A Computerized Analysis with Machine Learning Techniques for the Diagnosis of Parkinson’s Disease: Past Studies and Future Perspectives

According to the World Health Organization (WHO), Parkinson’s disease (PD) is a neurodegenerative disease of the brain that causes motor symptoms including slower movement, rigidity, tremor, and imbalance in addition to other problems like Alzheimer’s disease (AD), psychiatric problems, insomnia, anxiety, and sensory abnormalities. Techniques including artificial intelligence (AI), machine learning (ML), and deep learning (DL) have been established for the classification of PD and normal controls (NC) with similar therapeutic appearances in order to address these problems and improve the diagnostic procedure for PD. In this article, we examine a literature survey of research articles published up to September 2022 in order to present an in-depth analysis of the use of datasets, various modalities, experimental setups, and architectures that have been applied in the diagnosis of subjective disease. This analysis includes a total of 217 research publications with a list of the various datasets, methodologies, and features. These findings suggest that ML/DL methods and novel biomarkers hold promising results for application in medical decision-making, leading to a more methodical and thorough detection of PD. Finally, we highlight the challenges and provide appropriate recommendations on selecting approaches that might be used for subgrouping and connection analysis with structural magnetic resonance imaging (sMRI), DaTSCAN, and single-photon emission computerized tomography (SPECT) data for future Parkinson’s research.


Introduction
Parkinson's disease, commonly known as Tremor, is affected by a diminution in dopamine levels in the brain, which damages a person's motion functions, or physical functioning. It is one of the world's most common diseases. Intermittent neurological signs and symptoms result from these lesions, which become worse as the disease progresses [1]. Because aging causes changes in our brains, such as loss of synaptic connections and changes in neurotransmitters and neurohormones, this condition is more frequent among elders. With the passage of time, the neurons in a person's body begin to die and become inimitable. The consequences of neurological problems and the falling dopamine levels in Another issue is that, typically, the diagnosis of PD requires a number of steps, including gathering a thorough neurological history from the patient and examining their motor abilities in various environments.
The main purpose of this study is to summarize and assess the review of AI algorithms, data acquisition methods, and applications of AI in the diagnosis of subjective diseases and challenges. The majority of recent studies deal with the homo dataset (text, speech, video, or image). Many researchers applied voice data since they can only use so much data (single data type). Problems with dataset modification and multi-data handling procedures have been highlighted in the suggested study. The effectiveness of disease prediction is regulated as a result of the examination of a particular dataset. More real-time solutions are made possible by the use of ML-based techniques for multivariate data processing. Multi-variatevocal data analysis (MVDA) is driven to provide multiple dataset attributebased PD identification utilizing ML approaches. This study examines the potential for improving multi-variate and multimodal data processing, which aids in raising the disease detection rate. The existing research simultaneously concentrated on various ML-based such as support vector machines (SVM), naïve Bayes (NB), K nearest neighbor (K-NN), and artificial neural network (ANN) evaluations of Parkinson's data based on voice features. A larger number of patients were selected for the study of Parkinson's data in the experimental works of current systems. The MVDA employs extensive datasets and ML approaches to improve disease identification based on these works. The incorporation of numerous patients' multi-variate acoustic characteristics in the proposed MVDA is encouraged. The subjective disease has been diagnosed with the help of proposed ML techniques under the MVDA system.

Contribution
This research article covers the techniques of ML which are implemented in the auditory analysis of speech to diagnose this disease. The benefits and shortcomings of these algorithms in detecting the disease are thoroughly contrasted, and existing comparative studies' potential drawbacks are explored. The main contribution of this paper is as follows: a.
In this paper, we reviewed the significant statistics and relevant information collected from 217 articles (from various resources) published from 2015-2022 on the diagnosis and classification of PD. b.
The fundamental discussion on AI and ML with their significance in the field of medical healthcare. c.
In order to improve the prediction of PD, we also present recommendations for future perspectives to help researchers and scholars in recognizing various plausible paths for them to work in the future.

Structure of Proposed Work
The structure of the study is as follows (Figure 1): Section 2 describes the methods for the literature search strategy. Section 3 discusses the overview of AI, ML, and DL techniques. Section 4 defines an overview of Parkinson's disease. Section 5 illustrates the state of the art. Section 6 discusses the current limitations as challenges and future perspectives as recommendations. Finally, Section 7 defines the conclusion.

Methods
The present literature evaluation was carried out in a systematic manner that is generally in accordance with the most recent PRISMA criteria [40] as shown in Figure 2. By using the PRISMA technique, the authors can easily evaluate various studies as well as make decisions about the criteria for the final selection of the studies, the search strategy and data sources, and inclusion and the exclusion process.

Methods
The present literature evaluation was carried out in a systematic manner that is generally in accordance with the most recent PRISMA criteria [40] as shown in Figure 2. By using the PRISMA technique, the authors can easily evaluate various studies as well as make decisions about the criteria for the final selection of the studies, the search strategy and data sources, and inclusion and the exclusion process.

Literature Search Strategy
Firstly, the most appropriate literature databases were chosen to review the research article. Web of Science, Google Scholar, PubMed, Scopus, and IEEE are five significant databases that were examined to determine and screen relevance. A total of 303 records were identified through a web of science database searching by using the keyword "Parkinson's Disease and Artificial Intelligence" where 147 were review articles, 40 were proceeding papers, 7 were from early access, 109 were from open access, and 312 articles were found through other sources.
For choosing the relevant articles, the most appropriate keywords were selected, to ensure that papers that would address the presented research topics were included. The keywords such as PD, AI, ML, and ANN are combined using the logical expression "AND". After making multiple revisions to assure the inclusion of all the methodologies and techniques, the final technical keywords were chosen based on various approaches covered in earlier review articles. The articles considered in this study were published in journals, book chapters, abstract meetings, or conference proceedings from January 2015 to September 2022, and all the articles are written in English. After applying the logical operation in the keywords ("Parkinson's Disease" AND (Artificial Intelligence AND Machine Learning AND Artificial Neural Network OR Deep Learning) AND (gait analysis AND voice AND rigidity AND olfactory)) the total number of 33 research papers

Literature Search Strategy
Firstly, the most appropriate literature databases were chosen to review the research article. Web of Science, Google Scholar, PubMed, Scopus, and IEEE are five significant databases that were examined to determine and screen relevance. A total of 303 records were identified through a web of science database searching by using the keyword "Parkinson's Disease and Artificial Intelligence" where 147 were review articles, 40 were proceeding papers, 7 were from early access, 109 were from open access, and 312 articles were found through other sources.
For choosing the relevant articles, the most appropriate keywords were selected, to ensure that papers that would address the presented research topics were included. The keywords such as PD, AI, ML, and ANN are combined using the logical expression "AND". After making multiple revisions to assure the inclusion of all the methodologies and techniques, the final technical keywords were chosen based on various approaches covered in earlier review articles. The articles considered in this study were published in journals, book chapters, abstract meetings, or conference proceedings from January 2015 to September 2022, and all the articles are written in English. After applying the logical operation in the keywords ("Parkinson's Disease" AND (Artificial Intelligence AND Machine Learning AND Artificial Neural Network OR Deep Learning) AND (gait analysis AND voice AND rigidity AND olfactory)) the total number of 33 research papers were found from various reputed journals as shown in Figure 3. After refining the articles from Diagnostics 2022, 12, 2708 6 of 32 the web of science core collection, the next search query was submitted to the Google scholar database by using the same pattern and search strategy from 2015 to 2022 and a total 67 results were shorted out. By using the same pattern and strategy, the queries were submitted to the rest of the databases. were found from various reputed journals as shown in Figure 3. After refining the articles from the web of science core collection, the next search query was submitted to the Google scholar database by using the same pattern and search strategy from 2015 to 2022 and a total 67 results were shorted out. By using the same pattern and strategy, the queries were submitted to the rest of the databases.

Research Field: Journals
In this section, we organize a methodical study of the diagnosis of subjective disease using the AI/ML/DL approaches by analyzing 246 articles published during 2015-2022. The prominent journals publishing on the diagnosis of PD using different AI and ML approaches are shown in Table 1. In order to conduct this analysis, we analyzed the research papers from the MDPI (81), IEEE (86), Frontiers Media SA (31), PLOS (23), Nature (9), Springer (6), Hindawi (5), Nature Portfolio (3), and Elsevier (2), as shown in Figure 4. The research that has been included in this review shows that relevant information about the motor and non-motor symptoms of PD may be retrieved using feature selection approaches with the aid of ML algorithms, enabling clinicians to make decisions based on the given dataset. Table 2 summarizes the scope of the review article to diagnose PD (2015-2022) from the Web of Science Database.

Research Field: Journals
In this section, we organize a methodical study of the diagnosis of subjective disease using the AI/ML/DL approaches by analyzing 246 articles published during 2015-2022. The prominent journals publishing on the diagnosis of PD using different AI and ML approaches are shown in Table 1. In order to conduct this analysis, we analyzed the research papers from the MDPI (81), IEEE (86), Frontiers Media SA (31), PLOS (23), Nature (9), Springer (6), Hindawi (5), Nature Portfolio (3), and Elsevier (2), as shown in Figure 4. The research that has been included in this review shows that relevant information about the motor and non-motor symptoms of PD may be retrieved using feature selection approaches with the aid of ML algorithms, enabling clinicians to make decisions based on the given dataset. Table 2 summarizes the scope of the review article to diagnose PD (2015-2022) from the Web of Science Database.

Overview of Artificial Intelligence, Machine Learning, and Deep Learning
This section will provide a basic overview of how AI, ML, and DL work, the most popular AI, ML, and DL algorithms, and the various technologies that may be used to gather data to feed into these algorithms.
(a) Artificial Intelligence AI is the area of computer science that aims to reproduce human intellectual abilities in machines, especially computer systems. Some particular applications of AI include expert systems, ML, speech recognition, and natural language processing (NLP) including mundane tasks, formal tasks, and expert tasks [68]. Figure 5 represents the role of AI in medical healthcare.

Overview of Artificial Intelligence, Machine Learning, and Deep Learning
This section will provide a basic overview of how AI, ML, and DL work, the most popular AI, ML, and DL algorithms, and the various technologies that may be used to gather data to feed into these algorithms.
(a) Artificial Intelligence AI is the area of computer science that aims to reproduce human intellectual abilities in machines, especially computer systems. Some particular applications of AI include expert systems, ML, speech recognition, and natural language processing (NLP) including mundane tasks, formal tasks, and expert tasks [68]. Figure 5 represents the role of AI in medical healthcare. For the past 50 years, AI in healthcare has been primarily focused on the diagnosis and treatment of diseases. Early rule-based systems might have properly diagnosed and treated diseases, but they were not entirely adopted in clinical practice. In addition to having a less-than-perfect connection with clinical workflows and health record systems, they were not significantly more accurate at diagnosing than humans.

(b) Machine Learning
ML is a subset of AI in which the machine constructs a prediction model using historical data from its past experiences, predicts the outcome for new data, and becomes For the past 50 years, AI in healthcare has been primarily focused on the diagnosis and treatment of diseases. Early rule-based systems might have properly diagnosed and treated diseases, but they were not entirely adopted in clinical practice. In addition to having a less-than-perfect connection with clinical workflows and health record systems, they were not significantly more accurate at diagnosing than humans.

(b) Machine Learning
ML is a subset of AI in which the machine constructs a prediction model using historical data from its past experiences, predicts the outcome for new data, and becomes better at doing so. It is different from traditional programming. In traditional programming, rules are not explicitly learned from the data; rather, they are written in a computer language. Unlike traditional programming, ML creates predictive models using data that are then applied to predictions using data that have not yet been seen. Due to the intricacy of the code, it might be highly challenging to design a rule-based program for some problems. In these situations, ML can be employed if there are enough data available that is pertinent to the problem under consideration [69]. ML can be classified into supervised learning (SL), unsupervised learning (UL), and reinforcement learning (RL) as shown in Figure 6. In SL, sample-labeled data are given to the ML model as a training dataset, and on the basis of that, it predicts the outcome [70]. In UL, the ML model is trained with a collection of unlabeled, unclassified, or uncategorized data, and the algorithm is required to respond independently to that data. The main objective of UL is to reorganize the input data into new features or a collection of objects with related patterns [71]. RL is a form of ML approach where an intelligent agent (computer program) interacts with the environment and learns to function within that. By accumulating the greatest benefits over time, the goal of RL is to identify the "policy" that works best. The policy decides what should be conducted in a certain circumstance [72].
is pertinent to the problem under consideration [69]. ML can be classified into supervised learning (SL), unsupervised learning (UL), and reinforcement learning (RL) as shown in Figure 6. In SL, sample-labeled data aregiven to the ML model as a training dataset, and on the basis of that, it predicts the outcome [70]. In UL, the ML model is trained with a collection of unlabeled, unclassified, or uncategorized data, and the algorithm is required to respond independently to that data. The main objective of UL is to reorganize the input data into new features or a collection of objects with related patterns [71]. RL is a form of ML approach where an intelligent agent (computer program) interacts with the environment and learns to function within that. By accumulating the greatest benefits over time, the goal of RL is to identify the "policy" that works best. The policy decides what should be conductedin a certain circumstance [72].

(c) Deep Learning
Deep structured learning is a subfield of ML methods based on ANN with representation learning. DL has enormous potential in the fields of healthcare and medicine due to the sheer amount of data being produced (150 exabytes, or 10 18 bytes, in the United States alone, expanding 48% annually) as well as the rising number of medical equipment and electronic medical record systems [73]. DL algorithms are often effective with higher dimensional data, such as audio, video, and images. DL algorithms are created dynamically to run across several layers of neural networks, which are nothing more than a collection of decision-making networks that have been pre-trained to do a certain task. Each of them is then moved on to the following layer after being run through basic layered representations. On datasets containing hundreds of features or columns, however, most ML techniques are enhanced to perform quite well. ML often fails to detect a straightforward image with dimensions of 800×1000 in RGB, regardless of how organized or unstructured the data set is. A standard ML system would find handling such depths to be

(c) Deep Learning
Deep structured learning is a subfield of ML methods based on ANN with representation learning. DL has enormous potential in the fields of healthcare and medicine due to the sheer amount of data being produced (150 exabytes, or 10 18 bytes, in the United States alone, expanding 48% annually) as well as the rising number of medical equipment and electronic medical record systems [73]. DL algorithms are often effective with higher dimensional data, such as audio, video, and images. DL algorithms are created dynamically to run across several layers of neural networks, which are nothing more than a collection of decision-making networks that have been pre-trained to do a certain task. Each of them is then moved on to the following layer after being run through basic layered representations. On datasets containing hundreds of features or columns, however, most ML techniques are enhanced to perform quite well. ML often fails to detect a straightforward image with dimensions of 800 × 1000 in RGB, regardless of how organized or unstructured the data set is. A standard ML system would find handling such depths to be rather impractical [74]. Recently DL techniques have been introduced for the automatic detection and categorization of PD using speech patterns and handwriting patterns recorded by a smart pen [75][76][77][78]. Examples of DL include CNN and ANN. ANNs, which are networks of computing units that mimic the functioning of biological neural networks, are frequently employed for a variety of tasks, including classification, regression, and time series analysis. They are made up of several processing layers, with the input samples being held in the first layer and the prediction being provided in the last layer. Additionally, CNN is an adaptation of ANN that uses spatial filters (convolutional layers) to extract the textures, patterns, and intrinsic properties of images [79]. By subsampling the features obtained by the convolutional layers, pooling is also used to provide reliable features as shown in Figure 7. The main advantage of the aforementioned techniques is their capacity to automatically learn EEG parameters and detect anomalies based on those features. time series analysis. They are made up of several processing layers, with the input samples being held in the first layer and the prediction being provided in the last layer. Additionally, CNN is an adaptation of ANN that uses spatial filters (convolutional layers) to extract the textures, patterns, and intrinsic properties of images [79]. By subsampling the features obtained by the convolutional layers, pooling is also used to provide reliable features as shown in Figure 7. The main advantage of the aforementioned techniques is their capacity to automatically learn EEG parameters and detect anomalies based on those features.

Parkinson's Disease: An Overview
The brain's substantianigra area experiences nerve cell declension, which causes PD. A neurotransmitter called dopamine, which was created by nerve cells is produced by this region of the brain. Dopamine's function is to serve as a connection between the brain and the parts of the sensory organs that control and direct bodily movements. [80]. When these neurons die or are harmed, there is less dopamine in the brain. This suggests that the area of the brain responsible for controlling movement is dysfunctional, which causes sluggish, undesired, and erratic motions of the physical parts [81]. Nerve cell degeneration happens gradually. PD symptoms start to manifest after around 80% of the nerve cells in the substantianigra are destroyed [82]. Currently, a combination of ecological factors and genetic abnormalities are thought to be the disease's etiology. Although some inherited factors have been shown to increase a person's chance of developing PD, it is unclear how these factors make some people more sensitive to the disease [83]. PD can occur in families because the dysfunctional genes are passed down from parents to children. But given the condition, this is an uncommon type of legacy. Some specialists claim that ecological factors may potentially increase a person's chance of developing PD [84]. The AI-based algorithm may classify people as having PD or not (non-PD) based on their motor symptoms (risk factors). The training model may be developed using the dataset, which was created while evaluating the patients. Numerous PD risk factors, including both motor and non-motor variables, are formed. Although the symptomatic data cannot be statistically resolved, it is possible to improve PD detection by using an ML or DL to better grasp both data classes [85]. In-feature AI is the best choice to accurately predict PD while examining the symptomatic biology of the disease. Due to the

Parkinson's Disease: An Overview
The brain's substantianigra area experiences nerve cell declension, which causes PD. A neurotransmitter called dopamine, which was created by nerve cells is produced by this region of the brain. Dopamine's function is to serve as a connection between the brain and the parts of the sensory organs that control and direct bodily movements [80]. When these neurons die or are harmed, there is less dopamine in the brain. This suggests that the area of the brain responsible for controlling movement is dysfunctional, which causes sluggish, undesired, and erratic motions of the physical parts [81]. Nerve cell degeneration happens gradually. PD symptoms start to manifest after around 80% of the nerve cells in the substantianigra are destroyed [82]. Currently, a combination of ecological factors and genetic abnormalities are thought to be the disease's etiology. Although some inherited factors have been shown to increase a person's chance of developing PD, it is unclear how these factors make some people more sensitive to the disease [83]. PD can occur in families because the dysfunctional genes are passed down from parents to children. But given the condition, this is an uncommon type of legacy. Some specialists claim that ecological factors may potentially increase a person's chance of developing PD [84]. The AI-based algorithm may classify people as having PD or not (non-PD) based on their motor symptoms (risk factors). The training model may be developed using the dataset, which was created while evaluating the patients. Numerous PD risk factors, including both motor and non-motor variables, are formed. Although the symptomatic data cannot be statistically resolved, it is possible to improve PD detection by using an ML or DL to better grasp both data classes [85]. In-feature AI is the best choice to accurately predict PD while examining the symptomatic biology of the disease. Due to the progressive nature of PD, its severity and status have been assessed using the following subjective and varied assessment systems [86][87][88][89][90][91]-MoCA, screening questionnaire, GDS, RBD, UPDRS, STAI, PIGD score, SCOPA-AUT, and MMSE, [92]. Figure 8 depicts the category of PD.  Parkinsonian gait or festinating gait is the type of gait exhibited by patients with PD. Parkinson's patients frequently experience feeling as though they are "locked in place" when taking a stride or turning, which raises the possibility of falling. This disorder is brought on by a dopamine shortage in the basal ganglia circuit, which results in motor Parkinsonian gait or festinating gait is the type of gait exhibited by patients with PD. Parkinson's patients frequently experience feeling as though they are "locked in place" when taking a stride or turning, which raises the possibility of falling. This disorder is brought on by a dopamine shortage in the basal ganglia circuit, which results in motor impairments. Despite the fact that PD symptoms might vary, gait is one of the motor features of this disorder that is most commonly impaired. Small, shuffling steps, general slowness of movement (hypokinesia), and in severe cases, complete immobility (akinesia), are the characteristic of Parkinsonian gait [93][94][95]. In contrast to longer double supports, PD patients had shorter strides, slower walking speeds during spontaneous ambulation, and higher cyclic rates [96][97][98][99].
Another symptom of PD is a speech disorder. PD patients may stumble over their words, whisper, or falter toward the end of sentences. Whereas most people speak slowly, others do so quickly and even stutter or stammer. Speech issues may be intensified by PD motor symptoms such as less facial expression, slowness, and hunched posture. Deep brain stimulation, surgery, speech therapy, pharmaceutical intervention, and vocal fold augmentation are a few of the therapeutic approaches. Speech therapy for PD patients should be provided as part of a multidisciplinary approach to patient care, despite the fact that managing Parkinsonian dysarthria is clinically difficult [100].
Another feature of PD is small, cramped handwriting called micrographia, which is typically one of the early signs. Micrographia is a neurological condition that results in words that are often small and clustered together as well as other movement symptoms of the disorder. Additionally, symptoms including rigidity, tremor, and slowness of movement can all make it more difficult to write [101].
Recent research has shown that MRI can be used to detect and diagnose PD far sooner than conventional techniques. In order to detect PD, MRIs scan the brain for particular markers. These signs are frequently present even before Parkinson's symptoms appear [102].

Medical Approaches for Parkinson's Disease Diagnosis
Most cases of PD are identified based on their clinical symptoms. An X-ray or blood test cannot verify the disease. Nevertheless, non-invasive diagnostic imaging, such as positron emission tomography (PET), can help a surgeon make a diagnosis. Conventional methods for diagnosing Parkinsonism include the presence of two or more primary symptoms, the absence of additional neurological symptoms upon examination, the absence of a history of additional potential causes, such as the use of anesthetic drugs, head trauma, or stroke, and responsiveness to levodopa or other Parkinson's medications [66]. Following are some clinical methods that are used to diagnose PD: a.
Medical Treatment In most cases, medication is used to treat Parkinson's patients in order to reduce their disease symptoms. Levodopa drugs or anticholinergic pharmaceuticals stimulate the residual substantia nigra cells to create further dopamine whereas levodopa medications suppress part of the acetylcholine production, which restores the homeostasis of the brain's chemical production. There are a wide variety of side effects associated with each medication class [103]. Levodopa, which was created more than four decades ago, is frequently referred to as the standard of Parkinson's treatment. Levodopa is used in lower doses in order to reduce the symptoms. This development significantly lessens acute vomiting and nausea that are frequently encountered as levodopa side effects. Levodopa often lessens the tremor, stiffness, and slowness symptoms in individuals. Patients with a lack of spontaneous movement and muscular stiffness benefit the most from it [104].

b. Dopamine Agonists
The brain's chemical messengers are imitated by drugs like bromocriptine, pergolide, pramipexole, and ropinirole, which cause neurons to respond as they would to dopamine. Medications can be prescribed either alone or in combination with levodopa, and they can be given to patients in the early stages of the illness or to extend the time that levodopa will be effective. Before recommending dopamine agonists to patients, surgeons take into account the fact that these drugs often have greater adverse effects than levodopa [103].

c. COMT Inhibitors
Inhibitors of catechol-O-methyl transferase (COMT) are among the amino groups that contribute to the stability of levodopa levels. Entacapone, tolcapone, and opi-Capone are the three main COMT inhibitors. These medications work by inhibiting the COMT enzyme, which raises blood levels of levodopa without causing it to be peripherally degraded into 3-O-methyldopa (3-OMD) [105]. Dyskinesia and diarrhea may include possible side effects [103].

d. Selegiline
Monoamine oxidase B (MAO-B) is selectively inhibited by the drug selegiline. The major enzyme responsible for the catabolism of dopamine is MAOs, which are intracellular enzymes found on the mitochondrial membrane. It has been established that selegiline may be used safely in PD patients due to its ability to alleviate symptoms when taken as monotherapy, postpone the onset of levodopa, enhance wearing-off in individuals with motor irregularities, and perhaps even have neuroprotective effects. Selegiline may be considered an ancient medication, whereas rasagiline, a more contemporary MAO-B inhibitor, is equivalent in terms of effectiveness [106]. e.
Anticholinergic medications The function of the neurotransmitter acetylcholine (ACh) in the central and autonomic nervous systems is blocked by anticholinergic medicines, which leads to a wide range of both beneficial and undesirable consequences. Since many of the most often given medications for seniors are indicated for problems common to the aged, one-third to one-half of these medications contain anticholinergic effects [107]. These medications are particularly effective in treating tremors, stiffness of the muscles, and antidepressants for Parkinsonism. Due to difficulties and major adverse effects, they are typically not advised for prolonged use in elderly individuals [103].

f. Amantadine
Levodopa-related dyskinesia is typically treated with amantadine as an add-on medication, although more recently, novel long-acting amantadine formulations have been created with additional indications to treat motor fluctuations. Amantadine is hardly associated to impulse control problems and has not been found to produce dyskinesia [108]. Levodopa or anticholinergic medicine may occasionally be used with amantadine. Some of its adverse effects include confusion, sleeplessness, nightmares, irritability, and hallucinations. It may also cause leg swelling [103].

State of the Art
In order to distinguish PD cases from healthy controls, a variety of modern ML algorithms, including SVM, ANN, logistic regression, naïve Bayes, etc., were successfully used. In this study, numerous databases, including Web of Science, Elsevier, MDPI, Scopus, Science Direct, IEEE Xplore, Springer, and Google Scholar, were utilized to survey relevant papers on PD.

Literature Review Based on Speech, Gait, and Handwriting Patterns to Diagnose PD
E. Avuçlu et al. [109] proposed a method to detect PD using multiple classifiers. In their study, the authors used 195 sound samples and 22 acoustic vocal characteristics in a variety of 75% training and 25% of test data. The ML classifiers used to detect PD are naive Bayes, random forest, SVM, and KNN. According to research, the SVM accurately diagnoses PD patients with a test data accuracy of 67.27% and a training data accuracy of 87.06%. In a survey by KarimiRouzbahani, H et al. [110], the authors used KNN, SVM, and discrimination-based-function (DBF) classifiers for the diagnosis of PD. In their study, they used several parameters like jitter, fundamental frequency, pitch, shimmer, and other statistical measures. The best accuracy among these classifiers was obtained from KNN with a 93.83% accuracy rate and it also provides good performance in other parameters like sensitivity, specificity, and error rate also.
The authors Khamparia, A. et al. [111] used a CNN classifier applied to speech classification datasets. The accuracy reached throughout the training phase, which is over 77%, makes the results optimistic. In accordance with the works mentioned above, A. Bourouhou et al. [112] examined a variety of classifiers to identify individuals who were likely to have PD. They used 40 participants for their investigation, including 20 subjective patients and 20 NC. According to the experimental findings, the naive Bayes classifier has a detection accuracy, sensitivity rate, and specificity rate are 65%, 68%, and 66.6%, respectively. Sharma, A. et al. [113] used three types of classifiers based on KNN, SVM, and multilayer perceptron (MLP) to diagnose PD. Among all these ML classifiers, SVM using an RBF kernel outperformed with an overall classification accuracy rate of 85.294% percent.
A summary of the most recent DL methods for audio signal processing is given in another work by Purwins, H. et al. [114]. The works that have been examined include CNN as well as other long short-term memory (LSTM) architecture models and audiospecific neural network models. Similar to the previous studies, L. Zhang, Y. Qu, et al. [115] detected PD using naive Bayes and other ML approaches. In their method, relevant features were extracted from the voice signal of PD patients and healthy control (HC) subjects using signal processing techniques. The naive Bayes algorithm shows a 69.24% detection accuracy and 96.02% precision rate for the 22 voice characteristics. S. R. Kadiri et al. [116] suggested a technique for detecting PD using SVM on shifted delta cepstral (SDC) and single frequency filtering cepstral coefficients (SFFCC) features extracted from speech signals of PD patients and HC. Comparing the standard MFCC + SDC features with the SDC + SFFCC features, performance increases of 9% were observed. A 73.33% detection accuracy with a 73.32% F1-score is displayed by the conventional SVM on SDC + SFCC features. In addition to the naive Bayes classifier, several additional supervised methods, including but not restricted to well-known DL methods, have been suggested to identify PD patients among healthy controls.
In a survey conducted by M. Pramanik et al. [117], the authors examined two recognizing decision forests, i.e., SysFor and ForestPA, along with the most widely used random forest classifier, which has been utilized as a Parkinson's detector. In their study, as compared with SysFor and ForestPA, random forest's average detection accuracy on incremental trees shows 93.58%. For the purpose of classifying PD through sets of acoustic vocal (voice) characteristics, Gunduz, H. [118] suggested two frameworks based on CNN. Both frameworks are used for the mixing of different feature sets, although they combine feature sets in different ways. Although the second framework provides feature sets to the parallel input levels that are directly connected to convolution layers, the first framework first combines several feature sets before passing them as inputs to the nine-layered CNN.
One of the most important technological advancements of the twenty-first century has been the use of AI and ML in society. Basic AI systems were in use in the late 20th century, but during the past ten years, the creation of procedures and systems that employ ML and other functions has risen tremendously. These tools help researchers in a huge range of areas manage their data and work more efficiently. Clinical insights continue to employ AI and ML in a variety of ways. The fact that AI technologies don't only focus on one component of clinical findings is one of their strongest features. For handling massive and heterogeneous data sources, spotting complicated and hidden patterns, and forecasting difficult outcomes, many ML algorithms are available. Because of this, ML has much to offer in terms of clinical insights across the board, from preclinical drug development to pre-trial planning to study execution to data storage and analysis [119].
AI is assisting physicians in better diagnosing and treating diseases like postoperative hypotension, and more advanced future models may have even more widespread medical uses. The evolutionary step in the creation of therapeutic pathways and adherence is ML. The real benefit of ML, however, is that it enables provider organizations to use information about the patient population from their own systems of record to create therapeutic pathways that are unique to their procedures, clientele, and physicians [120].
It can be seen from the reviews above that all the research has been conductedand is only restricted to a small number of datasets. The above previous works inspired us to try a new methodology. In this study, we experimented with several feature selection methods before comparing the results with various ML classifiers. Table 3 represents the review of AI/ML/DL techniques used to diagnose major symptoms of PD i.e., speech recording, handwriting pattern, and gait features for 20 studies.

Literature Review on Neuroimagingto Diagnose PD
Since neuroimaging has demonstrated its efficacy in the diagnosis of PD, CAD that is based on neuroimaging has received a lot of attention. The classifier module is one of the important components of a CAD system that directly affects classification performance [141].
Chakraborty, S. et al. [142] discussed that a total of 906 people participated in the survey, of whom 203 served as controls, 66 as prodromal subjects, and 637 as symptoms of PD. Eight subcortical regions were separated from the obtained MRI scans by using atlasbased segmentation in order to examine the MRI scans for the diagnosis of the subjective disease. In addition, morphological, textural, and statistical information were recovered from the eight extracted subcortical structures using feature extraction. For each MRI scan, an exhaustive collection of 107 features was produced after the feature extraction procedure. In order to determine the optimal feature set for the identification of PD, a two-level feature extraction technique was used.
Bhan et al. [143] proposed a deep-learning methodology to diagnose PD among healthy controls and subjective disease. According to a study, taking the right actions early greatly improves the likelihood of healing, and using a machine to carry out the detection process might save a lot of time. The MRI data of PD participants were effectively separated from healthy controls using the CNN and the LeNet-5 architecture.
A methodology was suggested by Kumar, R. et al. [144] to use a discrete wavelet transform-based fusion of MRI sequences and radiomics feature extraction as the approach for a novel framework for classifying brain tumors. The performance evaluation of the authors' method was conducted using the Brain Tumor Segmentation 2018 Challenge training dataset, and features were taken from three areas of interest created by combining several tumor regions. The authors employed various ML classifiers to train the model. They also used filter and wrapper method-based feature selection strategies to choose a meaningful collection of features.
In a survey conducted by Pang, Y. et al. [145], the authors assessed a hand and finger motion capture wearable device that is basically used to record the hand and finger motion of HC and subjective disease patients. Using the DWT, the specific three-dimensional motion properties of each finger joint were recovered. By examining the motion variations in the frequency domain on four types of motion from 5 subjective patients and 22 healthy control subjects, the degree of tremor for each finger joint was measured.
According to A. Radziunas, et al. [146], the authors examined 28 PD patients for sleep problems using the PDSS and underwent brain MRIs conducted on 14 males and 14 females, all of whom were between the ages of 58. Using the FreeSurfer program, automated vowel-based image analysis was carried out.
D. Zhang, J. et al. [147] used multivariate pattern analysis to distinguish between subjective patients and NC by using the characteristics of the inconsistency of tremor using a multiple linear regression model. For this experiment, the data were collected from the 36 participants where 16 were affected by PD and 20 matched healthy controls. For each person, wavelet-based functional and morphological brain networks were then built. According to graph-based network analysis, individuals with PD had a disruption in information translation efficiency within the wavelet scale 2 of the functional brain network.
Magesh, P. R. et al. [149] suggested a CNN-based regression approach for differentiating between subjective patients and NC. Data for 252 patients were obtained for this study from the PPMI database. The trained network was tested using ten-fold cross-validation, and the performance parameter was the absolute difference between predicted and actual scores. Evaluation of prediction using inputs with and without DAT images.
Mabrouk, R. et al. [150] proposed five models of ML for distinguishing PD patients and HC using clinical evaluation and image-based features applied later on in the SWEDD group as a potential application of motor and non-motor data in understanding PD characteristics. In binary classification, the five models had a high degree of accuracy (75.4-78.4% for motor characteristics and 71-82.2% for non-motor data). In this manner, the authors have shown how ML models may be used to binary classify SPECT data, proving their applicability and utility.
The study proposed by Quan, J. et al. [151], demonstrates a deep-CNN methodology and assesses the effectiveness of the method for categorizing DaTSCAN SPECT images. The InceptionV3 architecture, which placed second in the 2015 ImageNet Large Scale Visual Recognition Competition (ILSVRC), serves as the foundation model for the deep neural network used in this study. On top of this basis, a unique, binary classifier block was created. The effectiveness of the model was assessed using ten-fold cross-validation in order to adjust for the short dataset size.
In accordance with the aforementioned studies, the authors Moon, S. et al. [152] proposed a number of ML methods, including an SVM, decision tree, gradient boosting, and neural network for the diagnosis of PD patients using an F1-score dummy model. For this study, authors used balance and gait variables collected during the instrumented stand and walk test from people with 524 PD patients and 43 essential tremors (ET).
In the study by Adams, M. P. et al. [153], the authors created a method based on CNN that predicts clinical motor function evaluation scores from longitudinal DAT SPECT images and clinical measurements that are not imaging-based.
In line with the above works, Khachnaoui, H. et al. [154] suggested an ML methodology used to differentiate PD patients from HC within a SWEDD group. The authors analyzed data from 548 participants using principal component analysis (PCA) and linear discriminant analysis (LDA) methods. The authors developed density-based spatial (DBSCAN), K-means, and hierarchical clustering using the results of the best reduction approach. In terms of accuracy, sensitivity, and specificity, hierarchical clustering outperformed DBSCAN and K-means algorithms by 64%, 78.13%, and 38.89%, respectively. The suggested approach showed that ML models could successfully separate PD patients from HC participants within a SWEDD group.
As stated by Oliveira, F. P. et al. [155], the authors aimed to evaluate the possibility of a collection of features derived from FP-CIT SPECT brain images to be employed in computer-aided "in vivo" confirmation of dopaminergic degradation and afterward to support clinical decision-making to diagnose Parkinson's disease.
According to Saponaro, S et al. [156], the authors addressed the case-control ML separation capacity in the analysis of a multi-center MRI dataset, the authors showed how the use of a harmonization strategy on brain structural variables unlocks this capability. On the ABIDE data collection, which included people across a wide age range, this impact is proven. Following data harmonization, the overall capacity of a random forest classifier to distinguish between autism spectrum disorders (ASD) and normal development (ND) increases from very poor performance (AUC = 0.58 ± 0.04) to a still low but reassuringly significant AUC = 0.67 ± 0.03. AUC = 0.62 ± 0.02, AUC = 0.65 ± 0.03, and AUC = 0.69 ± 0.06, respectively, were obtained when the RF classifier's performances were assessed in the age-specific subgroups of children, adolescents, and adults.
According to Tufail, A. B. et al. [157], using PET and SPECT neuroimaging modalities to separate Alzheimer's disease (AD), PD, and NC classes, the authors employed a 3D CNN to extract features for multiclass classification of both AD and PD. Along with random weak Gaussian blurring, random zooming in and out, and discrete cosine transform, both frequency and spatial domain learning techniques have been used.
In the study by Antikainen, E. et al. [158], the authors investigated 23 SPECT image characteristics on 646 individuals for the early detection of PD. The authors demonstrated that matching accuracy may be reached with only eight features, including unique features, and achieve 94% balanced classification accuracy in independent test data utilizing the whole feature space. All of the qualities that are being provided can be produced by commonly accessible clinical software, making it simple to extract and use them.
In the work by Salvatore, C. et al. [159], the authors proposed Morphological T1weighted MRIs of PD patients (28), PSP patients (28), and HC subjects (28) were used by a supervised machine learning algorithm based on the combination of PCA as feature extraction technique and on SVM as classification algorithm. The algorithm was able to obtain voxel-based morphological biomarkers of PD and PSP.
The authors Martínez-Ibañez, M., et al. [160] discussed computing isosurfaces as a method of removing pertinent information from 3D brain images. These isosurfaces are then used to implement a computer-aided diagnosis (CAD) system to help with the diagnosis of PD. This system uses the most well-known CNN architecture, LeNet, to classify DaTSCAN images with an average accuracy of 95.1% and AUC = 97%, obtaining comparable (slightly better) values to those obtained for the majority of the recently proposed systems. Therefore, it may be inferred that computing isosurfaces considerably decrease the complexity of the inputs, producing good classification accuracy with little processing load.
According to the above-mentioned work, Kurmi, A. et al. [161] suggested utilizing DaTSCAN images to predict Parkinson's using a collection of DL models. The classification of PD was initially conducted using four DL models: VGG16, ResNet50, Inception-V3, and Xception. To improve the classification model's overall performance, they used a Fuzzy Fusion logic-based ensemble technique in the next step. The suggested model outperforms the individual model in terms of attained recognition accuracy, precision, sensitivity, specificity, and F1-score, which are each 98.45%, 98.84%, 98.84%, 97.67%, and 98.84%, respectively. Additionally, they have created a software application with a graphical user interface (GUI) for the general public that accurately and promptly identifies all classes in MRI.
The evaluations mentioned above demonstrate that all the research has been completed and is only limited to a few datasets. The aforementioned earlier efforts motivated us to take a different approach. In this work, we evaluated a variety of feature selection techniques and then compared the outcomes with a number of ML classifiers. Tremor, DaTSCAN, SPECT, and MRI-T are some of the key symptoms of PD that may be diagnosed with ML methods, as shown in Table 4.

Discussion: Challenges and Recommendations
Although there is no known cure for PD, with an accurate and timely diagnosis, we can reduce and control its progression. Compared with traditional PD detection methods, AI is a strong choice for detecting early-stage PD. The use of AI can help with global epidemiology initiatives and patient symptom monitoring. Despite how stimulating these applications are, it is important to consider both the value and potential limitations of these cutting-edge analytical techniques. The most promising applications of AI are yet futuristic [180][181][182].
In this section, we summarized the current limitations and challenges and made prospective suggestions (recommendations) for the future that might lead to efficient AI and ML methods to address the issues.

Current Limitations and Challenges
Currently, DL-based CAD systems are usually applied as diagnostic aids or for educational purposes [183]. With the help of useful research, the software may now be developed in the real world to diagnose Parkinson's disease utilizing MRI modalities. However, there are still a few challenges faced by researchers which are listed as follows: Issues with Multimodality datasets for PD detection Another difficulty in diagnosing PD for researchers is the unavailability of multimodality neuroimaging datasets. The diagnosis of brain disorders including PD, AD, and schizophrenia (SZ) is often greatly aided by multimodality neuroimaging data [184]. Various clinical investigations have described the reliable detection of PD using a combination of neuroimaging modalities, such as EEG-fMRI [185][186][187][188], MRI-PET [189][190][191][192], fMRI-MEG [193,194], and fMRI-sMRI [195][196][197]. Diagnosis of PD using multimodality neuroimaging data is complicated and time-consuming for doctors, despite all the advantages. The lack of multimodality neuroimaging datasets for PD detection has been a significant problem for researchers. The accessibility of multimodality neuroimaging datasets might result in significant research on PD diagnosis utilizing AI methods.

b.
Issues with Machine Learning Techniques Another challenge to diagnosing PD is related to the use of ML techniques, including that the most essential aspect of CADS is extracting the distinctive features that might Diagnostics 2022, 12, 2708 21 of 32 result in useful PD biomarkers. An extensive understanding of the AI area is needed to implement CADS based on ML. It is challenging to choose the algorithms for each component of an ML-based CADS in order to make a highly precise PD diagnosis. MLbased CADS, however, is not a function that should be used for large amounts of data input. The fact that MRI uses several imaging procedures and does not have a function that is appropriate for processing these data simultaneously presents another challenge. Developing practical software for detecting Parkinson's disease is quite challenging due to these challenges.

c.
Issues with Clinical Validation The effectiveness of deep learning models for PD detection is mostly evaluated using standard ML parameters including accuracy, sensitivity, specificity, and/or area under the curve of a receiver operating characteristic curve. These measures may not accurately represent clinical effectiveness and anticipated positive adjustments to patient treatment. Additionally, before using AI-powered diagnostic tools in the clinic, physicians must receive training on how to use them because some of these metrics are difficult to interpret [198]. To support the validity of the suggested DL framework for PD diagnosis, recent studies [199][200][201][202] addressing this problem performed a connection of model predictions with neuropathological results as well as a head-to-head evaluation of the system performance with a team of neurologists. Due to data inconsistencies, it is important to do clinical validation to make sure that the imaging-specific features are appropriately compatible with the intended clinical adoption. A CAD system must be modified for a new community through clinical adoption [160], which requires certain members of the deployment population's intended target demographic.

Recommendations and Future Perspectives
We examined both the benefits and disadvantages of the selected articles. After taking into account the recommendations for critical evaluations [203], we began searching for relevant directions for further research. We divided our findings into groups that address the same or related problems and defined them as follows:

•
Make the most of all available data sources. It might be difficult to gather all available modalities for each subject. For instance, a PET scan, a costly neuroimaging modality, is not performed on some participants, although practically all of the subjects' clinical records list an MRI. This is true for publicly accessible data like ADNI. To fill up the gaps in the data, we advise employing certain methods. For instance, utilizing MRI data to complete missing PET data [204], creating CT scans from MRIs [205][206][207], cycle-consistent generative adversarial networks (GAN) [208], and feature-consistent GAN [207]. As an alternative, deep designs that include handling procedures for missing data can be applied [209][210][211][212][213]. Additionally, data augmentation might be useful in this context to increase the dataset and address unbalanced classes. Through image modification operators including rotation [214], scaling and shifting [215], changing intensity, contrast, and saturation [216], as well as noise injection and random translation [217], data augmentation may be accomplished.

•
The widespread usage of the CNN algorithm [218,219] on MRI image data is a significant discovery. Comparing these models to other algorithms, they frequently produce favorable outcomes. Researchers might wish to conduct further research and use more CNN-based hybrid algorithms [220][221][222]. Additionally, we found that classifying images has not frequently utilized the learning algorithm. This is an opportunity for future researchers to utilize attention to raise the precision of deep learning models. • At present, wearable sensors are only useful for using gait parameters to diagnose PD. The wearable device that can detect PD must have the other modules included in it. In addition to one symptom, researchers must concentrate on creating wearable sensor systems that can diagnose additional symptoms also. For instance, a wrist-worn sensor might be created that can track data constantly over a long period of time and recognize various PD symptoms. • Currently, various ML models have been developed by researchers that can diagnose PD based on a patient's specific symptoms. The researchers should focus on establishing an ML model for diagnosing PD that takes all the symptoms as input.
A lightweight portable device can be used to diagnose the various symptoms of PD by measuring several parameters such as accuracy, precision, sensitivity, recall, etc. This device should be easily wearable and washable, and it should be able to identify the different stages of the disease, along with analyzing the changes due to medication treatment.

Conclusions
AI and ML are revolutionizing healthcare since technologies assist in the diagnosis of any disease and have made it easier in recent years. This technique has the potential to revolutionize healthcare with more accuracy in diagnosing a disease. A computerized system aids doctors in making more precise diagnoses, forecasting patients' future health, and making better treatment recommendations. In this work, we conducted a comprehensive review of 217 research papers that addressed the application of various machine learning methods and deep neural network architectures to diagnose PD. We also thoroughly examined and analyzed the researcher's architectural designs. This review is significant for the advancements in neural networks and associated learning systems, which offer insightful information and recommendations for future growth.

Conflicts of Interest:
The authors declare no conflict of interest.

Abbreviations
The following abbreviations are used in this manuscript: