tree/Bernoulli NB/Gaussian NB/MLP/RF/Gradient Boosting/Bagging/AdaBoost/RCNNACC = 78%, F1-SCORE = 82%, AUROC = 0.87
Belgiovine et al. (2018) [99]L-dopa induced dyskinesia detection18 PD patientsAccelerometer and angular velocity data from smartphone placed on the wrist (for upper-limb experiment) or on the hip (for lower-limb experiment)z-score normalization + DT/Gaussian-SVM/linear-SVMBest performing: SVM (with both kernels) with ACC = 65.0–82.0%, MACRO F1-SCORE = 0.65–0.82
Table A2. Studies based on pressure and force sensors.
Table A2. Studies based on pressure and force sensors.
StudyProblemDataset PopulationInput DataAnalysis/AlgorithmsEvaluation
Aversano et al. (2020) [102]Diagnosis and severity estimation (H&Y scores classification)93 PD patients and 73 healthy controls from 3 merged datasets [106]Vertical GRF signals from 8 sensors located underneath each footFeed-forward DNNFor diagnosis: ACC = 99.29–99.52%; for severity estimation: ACC = 98.57–99.1%
El Maachi et al. (2020) [103]Diagnosis and severity estimation (H&Y scores classification)93 PD patients and 73 healthy controls from 3 merged datasets [106]Vertical GRF signals from 8 sensors located underneath each footGait segmentation + 1D-CNN/DNN/MLP/NB/RFBest performing for diagnosis: 1D-CNN with ACC = 98.7%, SENS = 98.1%, SPEC = 100%; for severity estimation: ACC = 85.3%, PREC = 87.3%, REC = 85.3%, F1-SCORE = 85.3%
Xia et al. (2020) [104]Diagnosis and severity estimation (H&Y scores classification)93 PD patients and 73 healthy controls from 3 merged datasets [106]Vertical GRF signals from 8 sensors located underneath each footDual-modal CNN + attention-enhanced LSTM (DCALSTM)/baseline DL models removing one of the previous stages (DCNN/DALSTM/DCLSTM/CNN-LSTM)/feature-based modelsBest performing for diagnosis: DCALSTM with ACC = 99.07–99.31%, SENS = 99.10–99.35%, SPEC = 98.98–99.35%; for severity estimation: ACC = 98.03–99.01%
Nancy Jane et al. (2016) [101]Severity estimation (H&Y scores classification)93 PD patients and 73 healthy controls from 3 merged datasets [106]Vertical GRF signals from 8 sensors located underneath each footQ-backpropagation/Levenberg–Marquardt/GD/GDM/SCG for a time delay NNBest performing: Q-BTDNN with ACC = 90.91–92.19%
Balaji et al. (2021) [105]Severity estimation (H&Y scores classification)93 PD patients and 73 healthy controls from 3 merged datasets [106]Vertical GRF signals from 8 sensors located underneath each footNormalization + spatiotemporal feature extraction + Shapiro–Wilk test + feature selection + kNN/NB/Bagging classifier/SVMBest performing: SVM with ACC = 98.8%, SENS = 96.6%, SPEC = 99.6%, PPV = 99.1%, FPR = 3.4%, PREC = 99.1%, F-SCORE = 97.8%, MCC = 0.98
Papavasileiou et al. (2017) [107]Differential diagnosis between PD, post-stroke patients and healthy controls5 PD patients, 3 post-stroke patients and 3 healthy controlsGround contact force data from barometric pressure sensors placed on both feetMultiplicative multi-task feature learning (MMTFL)/single task learning (STL)Best performing: MMTFL
AUROC = 0.880–0.994
Khoury et al. (2019) [108]Diagnosis and differential diagnosis between PD, HD, ALS and healthy controls93 PD patients and 73 healthy controls from 3 merged datasets [106] for diagnosis; 15 PD patients, 20 HD patients, 13 ALS patients and 16 healthy controls for differential diagnosisVertical GRF signals from 8 sensors located underneath each footFeature selection with RF-based wrapper method + kNN/DT/RF/NB/SVM/k-means/GMMBest performing for diagnosis: kNN/RF/SVM with ACC = 81.25–90.91%, PREC = 81.43–89.41%, REC = 71.48–88.35%, F-SCORE = 79.45–86.83%; for PD vs. ALL differential diagnosis: kNN with ACC = 90%, PREC = 90.18%, REC = 90%, F-SCORE = 90.09%
Table A3. Studies based on image and depth sensors.
Table A3. Studies based on image and depth sensors.
StudyProblemDataset PopulationInput DataAnalysis/AlgorithmsEvaluation
Guayacán et al. (2020) [111]Diagnosis11 PD patients and 11 healthy controlsVideo recordings while walking3D spatio-temporal CNNACC = 88–90%
Reyes et al. (2019) [109]Diagnosis88 PD patients and 94 healthy controlsGait samples from MS KinectCropping noisy parts + LSTM/1D-CNN/CNN-LSTMBest performing: CNN-LSTM with ACC = 83.1%, PREC = 83.5%, REC = 83.4%, F1-SCORE = 81%, Kappa = 64%
Buongiorno (2019) [110]Diagnosis and severity estimation (mild vs. moderate)16 PD patients and 14 healthy controlsPostural and kinematics features from MS Kinect v2 sensor while performing 3 motor exercises (gait, finger and foot tapping)Feature selection + SVM/ANNBest performing for diagnosis: gait-based ANN with ACC = 89.4%, SENS = 87.0%, SPEC = 91.8%; for severity estimation: ACC = 95.0%, SENS = 90.0%, SPEC = 99.0%
Grammatikopoulou et al. (2019) [117]Severity estimation (UPDRS scores classification)12 advanced PD patients and 6 PD patients in initial stageSkeletal features from MS Kinect v2 RGB videos while playing an exergameTransformation to local coordinate system + two parallel LSTMs (the 1st trained with raw joint coordinates and the 2nd with joint line distances)ACC = 77.7%
Tucker et al. (2015) [122]Medication adherence estimation (on/off medication classification)7 PD patientsSkeletal joints 3D position, velocity and acceleration from MS KinectC4.5 DT for generalized model; C4.5 DT, RF, SVM, IBk for personalized modelsfor generalized model: ACC = 36.2–77.9%; for personalized models: ACC = 67.7–100%
Li et al. (2018) [120]TUG subtasks segmentation and time estimation for each subtask24 PD patientsVideo recordings while performing TUG testsPose estimation with OpenPose/Iterator Error Feedback + SVM/LSTMBest performing: OpenPose + LSTM with ACC = 93.1%, PREC = 80.8–97.5%, REC = 86.3–97%, F1-SCORE = 83.5–97.3% for subtasks segmentation and MAE = 0.32–1.07 for time estimation
Wei et al. (2019) [123]Development of a virtual physical therapist: movement recognition (repetitions and sub-actions detection), patient’s errors identification (satisfactory/non-satisfactory performance), task recommendation (regress/repeat/progress)35 PD patientsMotion data recorded by MS Kinect v2 sensor while performing 3 balance/agility tasksHMM for repetitions and sub-actions detection + linear-SVM for movement errors identification + {majority undersampling/minority oversampling/decision threshold adjustment/hybrid oversampling with feature standardization and interpolation + RF} for task recommendationFor repetitions detection: ACC = 97.1–99.4%; for sub-actions segmentation: SENS = 88.4–96.9%, SPEC = 97.2–98.8%; for errors identification: ACC = 86.3–94.2%; best performing for tasks recommendation: hybrid oversampling + RF with ACC = 81.8–95.7%, FPR = 2.8–5.4%
Hu et al. (2019) [121]FoG detection45 PD patientsVideos collected while performing TUG testsGraph representation of videos + pretrained features (Res-Net 50 vertex, C3D vertex and context features) + graph sequence-RNN (Bi-directional GS-GRU/Bi-directional GS-LSTM/forward GS-GRU/forward GS-LSTM) + fusionBest performing: linear fusion of Bi-directional GS-GRU with context model with AUC = 0.90, SENS = 83.8%, SPEC = 82.3%, ACC = 82.5%, Youden’s J = 0.66, FPR = 17.7%, FNR = 16.2%
Li et al. (2018) [118]Binary classification between pathological (PD/LID) and normal motion; multiclass classification (PD with LID, PD without LID and normal); levodopa-induced dyskinesia severity (UDysRS-III scores) estimation; parkinsonism severity (UPDRS-III scores) estimation9 PD patients2D-Video recordings while performing communication and drinking tasks (for dyskinesia detection) and while performing leg agility and toe tapping tasks (for parkinsonism detection)Convolutional pose estimators + RFFor binary classification: AUC = 0.634–0.930, F1-SCORE = 50–90.6%; for multiclass classification: ACC = 71.4%, SENS = 83.5–96.2%, SPEC = 68.4–95.7%; for UDysRS-III estimation: RMSE = 2.906, r = 0.741; for UPDRS-III estimation: RMSE = 7.765, r = 0.53
Vivar-Estudillo et al. (2018) [112]Diagnosis18 PD patients and 22 healthy controlsPosition, velocity and rotation data regarding hand movements from leap motion sensorTexture features extraction with SDH + kNN/SVM/DT/LDA/LR/ensemblesBest performing: bagged tree with ACC = 98.62%, SENS = 98.43%, SPEC = 98.80%, PREC = 98.80%
Moshkova et al. (2020) [113]Diagnosis16 PD patients and 16 healthy controlsSignals from leap motion sensor while performing hand motor tasks according to the MDS-UPDRS-IIIFeatures extraction + kNN/SVM/DT/RFBest performing: SVM with ACC = 98.4% when features are extracted from all the tasks
Ali et al. (2020) [114]Diagnosis; classification between PD patients with medication, without medication and healthy controls87 PD patients with medication, 119 PD patients without medication and 139 healthy controlsVideos while performing hand motor tasksSegmentation to frames + temporal segmentation with CNN + spatial segmentation with CNN-AE + FFT for feature extraction + SVMBest performance when combining 2 tasks for diagnosis: ACC = 91.8%; for 3-class classification: ACC = 73.5%
Liu et al. (2019) [119]Severity estimation (Bradykinesia-related MDS-UPDRS scores classification)60 PD patientsVideo recordings while performing hand motor testsPose estimator NN + feature extraction + kNN/RF/linear-SVM/RBF-SVMBest performing: RBF-SVM with ACC = 89.7%, PREC = 20–100%, REC = 60–100%, F1-SCORE = 33.3–100%
Rajnoha et al. (2018) [116]Diagnosis50 PD patients and 50 age-matched healthy controlsFace images extracted from video recordingsHOG for face detection + CNN for embeddings generation + kNN/DT/RF/XGBoost/SVMBest performing: DT:ACC = 67.33% with leave-one-out cross validation, RF: ACC = 60.7–85.92% with train-test split
Jin et al. (2020) [115]Diagnosis33 PD patients and 31 elderly healthy subjectsShort videos while imitating images of smiley peopleSplitting videos to frames + coordinate points extraction with Face++ + transformation from absolute to relative coordinates + features extraction + LASSO + LR/SVM/DT/RF/LSTM/RNNBest performing: SVM with PREC = 99%, REC = 99%, F1-SCORE = 99%
Table A4. Studies based on voice or audio sensors.
Table A4. Studies based on voice or audio sensors.
StudyProblemDataset PopulationInput DataAnalysis/AlgorithmsEvaluation
Zhang et al. (2017) [127]Diagnosis23 PD patients and 4 healthy controls from the Oxford’s dataset (https://archive.ics.uci.edu/ml/datasets/parkinsons+telemonitoring, accessed on 17 February 2022); 20 PD patients and 20 healthy controls from the Istanbul dataset (https://archive.ics.uci.edu/ml/datasets/Parkinson%27s+Disease+Classification, accessed on 17 February 2022)Vocal measurements from a smartphone applicationStacked AEs + KELM/linear-SVM/MLP-SVM/RBF-SVM/kNN/NB/CART/LDABest performing: kNN with ACC = 94–98%
Zhang et al. (2018) [124]Diagnosis500 PD patients and 500 healthy controls from the mPower study [126]Vocal measurements from a smartphone applicationDT-STFT + LSTM/CNNBest performing: CNN with ACC = 90.45%
Tougui et al. (2020) [125]Diagnosis453 PD patients and 1037 healthy controls from the mPower study [126]Vocal measurements from a smartphone applicationTime, frequency and cepstral domain (with DFT) features + feature selection with ANOVA/LASSO + linear-SVM/kNN/RF/XGBoostBest performing: LASSO + XGBoost with ACC = 95.78%, SENS = 95.32%, SPEC = 96.23%, F1-SCORE = 95.74%
Almeida et al. (2019) [128]Diagnosis64 PD patients and 35 healthy controlsAudio recordings from acoustic cardioid and smartphonePhonation/speech/unvoiced/voiced features + kNN/MLP/OPF/SVMBest performing: kNN based on phonation features with ACC = 92.94–94.55%, SENS = 92.94–94.55%, SPEC = 89.21–94.26%, AUROC = 0.87–0.92
Arora et al. (2021) [129]Differential diagnosis between PD patients, patients with RBD and healthy controls; severity (MDS-UPDRS, MoCA, ESS, BDI and VAS scores) estimation335 PD patients, 112 patients with RBD and 92 healthy controlsSpeech recording from smartphonesSegmentation + feature extraction + feature selection + RFFor all the pairwise classifications: SENS = 59.4–74.9%, SPEC = 67.4–73.2%; for severity estimation: MAE = 1–8 (MDS-UPDRS), MAE = 1–14 (MDS-UPDRS I-III), MAE = 1–2 (MoCA), MAE = 2–3 (ESS), MAE = 1–5 (BDI), MAE = 6.5–10 (VAS)
Bayestehtashk et al. (2015) [132]Severity (mUPDRS 0–108 scores) estimation168 PD patientsSpeech recordings from a portable deviceFeature extraction with harmonic model + Ridge/LASSO regression/linear-SVRBest performing: ridge regression with MAE = 5.5, explained variance = 61%
Yoon et al. (2019) [130]Severity (UPDRS 0–176 scores) estimation42 PD patientsPhonation and speech recordings from an at-home testing deviceStandardization + feature extraction + {one-model-fits-all approaches with DT/GP/linear regression/SVR/ensemble}/{single learning approaches with DT/GP/linear regression/SVR/ensemble}/positive transfer learning based on Bayesian parameter transfer modelBest performing: positive transfer learning with MAE = 2–3 approximately
Raza et al. (2021) [131]Severity (UPDRS and mUPDRS scores) prediction in 6 months42 PD patientsPhonation and speech recordings from an at-home testing deviceFeature extraction + XGBoostMAE = 6.45 (UPDRS), MAE = 5.09 (mUPDRS)
Table A5. Studies based on other types of sensors.
Table A5. Studies based on other types of sensors.
StudyProblemDataset PopulationInput DataAnalysis/AlgorithmsEvaluation
Rahman et al. (2020) [133]Diagnosis5 PD patients and 5 healthy controlsEEG signals from a portable headset with sensors placed at the forehead while watching 4 videos which provoke 4 different emotionsFeed-forward NN trained with Adam optimization algorithmACC = 96.5%, PREC = 95.5%, REC = 97%, F1-SCORE = 97.6%
Kleinholdermann et al. (2021) [134]Severity (MDS-UPDRS III scores) estimation45 PD patientssEMG signals from a wrist-worn band while performing a simple tapping activityWindowing + feature extraction + linear regression/poly-SVM/kNN/RFBest performing: RF regression with r = 0.739
Capecci et al. (2019) [135]Emotion (positive/negative) recognition36 PD patientsBody temperature, heart rate and galvanic response from smartwatch sensorsLinear-SVM/poly-SVM/RBF-SVMBest performing: RBF-SVM with ACC = 88.6–91.3%
Lacy et al. (2018) [136]Diagnosis49 PD patients and 41 healthy controls (1st dataset); 58 PD patients and 29 healthy controls (2nd dataset)Position measures from 2 electromagnetic sensors located at the thumb and index finger while performing finger-tapping testsLow-pass Butterworth filtering + velocity and acceleration features extraction from derivatives + ESNAUROC = 0.802
Picardi et al. (2017) [137]Diagnosis; classification between different cognition levels (PD patients with normal cognition-PDNC, PD patients with mild cognitive impairment-PDMCI and PD patients with dementia-PDD)22 PDNC, 23 PDMCI, 10 PDD and 30 age-matched healthy controlsFlexion signals from a glove with finger-mounted sensors and position and orientation information from a wrist-worn tracking systemFeature extraction + Cartesian Genetic Programming/SVM/ANNSimilar performance for all the algorithms: AUROC = 0.72–0.99 for all the pair-wise classifications
Memedi et al. (2015) [138]Symptom detection (bradykinesia/dyskinesia)65 advanced PD patientsSpatiotemporal features from spiral drawings, produced with a touchscreen telemetry devicePCA + MLP/RF/RBF-SVM/linear-SVM/LRBest performing: MLP with ACC = 84.0%, SENS = 75.7%, SPEC = 88.9%, AUROC = 0.86, weighted Kappa = 0.65
Pham et al. (2019) [139]Diagnosis42 PD patients and 43 healthy controls from newQWERTY MIT-CSXPD database (https://www.physionet.org/content/nqmitcsxpd/1.0.0/, accessed on 17 February 2022)keystroke logs time seriesCNN/LSTM/CNN-GoogleNet/CNN-AlexNet/LSTM-fuzzy recurrent plots (FRP)Best performing: LSTM-FRP (m = 3) with ACC = 65.14–81.90%, SENS = 66.67–95%, SPEC = 63.33–66.67%
Matarazzo et al. (2019) [140]Medication response detection (improved/not changed) and medication response prediction in 21 weeks29 PD patients and 30 age-matched healthy controlsKeystroke logs with the help of neuroQWERTY softwareRNNFor medication response detection: ACC = 76.5%, AUROC = 0.75, kappa = 0.47; for medication response prediction: AUROC = 0.69–0.75
Table A6. Studies based on the combination of previous types of sensors.
Table A6. Studies based on the combination of previous types of sensors.
StudyProblemDataset PopulationInput DataAnalysis/AlgorithmsEvaluation
Aharonson et al. (2018) [141]Diagnosis22 PD patients and 20 healthy controlsSignals from sensors mounted on a support walker (2 encoders on the wheels, 2 force sensors underneath the hand grips and a tri-axial accelerometer) while performing two walking testsFiltering + wavelet denoising for accelerometer signals + differentiation for encoder signals + ranked feature selection/PCA + k-means/FLDABest performing: PCA + FLDA with SENS = 91–96%, SPEC = 95–100%
Pardoel et al. (2021) [142]FoG detection (pre-fog, transition, FoG, total-FoG, no-FoG events classification)11 PD patientsAccelerometer and gyroscope signals from 4 foot-mounted sensors and plantar pressure distribution data from in-sole sensorsData windowing + feature extraction in time and frequency domain with FFT and DWT + feature selection with mRMR/Relief-f + RUSBoostFor total-FoG/no-FoG classification: SENS = 61.9–78%, SPEC = 83.2–91.6%; for FoG/no-FoG classification: SENS = 81.4–98.5%, SPEC = 83.2–91.6%
Wu et al. (2020) [143]Severity estimation (UPDRS-III scores classification)17 PD patientsAcceleration signals from hand-mounted sensors and displacement signals from detection devicesDetrending + Wavelet transform + PCA + linear regression/SVM/NN/RFBest performing: NN with ACC = 91.18–95.30%
Cole et al. (2014) [144]Tremor and dyskinesia detection; severity estimation for tremor and dyskinesia8 PD patients and 4 healthy controlsAcceleration and electromyography signalsDynamical DNN/SVM/HMM for tremor and dyskinesia detection; Bayesian maximum likelihood classifier for severity estimationBest performing for tremor detection: HMM with global error = 6.1%; for dyskinesia detection: DNN with global error = 8.8%; for tremor severity estimation: SENS = 95.2–97.2%, SPEC = 97.1–99.3%; for dyskinesia severity estimation: SENS = 91.9–95%, SPEC = 94.6–98.6%
Hossen et al. (2012) [145]Differential diagnosis between PD and ET patients39 PD patients and 41 ET patientsSignals from a hand-mounted accelerometer and 2 sEMG sensors placed at the forearm flexors and extensorsFiltering + feature extraction with SDE based on wavelet decomposition + MLP trained with the back-propagation algorithmACC = 91.6%, SENS = 95%, SPEC = 88.2%
Tahafchi et al. (2019) [146]FoG detection4 PD patientsAccelerometer and gyroscope signals from 2 foot-mounted IMUs and EMG signals from 2 Shimmer modulesFully connected NNAUC = 0.906–0.963
Huo et al. (2020) [147]Diagnosis; Severity (UPDRS scores) estimation; PD patients with UPDRS > 0 and PD patients with UPDRS = 0 (due to DBS) classification23 PD patients and 10 healthy controlsBio-signals from hand-placed force sensor, 3 IMUs and 4 MMG sensors during different symptoms measurements (elbow rigidity, wrist rigidity, bradykinesia, kinetic tremor, postural tremor, rest tremor)Voting classifier of 3 best performing basic classifiers (kNN, MLP and AdaBoost)For severity estimation: ACC = 80.1–91.8% (avg = 85.4%); for diagnosis: ACC = 95.0–98.9% (avg = 96.6%); for PD patients with UPDRS > 0 and UPDRS = 0 classification: ACC = 85.2–91.1% (avg = 89.0%)
Yu et al. (2018) [148]Severity estimation; Fall risk detection22 PD patientsSignals from accelerometers, gyroscopes and thermometers embedded in 5 sensors attached to the chest, both thighs and feet while performing TUG testsCNN for multi-source multi-task learning/single feature-based assessment + kNN/SVM/NBBest performing for severity estimation: CNN for MML with RMSE = 0.060; for fall risk detection: PREC = 92.5%, REC = 95.8%, F-SCORE = 94%
Sajal et al. (2020) [149]Diagnosis and severity estimation (voice and rest tremor MDS-UPDRS scores classification)52 PD patients for tremor measurements; 23 PD patients and 8 healthy subjects for voice measurements (https://archive.ics.uci.edu/ml/datasets/parkinsons, accessed on 17 February 2022)Rest tremor data from a smartphone built-in accelerometer and vowel phonation recordings from a smartphoneFor tremor data: detrending + wavelet filtering
For vocal data: bandpass filter + down-sampling
+ mRMR feature selection algorithm + kNN/SVM/NB
+ majority voting
Best performing for severity estimation: kNN with ACC = 93.7%, SENS = 94.6%, SPEC = 93.7% (vocal features), ACC = 90.5–98.5%, SENS = 87.5–94%, SPEC = 96–100% (tremor features); for diagnosis: kNN + SVM + NB ensemble averaging with ACC = 99.8%
Oung et al. (2018) [150]Severity estimation (healthy/mild/moderate/severe classification)15 healthy controls, 20 PD patients with mild severity, 20 PD patients with moderate severity, 15 PD patients with severe symptomsAccelerometer, gyroscope and magnetometer data from 4 wrist- and limb-mounted IMUs and speech signals recorded with a headsetSegmentation + EWT for motor signals/EWPT for speech signals + Hilbert transformations + wavelet energy and entropy-based feature extraction + kNN/PNN/ELMBest performing: ELM with ACC = 92.45–95.93%
Papadopoulos et al. (2020) [151]Diagnosis; tremor/fine-motor impairment detection14 PD patients and 8 healthy controls (1st dataset), 26 PD patients and 131 healthy controls (2nd dataset)Accelerometer data from a smartphone sensor; typing dynamics from a smartphone virtual keyboardFiltering + feature extraction with 1D-CNN for accelerometer data and fully connected NN for keystroke data + attention pooling module + NN classifiersFor tremor/fine-motor impairment detection on the 1st dataset: SENS = 85.4–92.8%, SPEC = 84.2–93.6%, PREC = 92.1–93%; for diagnosis on the 2nd dataset: ensemble of 10 models with AUROC = 0.834–0.868, SENS/SPEC = 60%/91.7–92%/68.9%
Heidarivincheh et al. (2021) [152]Diagnosis5 PD patients (on medication) and 5 healthy controlsAcceleration signals from a wrist-worn sensor and silhouette images from an RGB-D camera while cooking{Convolutional VAEs + dense layers (MCPD-Net)}/CNN/unimodal VAE/RF/LSTM/other multimodal modelsBest performing: MCPD-Net with PREC = 71%, REC = 77%, F1-SCORE = 66%
Wahid et al. (2015) [153]Diagnosis23 PD patients and 26 healthy controlsSpatial-temporal gait features captured by a video-camera and GRF from force platformsFiltering + Normalization through multiple regression + KFD/NB/kNN/SVM/RFBest performing: RF with AUROC = 0.96, ACC = 92.6%, SENS = 96%, SPEC = 90%
Albani et al. (2019) [154]Diagnosis and severity estimation (UPDRS scores classification)25 PD patients and 15 healthy controls for the final testing of the pre-trained modelsVideo recordings of upper-limbs from an RGB-D camera and accelerometer, gyroscope and magnetometer signals from 3 wearable sensors attached to the thighs and the chestNB/LDA/MNR/kNN/poly-SVM for upper-limb classification (1st pretrained model); (PCA) + kNN/SVM for lower-limb classification (2nd pretrained model)Best performing for diagnosis: SVM (1st model) + kNN (2nd model) with ACC = 91.5–98.6%; for severity estimation: ACC = 60.7–79.1%
Joshi et al. (2018) [155]Facial expressivity estimation (classification and regression)117 PD patientsShort interview audio-video clipsAudio features extraction (MFCC) + visual features extraction (AU statistics) + HBNN classification/regression + contextual information (gender/sentiment)Best performing for classification: HBNN-sentiment with F1-SCORE = 0.55; for regression: MAE = 0.48
Barth et al. (2012) [159]Diagnosis18 PD patients and 17 healthy controlsSignals form a smart pen (acceleration, grip force, refill force, vibration sound); gyroscope and accelerometer data from a shoe-mounted IMUChebyshev low pass filter + linear forward feature selection with CFS/backtracking facility + LDA/linear-SVM/AdaBoostBest performing: linear forward feature selection with CFS + AdaBoost with classification rate = 97%, SENS = 100%, SPEC = 94%
Xu et al. (2020) [156]Diagnosis31 PD patients and 35 healthy controls from an extended HandPD dataset (http://wwwp.fc.unesp.br/~papa/pub/datasets/Handpd/, accessed on 17 February 2022)Pressure, tilt and acceleration signals from a smart pen while performing 6 different handwriting tasksPCA + RF + voting schemeACC = 88.8–89.4%, SENS = 83.7–84.5%, SPEC = 93.4–93.7%, F1-SCORE = 87.2–87.7%
Gallicchio et al. (2018) [157]Diagnosis61 PD patients and 15 healthy controls from a UCI dataset (https://archive.ics.uci.edu/ml/datasets/Parkinson+Disease+Spiral+Drawings+Using+Digitized+Graphics+Tablet, accessed on 17 February 2022)Pen position, pressure and grip angle while sketching spirals on a tabletDeep ESN/shallow ESN + ensemble learning or notBest performing: ensemble of DESNs with ACC = 89.33%, SENS = 90%, SPEC = 80%
Pereira et al. (2016) [158]Diagnosis14 PD patients and 21 healthy controls from an extended HandPD dataset (http://wwwp.fc.unesp.br/~papa/pub/datasets/Handpd/, accessed on 17 February 2022)Various signals from a smart pen (microphone, finger grip, axial pressure of ink refill, tilt and acceleration) while drawing spirals and meandersCNN architectures (ImageNet/CIFAR-10/LeNet)/OPFBest performing for meander dataset: ImageNet with ACC = 84.74–87.14%; for spiral dataset: OPF with ACC = 77.92–83.77%
Schwab et al. (2019) [160]Diagnosis1853 subjects (PD patients and healthy subjects) from the mPower study [126]Touchscreen data from a tapping activity, accelerometer data from a walking activity, performance in a memory game, vocal measurements and demographics from a mobile applicationRF/CNN/RNN for each test + evidence aggregation model combining evidence from all the different testsAUROC = 0.85, AUPR = 0.87, F1-SCORE = 82%
Prince et al. (2019) [161]Diagnosis1513 subjects (PD patients and healthy subjects)Touchscreen data from a tapping activity, accelerometer data from a walking activity, performance in a memory game and vocal measurements from a mobile appLR/RF/DNN/CNN for each test + ensemble learning combining the previous classifiersACC = 82.0%, F1-SCORE = 87.1%
Cook et al. (2015) [162]Diagnosis; classification between healthy older adults and PD patients with and without mild cognitive impairment (HOA/PDNC/HMCI/PDMCI)50 HOA and 25 PD patients for diagnosis; 18 HOA, 16 PDNC, 9 PDMCI, 9 HMCI for their classificationSignals from ambient sensors in a smart house environment (infrared motion sensors on the ceiling, light sensors, magnetic door sensors, temperature sensors and vibration sensors on selected items); accelerometer, gyroscope and magnetometer data from hand- and ankle- mounted sensorsNo dimensionality reduction technique/PCA/k-means clustering/random resampling + DT/NB/RF/SVM/Ada-DT/Ada-RFBest performing for diagnosis: k-means clustering + Ada-DT with ACC = 80%, AUC = 0.84; for the multiclass classification: random resampling + Ada-DT with ACC = 86%, AUC = 0.97

References

  1. Tysnes, O.-B.; Storstein, A. Epidemiology of Parkinson’s Disease. J. Neural. Transm. 2017, 124, 901–905. [Google Scholar] [CrossRef] [PubMed]
  2. Dorsey, E.R.; Elbaz, A.; Nichols, E.; Abd-Allah, F.; Abdelalim, A.; Adsuar, J.C.; Ansha, M.G.; Brayne, C.; Choi, J.-Y.J.; Collado-Mateo, D.; et al. Global, Regional, and National Burden of Parkinson’s Disease, 1990–2016: A Systematic Analysis for the Global Burden of Disease Study 2016. Lancet Neurol. 2018, 17, 939–953. [Google Scholar] [CrossRef] [Green Version]
  3. Kalia, L.V.; Lang, A.E. Parkinson’s Disease. Lancet 2015, 386, 896–912. [Google Scholar] [CrossRef]
  4. Wolters, E.C. Variability in the Clinical Expression of Parkinson’s Disease. J. Neurol. Sci. 2008, 266, 197–203. [Google Scholar] [CrossRef] [PubMed]
  5. Chaudhuri, K.R.; Healy, D.G.; Schapira, A.H. Non-Motor Symptoms of Parkinson’s Disease: Diagnosis and Management. Lancet Neurol. 2006, 5, 235–245. [Google Scholar] [CrossRef]
  6. Ho, A.K.; Iansek, R.; Marigliani, C.; Bradshaw, J.L.; Gates, S. Speech Impairment in a Large Sample of Patients with Parkinson’s Disease. Behav. Neurol. 1998, 11, 131–137. [Google Scholar] [CrossRef]
  7. Harel, B.; Cannizzaro, M.; Snyder, P.J. Variability in Fundamental Frequency during Speech in Prodromal and Incipient Parkinson’s Disease: A Longitudinal Case Study. Brain Cogn. 2004, 56, 24–29. [Google Scholar] [CrossRef]
  8. Marras, C. Subtypes of Parkinson’s Disease: State of the Field and Future Directions. Curr. Opin. Neurol. 2015, 28, 382–386. [Google Scholar] [CrossRef]
  9. Jankovic, J. Motor Fluctuations and Dyskinesias in Parkinson’s Disease: Clinical Manifestations. Mov. Disord. 2005, 20, S11–S16. [Google Scholar] [CrossRef]
  10. Tolosa, E.; Wenning, G.; Poewe, W. The Diagnosis of Parkinson’s Disease. Lancet Neurol. 2006, 5, 75–86. [Google Scholar] [CrossRef]
  11. Bhidayasiri, R.; Martinez-Martin, P. Chapter Six—Clinical Assessments in Parkinson’s Disease: Scales and Monitoring. In International Review of Neurobiology; Bhatia, K.P., Chaudhuri, K.R., Stamelou, M., Eds.; Parkinson’s Disease; Academic Press: Cambridge, MA, USA, 2017; Volume 132, pp. 129–182. [Google Scholar]
  12. Norvig, P.; Russel, S. Artificial Intelligence: A Modern Approach; Prentice Hall Upper: Saddle River, NJ, USA, 2002. [Google Scholar]
  13. Mitchell, T.M. Does Machine Learning Really Work? AI Mag. 1997, 18, 11. [Google Scholar] [CrossRef]
  14. LeCun, Y.; Bengio, Y.; Hinton, G. Deep Learning. Nature 2015, 521, 436–444. [Google Scholar] [CrossRef]
  15. Atzori, L.; Iera, A.; Morabito, G. The Internet of Things: A Survey. Comput. Netw. 2010, 54, 2787–2805. [Google Scholar] [CrossRef]
  16. Espay, A.J.; Bonato, P.; Nahab, F.B.; Maetzler, W.; Dean, J.M.; Klucken, J.; Eskofier, B.M.; Merola, A.; Horak, F.; Lang, A.E.; et al. Technology in Parkinson’s Disease: Challenges and Opportunities: Technology in PD. Mov. Disord. 2016, 31, 1272–1282. [Google Scholar] [CrossRef] [Green Version]
  17. Miele, G.; Straccia, G.; Moccia, M.; Leocani, L.; Tedeschi, G.; Bonavita, S.; Lavorgna, L.; Padovani, A.; Clerico, M.; Brigo, F.; et al. Telemedicine in Parkinson’s Disease: How to Ensure Patient Needs and Continuity of Care at the Time of COVID-19 Pandemic. Telemed. e-Health 2020, 26, 1533–1536. [Google Scholar] [CrossRef] [PubMed]
  18. Raghavendra, U.; Acharya, U.R.; Adeli, H. Artificial Intelligence Techniques for Automated Diagnosis of Neurological Disorders. Eur. Neurol. 2019, 82, 41–64. [Google Scholar] [CrossRef] [PubMed]
  19. Pasluosta, C.F.; Gassner, H.; Winkler, J.; Klucken, J.; Eskofier, B.M. An Emerging Era in the Management of Parkinson’s Disease: Wearable Technologies and the Internet of Things. IEEE J. Biomed. Health Inform. 2015, 19, 1873–1881. [Google Scholar] [CrossRef] [PubMed]
  20. Suzuki, M.; Mitoma, H.; Yoneyama, M. Quantitative Analysis of Motor Status in Parkinson’s Disease Using Wearable Devices: From Methodological Considerations to Problems in Clinical Applications. Park. Dis. 2017, 2017, e6139716. [Google Scholar] [CrossRef] [Green Version]
  21. Monje, M.H.G.; Foffani, G.; Obeso, J.; Sánchez-Ferro, Á. New Sensor and Wearable Technologies to Aid in the Diagnosis and Treatment Monitoring of Parkinson’s Disease. Annu. Rev. Biomed. Eng. 2019, 21, 111–143. [Google Scholar] [CrossRef]
  22. Channa, A.; Popescu, N.; Ciobanu, V. Wearable Solutions for Patients with Parkinson’s Disease and Neurocognitive Disorder: A Systematic Review. Sensors 2020, 20, 2713. [Google Scholar] [CrossRef]
  23. Abou, L.; Peters, J.; Wong, E.; Akers, R.; Dossou, M.S.; Sosnoff, J.J.; Rice, L.A. Gait and Balance Assessments Using Smartphone Applications in Parkinson’s Disease: A Systematic Review. J. Med. Syst. 2021, 45, 87. [Google Scholar] [CrossRef] [PubMed]
  24. Little, M.A. Smartphones for Remote Symptom Monitoring of Parkinson’s Disease. J. Park. Dis. 2021, 11, S49–S53. [Google Scholar] [CrossRef] [PubMed]
  25. Ireland, D.; Liddle, J.; Mcbride, S.; Ding, H.; Knuepffer, C. Chat-Bots for People with Parkinson’s Disease: Science Fiction or Reality? In Driving Reform: Digital Health Is Everyone’s Business; IOS Press: Amsterdam, The Netherlands, 2015; pp. 128–133. [Google Scholar]
  26. Dias, S.B.; Konstantinidis, E.; Diniz, J.A.; Bamidis, P.; Charisis, V.; Hadjidimitriou, S.; Stadtschnitzer, M.; Fagerberg, P.; Ioakeimidis, I.; Dimitropoulos, K.; et al. Serious Games as a Means for Holistically Supporting Parkinson’s Disease Patients: The i-PROGNOSIS Personalized Game Suite Framework. In Proceedings of the 2017 9th International Conference on Virtual Worlds and Games for Serious Applications (VS-Games), Athens, Greece, 6–8 September 2017; pp. 237–244. [Google Scholar]
  27. Zhou, Y.; Jenkins, M.E.; Naish, M.D.; Trejos, A.L. Development of a Wearable Tremor Suppression Glove. In Proceedings of the 2018 7th IEEE International Conference on Biomedical Robotics and Biomechatronics (Biorob), Enschede, The Netherlands, 26–29 August 2018; pp. 640–645. [Google Scholar]
  28. Maetzler, W.; Domingos, J.; Srulijes, K.; Ferreira, J.J.; Bloem, B.R. Quantitative Wearable Sensors for Objective Assessment of Parkinson’s Disease: Wearable Sensors in PD. Mov. Disord. 2013, 28, 1628–1637. [Google Scholar] [CrossRef] [PubMed]
  29. Ossig, C.; Antonini, A.; Buhmann, C.; Classen, J.; Csoti, I.; Falkenburger, B.; Schwarz, M.; Winkler, J.; Storch, A. Wearable Sensor-Based Objective Assessment of Motor Symptoms in Parkinson’s Disease. J. Neural Transm. 2016, 123, 57–64. [Google Scholar] [CrossRef] [PubMed]
  30. Godinho, C.; Domingos, J.; Cunha, G.; Santos, A.T.; Fernandes, R.M.; Abreu, D.; Gonçalves, N.; Matthews, H.; Isaacs, T.; Duffen, J.; et al. A Systematic Review of the Characteristics and Validity of Monitoring Technologies to Assess Parkinson’s Disease. J. Neuroeng. Rehabil. 2016, 13, 24. [Google Scholar] [CrossRef] [Green Version]
  31. Del Din, S.; Kirk, C.; Yarnall, A.J.; Rochester, L.; Hausdorff, J.M. Body-Worn Sensors for Remote Monitoring of Parkinson’s Disease Motor Symptoms: Vision, State of the Art, and Challenges Ahead. J. Park. Dis. 2021, 11, S35–S47. [Google Scholar] [CrossRef] [PubMed]
  32. Ferreira-Sánchez, M.D.R.; Moreno-Verdú, M.; Cano-de-la-Cuerda, R. Quantitative Measurement of Rigidity in Parkinson’s Disease: A Systematic Review. Sensors 2020, 20, 880. [Google Scholar] [CrossRef] [Green Version]
  33. Vienne, A.; Barrois, R.P.; Buffat, S.; Ricard, D.; Vidal, P.-P. Inertial Sensors to Assess Gait Quality in Patients with Neurological Disorders: A Systematic Review of Technical and Analytical Challenges. Front. Psychol. 2017, 8, 817. [Google Scholar] [CrossRef] [Green Version]
  34. De Oliveira Gondim, I.T.G.; de Souza, C.D.C.B.; Rodrigues, M.A.B.; Azevedo, I.M.; de Sales, M.D.G.W.; Lins, O.G. Portable Accelerometers for the Evaluation of Spatio-Temporal Gait Parameters in People with Parkinson’s Disease: An Integrative Review. Arch. Gerontol. Geriatr. 2020, 90, 104097. [Google Scholar] [CrossRef]
  35. di Biase, L.; Di Santo, A.; Caminiti, M.L.; De Liso, A.; Shah, S.A.; Ricci, L.; Di Lazzaro, V. Gait Analysis in Parkinson’s Disease: An Overview of the Most Accurate Markers for Diagnosis and Symptoms Monitoring. Sensors 2020, 20, 3529. [Google Scholar] [CrossRef]
  36. Silva de Lima, A.L.; Evers, L.J.W.; Hahn, T.; Bataille, L.; Hamilton, J.L.; Little, M.A.; Okuma, Y.; Bloem, B.R.; Faber, M.J. Freezing of Gait and Fall Detection in Parkinson’s Disease Using Wearable Sensors: A Systematic Review. J. Neurol. 2017, 264, 1642–1654. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  37. Pardoel, S.; Kofman, J.; Nantel, J.; Lemaire, E.D. Wearable-Sensor-Based Detection and Prediction of Freezing of Gait in Parkinson’s Disease: A Review. Sensors 2019, 19, 5141. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  38. Sun, M.; Watson, A.; Zhou, G. Wearable Computing of Freezing of Gait in Parkinson’s Disease: A Survey. Smart Health 2020, 18, 100143. [Google Scholar] [CrossRef]
  39. Maitín, A.M.; García-Tejedor, A.J.; Muñoz, J.P.R. Machine Learning Approaches for Detecting Parkinson’s Disease from EEG Analysis: A Systematic Review. Appl. Sci. 2020, 10, 8662. [Google Scholar] [CrossRef]
  40. Mei, J.; Desrosiers, C.; Frasnelli, J. Machine Learning for the Diagnosis of Parkinson’s Disease: A Review of Literature. Front. Aging Neurosci. 2021, 13, 184. [Google Scholar] [CrossRef]
  41. Rovini, E.; Maremmani, C.; Cavallo, F. How Wearable Sensors Can Support Parkinson’s Disease Diagnosis and Treatment: A Systematic Review. Front. Neurosci. 2017, 11, 555. [Google Scholar] [CrossRef]
  42. Thorp, J.E.; Adamczyk, P.G.; Ploeg, H.-L.; Pickett, K.A. Monitoring Motor Symptoms During Activities of Daily Living in Individuals with Parkinson’s Disease. Front. Neurol. 2018, 9, 1036. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  43. Ramdhani, R.A.; Khojandi, A.; Shylo, O.; Kopell, B.H. Optimizing Clinical Assessments in Parkinson’s Disease Through the Use of Wearable Sensors and Data Driven Modeling. Front. Comput. Neurosci. 2018, 12, 72. [Google Scholar] [CrossRef] [Green Version]
  44. Belić, M.; Bobić, V.; Badža, M.; Šolaja, N.; Đurić-Jovičić, M.; Kostić, V.S. Artificial Intelligence for Assisting Diagnostics and Assessment of Parkinson’s Disease—A Review. Clin. Neurol. Neurosurg. 2019, 184, 105442. [Google Scholar] [CrossRef] [PubMed]
  45. Zhang, H.; Song, C.; Rathore, A.S.; Huang, M.-C.; Zhang, Y.; Xu, W. MHealth Technologies Towards Parkinson’s Disease Detection and Monitoring in Daily Life: A Comprehensive Review. IEEE Rev. Biomed. Eng. 2021, 14, 71–81. [Google Scholar] [CrossRef]
  46. Sica, M.; Tedesco, S.; Crowe, C.; Kenny, L.; Moore, K.; Timmons, S.; Barton, J.; O’Flynn, B.; Komaris, D.-S. Continuous Home Monitoring of Parkinson’s Disease Using Inertial Sensors: A Systematic Review. PLoS ONE 2021, 16, e0246528. [Google Scholar] [CrossRef]
  47. Barrachina-Fernández, M.; Maitín, A.M.; Sánchez-Ávila, C.; Romero, J.P. Wearable Technology to Detect Motor Fluctuations in Parkinson’s Disease Patients: Current State and Challenges. Sensors 2021, 21, 4188. [Google Scholar] [CrossRef]
  48. Page, M.J.; McKenzie, J.E.; Bossuyt, P.M.; Boutron, I.; Hoffmann, T.C.; Mulrow, C.D.; Shamseer, L.; Tetzlaff, J.M.; Akl, E.A.; Brennan, S.E.; et al. The PRISMA 2020 Statement: An Updated Guideline for Reporting Systematic Reviews. BMJ 2021, 372, n71. [Google Scholar] [CrossRef]
  49. Rastegari, E.; Ali, H. A Bag-of-Words Feature Engineering Approach for Assessing Health Conditions Using Accelerometer Data. Smart Health 2020, 16, 100116. [Google Scholar] [CrossRef]
  50. Juutinen, M.; Wang, C.; Zhu, J.; Haladjian, J.; Ruokolainen, J.; Puustinen, J.; Vehkaoja, A. Parkinson’s Disease Detection from 20-Step Walking Tests Using Inertial Sensors of a Smartphone: Machine Learning Approach Based on an Observational Case-Control Study. PLoS ONE 2020, 15, e0236258. [Google Scholar] [CrossRef]
  51. Fernandes, C.; Fonseca, L.; Ferreira, F.; Gago, M.; Costa, L.; Sousa, N.; Ferreira, C.; Gama, J.; Erlhagen, W.; Bicho, E. Artificial Neural Networks Classification of Patients with Parkinsonism Based on Gait. In Proceedings of the 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), Madrid, Spain, 3–6 December 2018; pp. 2024–2030. [Google Scholar]
  52. Cuzzolin, F.; Sapienza, M.; Esser, P.; Saha, S.; Franssen, M.M.; Collett, J.; Dawes, H. Metric Learning for Parkinsonian Identification from IMU Gait Measurements. Gait Posture 2017, 54, 127–132. [Google Scholar] [CrossRef] [Green Version]
  53. Abujrida, H.; Agu, E.; Pahlavan, K. Machine Learning-Based Motor Assessment of Parkinson’s Disease Using Postural Sway, Gait and Lifestyle Features on Crowdsourced Smartphone Data. Biomed. Phys. Eng. Express 2020, 6, 035005. [Google Scholar] [CrossRef] [PubMed]
  54. Zhang, H.; Deng, K.; Li, H.; Albin, R.L.; Guan, Y. Deep Learning Identifies Digital Biomarkers for Self-Reported Parkinson’s Disease. Patterns 2020, 1, 100042. [Google Scholar] [CrossRef]
  55. Kostikis, N.; Hristu-Varsakelis, D.; Arnaoutoglou, M.; Kotsavasiloglou, C. A Smartphone-Based Tool for Assessing Parkinsonian Hand Tremor. IEEE J. Biomed. Health Inform. 2015, 19, 1835–1842. [Google Scholar] [CrossRef] [PubMed]
  56. Koçer, A.; Oktay, A.B. Nintendo Wii Assessment of Hoehn and Yahr Score with Parkinson’s Disease Tremor. Technol. Health Care 2016, 24, 185–191. [Google Scholar] [CrossRef]
  57. Li, N.; Tian, F.; Fan, X.; Zhu, Y.; Wang, H.; Dai, G. Monitoring Motor Symptoms in Parkinson’s Disease via Instrumenting Daily Artifacts with Inertia Sensors. CCF Trans. Pervasive Comput. Interact. 2019, 1, 100–113. [Google Scholar] [CrossRef] [Green Version]
  58. Javed, F.; Thomas, I.; Memedi, M. A Comparison of Feature Selection Methods When Using Motion Sensors Data: A Case Study in Parkinson’s Disease. In Proceedings of the 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Honolulu, HI, USA, 18–21 July 2018; pp. 5426–5429. [Google Scholar]
  59. Williamson, J.R.; Telfer, B.; Mullany, R.; Friedl, K.E. Detecting Parkinson’s Disease from Wrist-Worn Accelerometry in the U.K. Biobank. Sensors 2021, 21, 2047. [Google Scholar] [CrossRef] [PubMed]
  60. Park, D.J.; Lee, J.W.; Lee, M.J.; Ahn, S.J.; Kim, J.; Kim, G.L.; Ra, Y.J.; Cho, Y.N.; Jeong, W.B. Evaluation for Parkinsonian Bradykinesia by Deep Learning Modeling of Kinematic Parameters. J. Neural Transm. 2021, 128, 181–189. [Google Scholar] [CrossRef]
  61. Som, A.; Krishnamurthi, N.; Buman, M.; Turaga, P. Unsupervised Pre-Trained Models from Healthy ADLs Improve Parkinson’s Disease Classification of Gait Patterns. In Proceedings of the 2020 42nd Annual International Conference of the IEEE Engineering in Medicine Biology Society (EMBC), Montreal, QC, Canada, 20–24 July 2020; pp. 784–788. [Google Scholar]
  62. Ricci, M.; Di Lazzaro, G.; Pisani, A.; Mercuri, N.B.; Giannini, F.; Saggio, G. Assessment of Motor Impairments in Early Untreated Parkinson’s Disease Patients: The Wearable Electronics Impact. IEEE J. Biomed. Health Inform. 2020, 24, 120–130. [Google Scholar] [CrossRef] [PubMed]
  63. Kuhner, A.; Schubert, T.; Cenciarini, M.; Wiesmeier, I.K.; Coenen, V.A.; Burgard, W.; Weiller, C.; Maurer, C. Correlations between Motor Symptoms across Different Motor Tasks, Quantified via Random Forest Feature Classification in Parkinson’s Disease. Front. Neurol. 2017, 8, 607. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  64. Caramia, C.; Torricelli, D.; Schmid, M.; Munoz-Gonzalez, A.; Gonzalez-Vargas, J.; Grandas, F.; Pons, J.L. IMU-Based Classification of Parkinson’s Disease from Gait: A Sensitivity Analysis on Sensor Location and Feature Selection. IEEE J. Biomed. Health Inform. 2018, 22, 1765–1774. [Google Scholar] [CrossRef]
  65. Talitckii, A.; Kovalenko, E.; Anikina, A.; Zimniakova, O.; Semenov, M.; Bril, E.; Shcherbak, A.; Dylov, D.V.; Somov, A. Avoiding Misdiagnosis of Parkinson’s Disease with the Use of Wearable Sensors and Artificial Intelligence. IEEE Sens. J. 2021, 21, 3738–3747. [Google Scholar] [CrossRef]
  66. Varghese, J.; Fujarski, M.; Hahn, T.; Dugas, M.; Warnecke, T. The Smart Device System for Movement Disorders: Preliminary Evaluation of Diagnostic Accuracy in a Prospective Study. In Digital Personalized Health and Medicine; Series of Studies in Health Technology and Informatics; IOS Press: Amsterdam, The Netherlands, 2020; Volume 270, pp. 889–893. [Google Scholar] [CrossRef]
  67. Duque, J.D.L.; Egea, A.J.S.; Reeb, T.; Rojas, H.A.G.; González-Vargas, A.M. Angular Velocity Analysis Boosted by Machine Learning for Helping in the Differential Diagnosis of Parkinson’s Disease and Essential Tremor. IEEE Access 2020, 8, 88866–88875. [Google Scholar] [CrossRef]
  68. Moon, S.; Song, H.-J.; Sharma, V.D.; Lyons, K.E.; Pahwa, R.; Akinwuntan, A.E.; Devos, H. Classification of Parkinson’s Disease and Essential Tremor Based on Balance and Gait Characteristics from Wearable Motion Sensors via Machine Learning Techniques: A Data-Driven Approach. J. Neuroeng. Rehabil. 2020, 17, 125. [Google Scholar] [CrossRef]
  69. De Vos, M.; Prince, J.; Buchanan, T.; FitzGerald, J.J.; Antoniades, C.A. Discriminating Progressive Supranuclear Palsy from Parkinson’s Disease Using Wearable Technology and Machine Learning. Gait Posture 2020, 77, 257–263. [Google Scholar] [CrossRef]
  70. Borzì, L.; Varrecchia, M.; Sibille, S.; Olmo, G.; Artusi, C.A.; Fabbri, M.; Rizzone, M.G.; Romagnolo, A.; Zibetti, M.; Lopiano, L. Smartphone-Based Estimation of Item 3.8 of the MDS-UPDRS-III for Assessing Leg Agility in People with Parkinson’s Disease. IEEE Open J. Eng. Med. Biol. 2020, 1, 140–147. [Google Scholar] [CrossRef]
  71. Bazgir, O.; Frounchi, J.; Habibi, S.A.H.; Palma, L.; Pierleoni, P. A Neural Network System for Diagnosis and Assessment of Tremor in Parkinson Disease Patients. In Proceedings of the 2015 22nd Iranian Conference on Biomedical Engineering (ICBME), Tehran, Iran, 25–27 November 2015; pp. 1–5. [Google Scholar]
  72. Kim, H.B.; Lee, W.W.; Kim, A.; Lee, H.J.; Park, H.Y.; Jeon, H.S.; Kim, S.K.; Jeon, B.; Park, K.S. Wrist Sensor-Based Tremor Severity Quantification in Parkinson’s Disease Using Convolutional Neural Network. Comput. Biol. Med. 2018, 95, 140–146. [Google Scholar] [CrossRef] [PubMed]
  73. Dai, H.; Cai, G.; Lin, Z.; Wang, Z.; Ye, Q. Validation of Inertial Sensing-Based Wearable Device for Tremor and Bradykinesia Quantification. IEEE J. Biomed. Health Inform. 2021, 25, 997–1005. [Google Scholar] [CrossRef] [PubMed]
  74. Khodakarami, H.; Ricciardi, L.; Contarino, M.F.; Pahwa, R.; Lyons, K.E.; Geraedts, V.J.; Morgante, F.; Leake, A.; Paviour, D.; De Angelis, A.; et al. Prediction of the Levodopa Challenge Test in Parkinson’s Disease Using Data from a Wrist-Worn Sensor. Sensors 2019, 19, 5153. [Google Scholar] [CrossRef] [Green Version]
  75. Mirelman, A.; Ben or Frank, M.; Melamed, M.; Granovsky, L.; Nieuwboer, A.; Rochester, L.; Del Din, S.; Avanzino, L.; Pelosin, E.; Bloem, B.R.; et al. Detecting Sensitive Mobility Features for Parkinson’s Disease Stages Via Machine Learning. Mov. Disord. 2021, 36, 2144–2155. [Google Scholar] [CrossRef]
  76. Hssayeni, M.D.; Jimenez-Shahed, J.; Burack, M.A.; Ghoraani, B. Ensemble Deep Model for Continuous Estimation of Unified Parkinson’s Disease Rating Scale III. Biomed. Eng. Online 2021, 20, 32. [Google Scholar] [CrossRef]
  77. Butt, A.H.; Rovini, E.; Fujita, H.; Maremmani, C.; Cavallo, F. Data-Driven Models for Objective Grading Improvement of Parkinson’s Disease. Ann. Biomed. Eng. 2020, 48, 2976–2987. [Google Scholar] [CrossRef]
  78. Stamate, C.; Magoulas, G.D.; Kueppers, S.; Nomikou, E.; Daskalopoulos, I.; Jha, A.; Pons, J.S.; Rothwell, J.; Luchini, M.U.; Moussouri, T.; et al. The CloudUPDRS App: A Medical Device for the Clinical Assessment of Parkinson’s Disease. Pervasive Mob. Comput. 2018, 43, 146–166. [Google Scholar] [CrossRef]
  79. Eskofier, B.M.; Lee, S.I.; Daneault, J.-F.; Golabchi, F.N.; Ferreira-Carvalho, G.; Vergara-Diaz, G.; Sapienza, S.; Costante, G.; Klucken, J.; Kautz, T.; et al. Recent Machine Learning Advancements in Sensor-Based Mobility Analysis: Deep Learning for Parkinson’s Disease Assessment. In Proceedings of the 2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Orlando, FL, USA, 16–20 August 2016; pp. 655–658. [Google Scholar]
  80. Shawen, N.; O’Brien, M.K.; Venkatesan, S.; Lonini, L.; Simuni, T.; Hamilton, J.L.; Ghaffari, R.; Rogers, J.A.; Jayaraman, A. Role of Data Measurement Characteristics in the Accurate Detection of Parkinson’s Disease Symptoms Using Wearable Sensors. J. Neuroeng. Rehabil. 2020, 17, 52. [Google Scholar] [CrossRef] [Green Version]
  81. San-Segundo, R.; Zhang, A.; Cebulla, A.; Panev, S.; Tabor, G.; Stebbins, K.; Massa, R.E.; Whitford, A.; de la Torre, F.; Hodgins, J. Parkinson’s Disease Tremor Detection in the Wild Using Wearable Accelerometers. Sensors 2020, 20, 5817. [Google Scholar] [CrossRef]
  82. Ibrahim, A.; Zhou, Y.; Jenkins, M.E.; Naish, M.D.; Trejos, A.L. Parkinson’s Tremor Onset Detection and Active Tremor Classification Using a Multilayer Perceptron. In Proceedings of the 2020 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE), London, ON, Canada, 30 August–2 September 2020; pp. 1–4. [Google Scholar]
  83. Channa, A.; Ifrim, R.-C.; Popescu, D.; Popescu, N. A-WEAR Bracelet for Detection of Hand Tremor and Bradykinesia in Parkinson’s Patients. Sensors 2021, 21, 981. [Google Scholar] [CrossRef] [PubMed]
  84. Kim, H.; Lee, H.J.; Lee, W.; Kwon, S.; Kim, S.K.; Jeon, H.S.; Park, H.; Shin, C.W.; Yi, W.J.; Jeon, B.S.; et al. Unconstrained Detection of Freezing of Gait in Parkinson’s Disease Patients Using Smartphone. In Proceedings of the 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Milan, Italy, 25–29 August 2015; pp. 3751–3754. [Google Scholar]
  85. Shi, B.; Yen, S.C.; Tay, A.; Tan, D.M.L.; Chia, N.S.Y.; Au, W.L. Convolutional Neural Network for Freezing of Gait Detection Leveraging the Continuous Wavelet Transform on Lower Extremities Wearable Sensors Data. In Proceedings of the 2020 42nd Annual International Conference of the IEEE Engineering in Medicine Biology Society (EMBC), Montreal, QC, Canada, 20–24 July 2020; pp. 5410–5415. [Google Scholar]
  86. Camps, J.; Samà, A.; Martín, M.; Rodríguez-Martín, D.; Pérez-López, C.; Moreno Arostegui, J.M.; Cabestany, J.; Català, A.; Alcaine, S.; Mestre, B.; et al. Deep Learning for Freezing of Gait Detection in Parkinson’s Disease Patients in Their Homes Using a Waist-Worn Inertial Measurement Unit. Knowl.-Based Syst. 2018, 139, 119–131. [Google Scholar] [CrossRef]
  87. Ashour, A.S.; El-Attar, A.; Dey, N.; El-Kader, H.A.; Abd El-Naby, M.M. Long short term Memory Based Patient-Dependent Model for FOG Detection in Parkinson’s Disease. Pattern Recognit. Lett. 2020, 131, 23–29. [Google Scholar] [CrossRef]
  88. Li, B.; Sun, Y.; Yao, Z.; Wang, J.; Wang, S.; Yang, X. Improved Deep Learning Technique to Detect Freezing of Gait in Parkinson’s Disease Based on Wearable Sensors. Electronics 2020, 9, 1919. [Google Scholar] [CrossRef]
  89. Torvi, V.G.; Bhattacharya, A.; Chakraborty, S. Deep Domain Adaptation to Predict Freezing of Gait in Patients with Parkinson’s Disease. In Proceedings of the 2018 17th IEEE International Conference on Machine Learning and Applications (ICMLA), Orlando, FL, USA, 17–20 December 2018; pp. 1001–1006. [Google Scholar]
  90. Arami, A.; Poulakakis-Daktylidis, A.; Tai, Y.F.; Burdet, E. Prediction of Gait Freezing in Parkinsonian Patients: A Binary Classification Augmented with Time Series Prediction. IEEE Trans. Neural Syst. Rehabil. Eng. 2019, 27, 1909–1919. [Google Scholar] [CrossRef]
  91. Kleanthous, N.; Hussain, A.J.; Khan, W.; Liatsis, P. A New Machine Learning Based Approach to Predict Freezing of Gait. Pattern Recognit. Lett. 2020, 140, 119–126. [Google Scholar] [CrossRef]
  92. Halder, A.; Singh, R.; Suri, A.; Joshi, D. Predicting State Transition in Freezing of Gait via Acceleration Measurements for Controlled Cueing in Parkinson’s Disease. IEEE Trans. Instrum. Meas. 2021, 70, 1–16. [Google Scholar] [CrossRef]
  93. Palmerini, L.; Rocchi, L.; Mazilu, S.; Gazit, E.; Hausdorff, J.M.; Chiari, L. Identification of Characteristic Motor Patterns Preceding Freezing of Gait in Parkinson’s Disease Using Wearable Sensors. Front. Neurol. 2017, 8, 394. [Google Scholar] [CrossRef] [Green Version]
  94. Borzì, L.; Mazzetta, I.; Zampogna, A.; Suppa, A.; Olmo, G.; Irrera, F. Prediction of Freezing of Gait in Parkinson’s Disease Using Wearables and Machine Learning. Sensors 2021, 21, 614. [Google Scholar] [CrossRef]
  95. Haji Ghassemi, N.; Hannink, J.; Martindale, C.F.; Gaßner, H.; Müller, M.; Klucken, J.; Eskofier, B.M. Segmentation of Gait Sequences in Sensor-Based Movement Analysis: A Comparison of Methods in Parkinson’s Disease. Sensors 2018, 18, 145. [Google Scholar] [CrossRef] [Green Version]
  96. Hssayeni, M.D.; Adams, J.L.; Ghoraani, B. Deep Learning for Medication Assessment of Individuals with Parkinson’s Disease Using Wearable Sensors. In Proceedings of the 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Honolulu, HI, USA, 18–21 July 2018; pp. 1–4. [Google Scholar]
  97. Aich, S.; Youn, J.; Chakraborty, S.; Pradhan, P.M.; Park, J.-H.; Park, S.; Park, J. A Supervised Machine Learning Approach to Detect the On/Off State in Parkinson’s Disease Using Wearable Based Gait Signals. Diagnostics 2020, 10, 421. [Google Scholar] [CrossRef]
  98. Pfister, F.M.J.; Um, T.T.; Pichler, D.C.; Goschenhofer, J.; Abedinpour, K.; Lang, M.; Endo, S.; Ceballos-Baumann, A.O.; Hirche, S.; Bischl, B.; et al. High-Resolution Motor State Detection in Parkinson’s Disease Using Convolutional Neural Networks. Sci. Rep. 2020, 10, 5860. [Google Scholar] [CrossRef] [Green Version]
  99. Belgiovine, G.; Capecci, M.; Ciabattoni, L.; Fiorentino, M.C.; Foresi, G.; Monteriù, A.; Pepa, L. Upper and Lower Limbs Dyskinesia Detection for Patients with Parkinson’s Disease. In Proceedings of the 2018 IEEE 7th Global Conference on Consumer Electronics (GCCE), Nara, Japan, 9–12 October 2018; pp. 704–705. [Google Scholar]
  100. Watts, J.; Khojandi, A.; Vasudevan, R.; Nahab, F.B.; Ramdhani, R.A. Improving Medication Regimen Recommendation for Parkinson’s Disease Using Sensor Technology. Sensors 2021, 21, 3553. [Google Scholar] [CrossRef] [PubMed]
  101. Nancy Jane, Y.; Khanna Nehemiah, H.; Arputharaj, K. A Q-Backpropagated Time Delay Neural Network for Diagnosing Severity of Gait Disturbances in Parkinson’s Disease. J. Biomed. Inform. 2016, 60, 169–176. [Google Scholar] [CrossRef]
  102. Aversano, L.; Bernardi, M.L.; Cimitile, M.; Pecori, R. Early Detection of Parkinson Disease Using Deep Neural Networks on Gait Dynamics. In Proceedings of the 2020 International Joint Conference on Neural Networks (IJCNN), Glasgow, UK, 19–24 July 2020; pp. 1–8. [Google Scholar]
  103. El Maachi, I.; Bilodeau, G.-A.; Bouachir, W. Deep 1D-Convnet for Accurate Parkinson Disease Detection and Severity Prediction from Gait. Expert Syst. Appl. 2020, 143, 113075. [Google Scholar] [CrossRef]
  104. Xia, Y.; Yao, Z.; Ye, Q.; Cheng, N. A Dual-Modal Attention-Enhanced Deep Learning Network for Quantification of Parkinson’s Disease Characteristics. IEEE Trans. Neural Syst. Rehabil. Eng. 2020, 28, 42–51. [Google Scholar] [CrossRef] [PubMed]
  105. Balaji, E.; Brindha, D.; Vinodh Kumar, E.; Umesh, K. Data-Driven Gait Analysis for Diagnosis and Severity Rating of Parkinson’s Disease. Med. Eng. Phys. 2021, 91, 54–64. [Google Scholar] [CrossRef]
  106. Goldberger, A.L.; Amaral, L.A.N.; Glass, L.; Hausdorff, J.M.; Ivanov, P.C.; Mark, R.G.; Mietus, J.E.; Moody, G.B.; Peng, C.-K.; Stanley, H.E. PhysioBank, PhysioToolkit, and PhysioNet. Circulation 2000, 101, e215–e220. [Google Scholar] [CrossRef] [Green Version]
  107. Papavasileiou, I.; Zhang, W.; Wang, X.; Bi, J.; Zhang, L.; Han, S. Classification of Neurological Gait Disorders Using Multi-Task Feature Learning. In Proceedings of the 2017 IEEE/ACM International Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE), Philadelphia, PA, USA, 17–19 July 2017; pp. 195–204. [Google Scholar]
  108. Khoury, N.; Attal, F.; Amirat, Y.; Oukhellou, L.; Mohammed, S. Data-Driven Based Approach to Aid Parkinson’s Disease Diagnosis. Sensors 2019, 19, 242. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  109. Reyes, J.F.; Steven Montealegre, J.; Castano, Y.J.; Urcuqui, C.; Navarro, A. LSTM and Convolution Networks Exploration for Parkinson’s Diagnosis. In Proceedings of the 2019 IEEE Colombian Conference on Communications and Computing (COLCOM), Barranquilla, Colombia, 5–7 June 2019; pp. 1–4. [Google Scholar]
  110. Buongiorno, D.; Bortone, I.; Cascarano, G.D.; Trotta, G.F.; Brunetti, A.; Bevilacqua, V. A Low-Cost Vision System Based on the Analysis of Motor Features for Recognition and Severity Rating of Parkinson’s Disease. BMC Med. Inform. Decis. Mak. 2019, 19, 243. [Google Scholar] [CrossRef]
  111. Guayacán, L.C.; Rangel, E.; Martínez, F. Towards Understanding Spatio-Temporal Parkinsonian Patterns from Salient Regions of a 3D Convolutional Network. In Proceedings of the 2020 42nd Annual International Conference of the IEEE Engineering in Medicine Biology Society (EMBC), Montreal, QC, Canada, 20–24 July 2020; pp. 3688–3691. [Google Scholar]
  112. Vivar-Estudillo, G.; Ibarra-Manzano, M.-A.; Almanza-Ojeda, D.-L. Tremor Signal Analysis for Parkinson’s Disease Detection Using Leap Motion Device. In Advances in Soft Computing; Lecture Notes in Computer Science; Springer: Cham, Switzerland, 2018; Volume 11288, p. 353. [Google Scholar]
  113. Moshkova, A.; Samorodov, A.; Voinova, N.; Volkov, A.; Ivanova, E.; Fedotova, E. Parkinson’s Disease Detection by Using Machine Learning Algorithms and Hand Movement Signal from LeapMotion Sensor. In Proceedings of the 2020 26th Conference of Open Innovations Association (FRUCT), Yaroslavl, Russia, 20–24 April 2020; pp. 321–327. [Google Scholar]
  114. Ali, M.R.; Hernandez, J.; Dorsey, E.R.; Hoque, E.; McDuff, D. Spatio-Temporal Attention and Magnification for Classification of Parkinson’s Disease from Videos Collected via the Internet. In Proceedings of the 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), Buenos Aires, Argentina, 16–20 November 2020; pp. 207–214. [Google Scholar]
  115. Jin, B.; Qu, Y.; Zhang, L.; Gao, Z. Diagnosing Parkinson Disease through Facial Expression Recognition: Video Analysis. J. Med. Internet Res. 2020, 22, e18697. [Google Scholar] [CrossRef] [PubMed]
  116. Rajnoha, M.; Mekyska, J.; Burget, R.; Eliasova, I.; Kostalova, M.; Rektorova, I. Towards Identification of Hypomimia in Parkinson’s Disease Based on Face Recognition Methods. In Proceedings of the 2018 10th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), Moscow, Russia, 5–9 November 2018; pp. 1–4. [Google Scholar]
  117. Grammatikopoulou, A.; Dimitropoulos, K.; Bostantjopoulou, S.; Katsarou, Z.; Grammalidis, N. Motion Analysis of Parkinson Diseased Patients Using a Video Game Approach. In Proceedings of the 12th ACM International Conference on PErvasive Technologies Related to Assistive Environments, Rhodes, Greece, 5–7 June 2019; pp. 523–527. [Google Scholar]
  118. Li, M.H.; Mestre, T.A.; Fox, S.H.; Taati, B. Vision-Based Assessment of Parkinsonism and Levodopa-Induced Dyskinesia with Pose Estimation. J. Neuroeng. Rehabil. 2018, 15, 97. [Google Scholar] [CrossRef] [Green Version]
  119. Liu, Y.; Chen, J.; Hu, C.; Ma, Y.; Ge, D.; Miao, S.; Xue, Y.; Li, L. Vision-Based Method for Automatic Quantification of Parkinsonian Bradykinesia. IEEE Trans. Neural Syst. Rehabil. Eng. 2019, 27, 1952–1961. [Google Scholar] [CrossRef] [PubMed]
  120. Li, T.; Chen, J.; Hu, C.; Ma, Y.; Wu, Z.; Wan, W.; Huang, Y.; Jia, F.; Gong, C.; Wan, S.; et al. Automatic Timed Up-and-Go Sub-Task Segmentation for Parkinson’s Disease Patients Using Video-Based Activity Classification. IEEE Trans. Neural Syst. Rehabil. Eng. 2018, 26, 2189–2199. [Google Scholar] [CrossRef] [PubMed]
  121. Hu, K.; Wang, Z.; Wang, W.; Ehgoetz Martens, K.A.; Wang, L.; Tan, T.; Lewis, S.J.G.; Feng, D.D. Graph Sequence Recurrent Neural Network for Vision-Based Freezing of Gait Detection. IEEE Trans. Image Process. 2020, 29, 1890–1901. [Google Scholar] [CrossRef] [PubMed]
  122. Tucker, C.S.; Behoora, I.; Nembhard, H.B.; Lewis, M.; Sterling, N.W.; Huang, X. Machine Learning Classification of Medication Adherence in Patients with Movement Disorders Using Non-Wearable Sensors. Comput. Biol. Med. 2015, 66, 120–134. [Google Scholar] [CrossRef] [Green Version]
  123. Wei, W.; McElroy, C.; Dey, S. Towards On-Demand Virtual Physical Therapist: Machine Learning-Based Patient Action Understanding, Assessment and Task Recommendation. IEEE Trans. Neural Syst. Rehabil. Eng. 2019, 27, 1824–1835. [Google Scholar] [CrossRef]
  124. Zhang, H.; Wang, A.; Li, D.; Xu, W. DeepVoice: A Voiceprint-Based Mobile Health Framework for Parkinson’s Disease Identification. In Proceedings of the 2018 IEEE EMBS International Conference on Biomedical Health Informatics (BHI), Las Vegas, NV, USA, 4–7 March 2018; pp. 214–217. [Google Scholar]
  125. Tougui, I.; Jilbab, A.; Mhamdi, J.E. Analysis of Smartphone Recordings in Time, Frequency, and Cepstral Domains to Classify Parkinson’s Disease. Healthc. Inform. Res. 2020, 26, 274–283. [Google Scholar] [CrossRef] [PubMed]
  126. Bot, B.M.; Suver, C.; Neto, E.C.; Kellen, M.; Klein, A.; Bare, C.; Doerr, M.; Pratap, A.; Wilbanks, J.; Dorsey, E.R.; et al. The MPower Study, Parkinson Disease Mobile Data Collected Using ResearchKit. Sci. Data 2016, 3, 160011. [Google Scholar] [CrossRef]
  127. Zhang, Y.N. Can a Smartphone Diagnose Parkinson Disease? A Deep Neural Network Method and Telediagnosis System Implementation. Park. Dis. 2017, 2017, 6209703. [Google Scholar] [CrossRef] [Green Version]
  128. Almeida, J.S.; Rebouças Filho, P.P.; Carneiro, T.; Wei, W.; Damaševičius, R.; Maskeliūnas, R.; de Albuquerque, V.H.C. Detecting Parkinson’s Disease with Sustained Phonation and Speech Signals Using Machine Learning Techniques. Pattern Recognit. Lett. 2019, 125, 55–62. [Google Scholar] [CrossRef] [Green Version]
  129. Arora, S.; Lo, C.; Hu, M.; Tsanas, A. Smartphone Speech Testing for Symptom Assessment in Rapid Eye Movement Sleep Behavior Disorder and Parkinson’s Disease. IEEE Access 2021, 9, 44813–44824. [Google Scholar] [CrossRef]
  130. Yoon, H.; Li, J. A Novel Positive Transfer Learning Approach for Telemonitoring of Parkinson’s Disease. IEEE Trans. Autom. Sci. Eng. 2019, 16, 180–191. [Google Scholar] [CrossRef]
  131. Raza, M.; Awais, M.; Singh, N.; Imran, M.; Hussain, S. Intelligent IoT Framework for Indoor Healthcare Monitoring of Parkinson’s Disease Patient. IEEE J. Sel. Areas Commun. 2021, 39, 593–602. [Google Scholar] [CrossRef]
  132. Bayestehtashk, A.; Asgari, M.; Shafran, I.; McNames, J. Fully Automated Assessment of the Severity of Parkinson’s Disease from Speech. Comput. Speech Lang. 2015, 29, 172–185. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  133. Rahman, M.A.; Tutul, A.A.; Islam, A.B.M.A.A. Solving the Maze of Diagnosing Parkinson’s Disease Based on Portable EEG Sensing to Be Adaptable to Go In-The-Wild. In Proceedings of the 7th International Conference on Networking, Systems and Security, Dhaka, Bangladesh, 22–24 December 2020; pp. 65–73. [Google Scholar]
  134. Kleinholdermann, U.; Wullstein, M.; Pedrosa, D. Prediction of Motor Unified Parkinson’s Disease Rating Scale Scores in Patients with Parkinson’s Disease Using Surface Electromyography. Clin. Neurophysiol. 2021, 132, 1708–1713. [Google Scholar] [CrossRef] [PubMed]
  135. Capecci, M.; Ciabattoni, L.; Foresi, G.; Monteriù, A.; Pepa, L. A Machine-Learning Based Emotion Recognition System in Patients with Parkinson’s Disease. In Proceedings of the 2019 IEEE 9th International Conference on Consumer Electronics (ICCE-Berlin), Berlin, Germany, 8–11 September 2019; pp. 20–21. [Google Scholar]
  136. Lacy, S.E.; Smith, S.L.; Lones, M.A. Using Echo State Networks for Classification: A Case Study in Parkinson’s Disease Diagnosis. Artif. Intell. Med. 2018, 86, 53–59. [Google Scholar] [CrossRef]
  137. Picardi, C.; Cosgrove, J.; Smith, S.L.; Jamieson, S.; Alty, J.E. Objective Assessment of Cognitive Impairment in Parkinson’s Disease Using Evolutionary Algorithm. In Applications of Evolutionary Computation; Lecture Notes in Computer Science; Springer: Cham, Switzerland, 2017; Volume 10199, p. 124. [Google Scholar]
  138. Memedi, M.; Sadikov, A.; Groznik, V.; Žabkar, J.; Možina, M.; Bergquist, F.; Johansson, A.; Haubenberger, D.; Nyholm, D. Automatic Spiral Analysis for Objective Assessment of Motor Symptoms in Parkinson’s Disease. Sensors 2015, 15, 23727–23744. [Google Scholar] [CrossRef]
  139. Pham, T.D.; Wårdell, K.; Eklund, A.; Salerud, G. Classification of Short Time Series in Early Parkinsonʼ s Disease with Deep Learning of Fuzzy Recurrence Plots. IEEECAA J. Autom. Sin. 2019, 6, 1306–1317. [Google Scholar] [CrossRef]
  140. Matarazzo, M.; Arroyo-Gallego, T.; Montero, P.; Puertas-Martín, V.; Butterworth, I.; Mendoza, C.S.; Ledesma-Carbayo, M.J.; Catalán, M.J.; Molina, J.A.; Bermejo-Pareja, F.; et al. Remote Monitoring of Treatment Response in Parkinson’s Disease: The Habit of Typing on a Computer. Mov. Disord. 2019, 34, 1488–1495. [Google Scholar] [CrossRef]
  141. Aharonson, V.; Schlesinger, I.; McDonald, A.M.; Dubowsky, S.; Korczyn, A.D. A Practical Measurement of Parkinson’s Patients Gait Using Simple Walker-Based Motion Sensing and Data Analysis. J. Med. Devices Trans. ASME 2018, 12. [Google Scholar] [CrossRef]
  142. Pardoel, S.; Shalin, G.; Nantel, J.; Lemaire, E.D.; Kofman, J. Early Detection of Freezing of Gait during Walking Using Inertial Measurement Unit and Plantar Pressure Distribution Data. Sensors 2021, 21, 2246. [Google Scholar] [CrossRef] [PubMed]
  143. Wu, H.; Zhang, Y.; Wu, X.; Yang, F. Assessment of Upper Limb Tremors in Patients with Parkinson’s Disease Based on Displacement and Acceleration Information. In Proceedings of the 2020 5th International Conference on Automation, Control and Robotics Engineering (CACRE), Dalian, China, 19–20 September 2020; pp. 177–182. [Google Scholar]
  144. Cole, B.T.; Roy, S.H.; De Luca, C.J.; Nawab, S.H. Dynamical Learning and Tracking of Tremor and Dyskinesia from Wearable Sensors. IEEE Trans. Neural Syst. Rehabil. Eng. 2014, 22, 982–991. [Google Scholar] [CrossRef] [PubMed]
  145. Hossen, A.; Muthuraman, M.; Raethjen, J.; Deuschl, G.; Heute, U. A Neural Network Approach to Distinguish Parkinsonian Tremor from Advanced Essential Tremor. In Proceedings of the International Conference on Soft Computing for Problem Solving (SocProS 2011), AISC, Roorkee, India, 20–22 December 2012; Volume 130. [Google Scholar]
  146. Tahafchi, P.; Judy, J.W. Freezing-of-Gait Detection Using Wearable-Sensor Technology and Neural-Network Classifier. In Proceedings of the 2019 IEEE Sensors, Montreal, QC, Canada, 27–30 October 2019; pp. 1–4. [Google Scholar]
  147. Huo, W.; Angeles, P.; Tai, Y.F.; Pavese, N.; Wilson, S.; Hu, M.T.; Vaidyanathan, R. A Heterogeneous Sensing Suite for Multisymptom Quantification of Parkinson’s Disease. IEEE Trans. Neural Syst. Rehabil. Eng. 2020, 28, 1397–1406. [Google Scholar] [CrossRef] [PubMed]
  148. Yu, S.; Chen, H.; Brown, R.; Sherman, S. Motion Sensor-Based Assessment on Fall Risk and Parkinson’s Disease Severity: A Deep Multi-Source Multi-Task Learning (DMML) Approach. In Proceedings of the 2018 IEEE International Conference on Healthcare Informatics (ICHI), New York, NY, USA, 4–7 June 2018; pp. 174–179. [Google Scholar]
  149. Sajal, M.S.R.; Ehsan, M.T.; Vaidyanathan, R.; Wang, S.; Aziz, T.; Mamun, K.A.A. Telemonitoring Parkinson’s Disease Using Machine Learning by Combining Tremor and Voice Analysis. Brain Inform. 2020, 7, 12. [Google Scholar] [CrossRef]
  150. Oung, Q.W.; Muthusamy, H.; Basah, S.N.; Lee, H.; Vijean, V. Empirical Wavelet Transform Based Features for Classification of Parkinson’s Disease Severity. J. Med. Syst. 2018, 42, 1–17. [Google Scholar] [CrossRef]
  151. Papadopoulos, A.; Iakovakis, D.; Klingelhoefer, L.; Bostantjopoulou, S.; Chaudhuri, K.R.; Kyritsis, K.; Hadjidimitriou, S.; Charisis, V.; Hadjileontiadis, L.J.; Delopoulos, A. Unobtrusive Detection of Parkinson’s Disease from Multi-Modal and in-the-Wild Sensor Data Using Deep Learning Techniques. Sci. Rep. 2020, 10, 21370. [Google Scholar] [CrossRef]
  152. Heidarivincheh, F.; McConville, R.; Morgan, C.; McNaney, R.; Masullo, A.; Mirmehdi, M.; Whone, A.L.; Craddock, I. Multimodal Classification of Parkinson’s Disease in Home Environments with Resiliency to Missing Modalities. Sensors 2021, 21, 4133. [Google Scholar] [CrossRef]
  153. Wahid, F.; Begg, R.K.; Hass, C.J.; Halgamuge, S.; Ackland, D.C. Classification of Parkinson’s Disease Gait Using Spatial-Temporal Gait Features. IEEE J. Biomed. Health Inform. 2015, 19, 1794–1802. [Google Scholar] [CrossRef]
  154. Albani, G.; Ferraris, C.; Nerino, R.; Chimienti, A.; Pettiti, G.; Parisi, F.; Ferrari, G.; Cau, N.; Cimolin, V.; Azzaro, C.; et al. An Integrated Multi-Sensor Approach for the Remote Monitoring of Parkinson’s Disease. Sensors 2019, 19, 4764. [Google Scholar] [CrossRef] [Green Version]
  155. Joshi, A.; Ghosh, S.; Gunnery, S.; Tickle-Degnen, L.; Sclaroff, S.; Betke, M. Context-Sensitive Prediction of Facial Expressivity Using Multimodal Hierarchical Bayesian Neural Networks. In Proceedings of the 2018 13th IEEE International Conference on Automatic Face Gesture Recognition (FG 2018), Xi’an, China, 15–19 May 2018; pp. 278–285. [Google Scholar]
  156. Xu, S.; Pan, Z. A Novel Ensemble of Random Forest for Assisting Diagnosis of Parkinson’s Disease on Small Handwritten Dynamics Dataset. Int. J. Med. Inf. 2020, 144, 104283. [Google Scholar] [CrossRef] [PubMed]
  157. Gallicchio, C.; Micheli, A.; Pedrelli, L. Deep Echo State Networks for Diagnosis of Parkinson’s Disease. In Proceedings of the ESANN—European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, Bruges, Belgium, 25–27 April 2018; pp. 397–402. [Google Scholar]
  158. Pereira, C.R.; Weber, S.A.T.; Hook, C.; Rosa, G.H.; Papa, J.P. Deep Learning-Aided Parkinson’s Disease Diagnosis from Handwritten Dynamics. In Proceedings of the 2016 29th SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI), Sao Paulo, Brazil, 4–7 October 2016; pp. 340–346. [Google Scholar]
  159. Barth, J.; Sünkel, M.; Bergner, K.; Schickhuber, G.; Winkler, J.; Klucken, J.; Eskofier, B. Combined Analysis of Sensor Data from Hand and Gait Motor Function Improves Automatic Recognition of Parkinson’s Disease. In Proceedings of the 2012 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, San Diego, CA, USA, 28 August—1 September 2012; pp. 5122–5125. [Google Scholar]
  160. Schwab, P.; Karlen, W. PhoneMD: Learning to Diagnose Parkinson’s Disease from Smartphone Data. Proc. AAAI Conf. Artif. Intell. 2019, 33, 1118–1125. [Google Scholar] [CrossRef] [Green Version]
  161. Prince, J.; Andreotti, F.; De Vos, M. Multi-Source Ensemble Learning for the Remote Prediction of Parkinson’s Disease in the Presence of Source-Wise Missing Data. IEEE Trans. Biomed. Eng. 2019, 66, 1402–1411. [Google Scholar] [CrossRef]
  162. Cook, D.J.; Schmitter-Edgecombe, M.; Dawadi, P. Analyzing Activity Behavior and Movement in a Naturalistic Environment Using Smart Home Techniques. IEEE J. Biomed. Health Inform. 2015, 19, 1882–1892. [Google Scholar] [CrossRef] [PubMed]
  163. Habibzadeh, H.; Dinesh, K.; Rajabi Shishvan, O.; Boggio-Dandry, A.; Sharma, G.; Soyata, T. A Survey of Healthcare Internet of Things (HIoT): A Clinical Perspective. IEEE Internet Things J. 2020, 7, 53–71. [Google Scholar] [CrossRef] [PubMed]
  164. Barakat, B.; Taha, A.; Samson, R.; Steponenaite, A.; Ansari, S.; Langdon, P.M.; Wassell, I.J.; Abbasi, Q.H.; Imran, M.A.; Keates, S. 6G Opportunities Arising from Internet of Things Use Cases: A Review Paper. Future Internet 2021, 13, 159. [Google Scholar] [CrossRef]
  165. Srivastava, G.; Parizi, R.M.; Dehghantanha, A. The Future of Blockchain Technology in Healthcare Internet of Things Security. In Blockchain Cybersecurity, Trust and Privacy; Springer: Cham, Switzerland, 2020; pp. 161–184. [Google Scholar]
  166. Barth, J.; Oberndorfer, C.; Pasluosta, C.; Schülein, S.; Gassner, H.; Reinfelder, S.; Kugler, P.; Schuldhaus, D.; Winkler, J.; Klucken, J.; et al. Stride Segmentation during Free Walk Movements Using Multi-Dimensional Subsequence Dynamic Time Warping on Inertial Sensor Data. Sensors 2015, 15, 6419–6440. [Google Scholar] [CrossRef] [PubMed]
Figure 1. Flow diagram of the PRISMA methodology followed in the literature review.
Figure 1. Flow diagram of the PRISMA methodology followed in the literature review.
Sensors 22 01799 g001
Figure 2. Distribution of (a) studies retrieved per academic citation database and (b) studies considered per year.
Figure 2. Distribution of (a) studies retrieved per academic citation database and (b) studies considered per year.
Sensors 22 01799 g002
Figure 3. Distribution of the deployed types of sensors over the considered studies.
Figure 3. Distribution of the deployed types of sensors over the considered studies.
Sensors 22 01799 g003
Figure 4. Distribution of different devices deployed over the considered studies.
Figure 4. Distribution of different devices deployed over the considered studies.
Sensors 22 01799 g004
Figure 5. PD problems addressed based on inertial sensors and the respective obtained accuracies. (a) Number of studies per task; (b) accuracies obtained from the proposed models for each category of addressed tasks. Number of studies considered in this boxplot per task: n = 12 for diagnosis, n = 6 for differential diagnosis, n = 8 for severity estimation, n = 3 for fluctuations detection, n = 11 for FoG detection, n = 4 for tremor/bradykinesia/dyskinesia detection and n = 1 for medication response estimation.
Figure 5. PD problems addressed based on inertial sensors and the respective obtained accuracies. (a) Number of studies per task; (b) accuracies obtained from the proposed models for each category of addressed tasks. Number of studies considered in this boxplot per task: n = 12 for diagnosis, n = 6 for differential diagnosis, n = 8 for severity estimation, n = 3 for fluctuations detection, n = 11 for FoG detection, n = 4 for tremor/bradykinesia/dyskinesia detection and n = 1 for medication response estimation.
Sensors 22 01799 g005
Figure 6. PD problems addressed based on pressure sensors and the respective obtained accuracies. (a) Number of studies per task; (b) accuracies obtained from the proposed models for each category of addressed tasks. Number of studies considered in this boxplot per task: n = 4 for diagnosis, n = 1 for differential diagnosis and n = 5 for severity estimation.
Figure 6. PD problems addressed based on pressure sensors and the respective obtained accuracies. (a) Number of studies per task; (b) accuracies obtained from the proposed models for each category of addressed tasks. Number of studies considered in this boxplot per task: n = 4 for diagnosis, n = 1 for differential diagnosis and n = 5 for severity estimation.
Sensors 22 01799 g006
Figure 7. PD problems addressed based on image and depth sensors and the respective obtained accuracies. (a) Number of studies per task; (b) accuracies obtained from the proposed models for each category of addressed tasks. Number of studies considered in this boxplot per task: n = 7 for diagnosis, n = 3 for severity estimation, n = 2 for medication adherence estimation, n = 2 for subtasks segmentation, n = 1 for rehabilitation assessment, n = 1 for FoG detection and n = 1 for dyskinesia detection.
Figure 7. PD problems addressed based on image and depth sensors and the respective obtained accuracies. (a) Number of studies per task; (b) accuracies obtained from the proposed models for each category of addressed tasks. Number of studies considered in this boxplot per task: n = 7 for diagnosis, n = 3 for severity estimation, n = 2 for medication adherence estimation, n = 2 for subtasks segmentation, n = 1 for rehabilitation assessment, n = 1 for FoG detection and n = 1 for dyskinesia detection.
Sensors 22 01799 g007
Figure 8. PD problems addressed based on data collected via a combination of various types of sensors and the respective obtained accuracies. (a) Number of studies per task; (b) accuracies obtained from the proposed models for each category of addressed tasks. Number of studies considered in this boxplot per task: n = 9 for diagnosis, n = 1 for differential diagnosis, n = 5 for severity estimation and n = 1 for cognitive impairment detection.
Figure 8. PD problems addressed based on data collected via a combination of various types of sensors and the respective obtained accuracies. (a) Number of studies per task; (b) accuracies obtained from the proposed models for each category of addressed tasks. Number of studies considered in this boxplot per task: n = 9 for diagnosis, n = 1 for differential diagnosis, n = 5 for severity estimation and n = 1 for cognitive impairment detection.
Sensors 22 01799 g008
Figure 9. Frequency of PD-related problems studied independently of type of input data used.
Figure 9. Frequency of PD-related problems studied independently of type of input data used.
Sensors 22 01799 g009
Figure 10. Comparison between the best performing proposed models among all the different types of sensors (a) regarding diagnosis; (b) regarding severity estimation. In both cases, the horizontal axis refers to the indices of the studies.
Figure 10. Comparison between the best performing proposed models among all the different types of sensors (a) regarding diagnosis; (b) regarding severity estimation. In both cases, the horizontal axis refers to the indices of the studies.
Sensors 22 01799 g010
Figure 11. Distribution of the deployed ML algorithms over the considered studies. The blue bar depicts how many times each algorithm was tested, while the orange bar depicts how many times the same algorithm was selected as the best performing model or as a part of the final proposed solution. For the most widely tested algorithms, the ratio of the two previous values (grey bar) is also depicted (%).
Figure 11. Distribution of the deployed ML algorithms over the considered studies. The blue bar depicts how many times each algorithm was tested, while the orange bar depicts how many times the same algorithm was selected as the best performing model or as a part of the final proposed solution. For the most widely tested algorithms, the ratio of the two previous values (grey bar) is also depicted (%).
Sensors 22 01799 g011
Figure 12. Distribution of the number of the enrolled subjects across the studies considered. (a) Total population size across the studies considered in logarithmic scale-based bins. (b) The rate of PD patients to total subjects enrolled across the studies considered in 20-width bins.
Figure 12. Distribution of the number of the enrolled subjects across the studies considered. (a) Total population size across the studies considered in logarithmic scale-based bins. (b) The rate of PD patients to total subjects enrolled across the studies considered in 20-width bins.
Sensors 22 01799 g012
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share and Cite

MDPI and ACS Style

Giannakopoulou, K.-M.; Roussaki, I.; Demestichas, K. Internet of Things Technologies and Machine Learning Methods for Parkinson’s Disease Diagnosis, Monitoring and Management: A Systematic Review. Sensors 2022, 22, 1799. https://doi.org/10.3390/s22051799

AMA Style

Giannakopoulou K-M, Roussaki I, Demestichas K. Internet of Things Technologies and Machine Learning Methods for Parkinson’s Disease Diagnosis, Monitoring and Management: A Systematic Review. Sensors. 2022; 22(5):1799. https://doi.org/10.3390/s22051799

Chicago/Turabian Style

Giannakopoulou, Konstantina-Maria, Ioanna Roussaki, and Konstantinos Demestichas. 2022. "Internet of Things Technologies and Machine Learning Methods for Parkinson’s Disease Diagnosis, Monitoring and Management: A Systematic Review" Sensors 22, no. 5: 1799. https://doi.org/10.3390/s22051799

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop