4.1. Asthma/COPD
The works reported upon in this section are mainly related to chronic obstructive pulmonary disease (COPD). Counteracting the effects of having COPD was in focus in [
6,
7,
8]. Work on smart vests for lowering the number of readmissions to hospitals [
9] and for extracting RR, inspiration time and expiration time during pulmonary rehabilitation was reported in [
8]. Another work extracting RR and HR noninvasively was presented in [
10]. Despite the fact that asthma is the most common NCD in the world among children and that 235 million people live with the disease [
11], only two articles reporting on the use of wearable sensors in the assessment of asthma were retrieved [
12,
13] in this literature search.
Table 4 shows that only four studies [
6,
7,
8,
12] reported on the participants’ age. The majority of the participants were men, and three studies [
9,
10,
12] involved only men. The majority of the studies included only patients, exceptions being [
10] and a sub-study in [
8]. One article [
12] did not report this information. Four articles [
7,
8,
10,
12] reported on studies with a maximum of 10 participants. Two articles [
6,
9] reported on studies with more than 45 participants. The majority of the studies were observational, the exception being [
9], which reported on a randomized control study. As shown in
Table 5, there is no consistency in the types of sensors used, their location and the study aim in the article category Asthma/COPD.
Pulmonary rehabilitation is a means to counteract the effects of having COPD. Aiming at understanding the potential and validity of conducting pulmonary rehabilitation remotely at home, Bonnevie et al. (2019) [
6] conducted an observational study on patients with chronic respiratory disease who were referred to pulmonary rehabilitation. First, they underwent an assessment comprising pulmonary function tests, two sets of a six minute walk test (6MWT) and a cardiopulmonary exercise test. Second, they were instructed how to record and transmit data while using a telemonitoring system consisting of a pulse oximeter. Thereafter, they conducted prescribed exercises on a cycle while wearing the pulse oximeter. At a later exercise session, the participants were told to record and transmit data independently. Failing in this process resulted in having the instructions repeated and participation in a later exercise session. The majority of the participants were recording and transmitting data autonomously already after one exercise session. The validity of the data was checked by comparing it with data from a subgroup of five participants who conducted five exercise sessions during a period of ten days. The data from 98% of the exercise sessions were usable and had few artefacts (0.9%); these were mainly due to movement of the pulse oximeter’s finger probe during the exercising. The participants were satisfied with the system that they found easy to use, and ninety-eight percent of them agreed to use it throughout their pulmonary rehabilitation program [
6].
In addition, aiming at counteracting the effects of COPD, Caulfield et al. (2014) [
7] conducted a study where COPD patients were instructed to wear Fitbit One during waking hours for a period of six weeks. Reminders to wear the device were provided via phone calls and text messages. During Weeks 1–2, the wearable device’s display was occluded; hence, no feedback was provided. After Week 2, the label occluding the display was removed, and the participants could access online and on-device feedback. Exercise capability was assessed using 6MWT at baseline, after Week 2 and after Week 6. Health-related quality was simultaneously assessed using the St Georges Respiratory Disease Questionnaire. The participants were visited twice: once for installing and testing the wearable device and once for removing the occluding label and for providing information on how the device was working and what feedback it provided. The participants were encouraged to look at the feedback, but not asked to conduct physical exercises. Comparing the average number of steps taken per hour during Weeks 1–2 and Weeks 3–6, the number of steps taken increased significantly after Week 2. The results on the 6MWT also improved significantly while health-related quality remained unchanged. The participants found the wearable device easy to use, but were concerned about remembering to wear it every day [
7].
Aiming at lowering the number of readmissions after being hospitalized due to a COPD exacerbation, Katsaras et al. (2011) [
9] developed the knitted wearable clothing “Healthwear” and a portable patient unit (PPU). The authors conducted a randomized control study. One group received conventional care and was discharged using current criteria and practices, while the other group was discharged early (after 3–5 days), but monitored at home using Healthwear, the PPU and regular video checkups; the lengths of the hospital stays were 6.8 and 3.6 days, respectively. The number of readmissions and emergency room visits were also lower among the participants in the early discharge group. However, they required more visits from nurses in accordance with patient needs. The majority found Healthwear acceptable, although approximately 17% found it inconvenient.
In another work, Naranjo-Hernández et al. (2018) used a smart vest equipped with two electrodes for taking physiological measures on COPD patients while resting between respiratory rehabilitation exercises [
8]. The authors conducted studies with two healthy subjects and nine COPD patients to validate the sensors. The mobile and wireless clinical system Oxycon from CareFusion was used as a reference measurement system. Oxycon is normally used for cardiopulmonary stress testing. The healthy participants performed a sequence of actions while being seated after which they stood up and increased inspiration and expiration times in a controlled way. The experiment was conducted twice. The results were mixed; the mean error for inspiration time was positive, which indicates an overestimation, while the mean error for expiration time was negative, which indicates an underestimation. The COPD patients followed a respiratory rehabilitation program and measures were taken for 2 min between exercises. This 2 min measure has been proposed as a method for evaluating respiratory rehabilitation efficacy [
14] and improvement of breathing skills [
15]. The correlation between the smart vest and reference system measures for the COPD patients was high.
Preliminary work on noninvasive recording of the electrical respiratory muscle activity of the diaphragm (EMGdi) was presented by Estrada et al. (2016) [
10]. A bipolar Shimmer3 and lab equipment from Biopac were used to collect EMGdi data from one healthy participant. Simultaneously, inspiratory mouth pressure was measured using Pmouth. The correlation between the EMGdi signals increased with increased respiratory load. The extracted RR and HR from the EMGdi signals showed high accuracy. Estrada et al. (2016) [
10] reported that previous works have found that RR is a parameter for detecting COPD exacerbation episodes [
16] and that resting HR is used in risk prediction [
17]. Future studies are planned that include assessing the potential of using the technology with more participants.
Moving on to the use of wearable sensors to detect the asthma symptom of wheezing, Khan et al. (2020) [
12] presented work on a flexible acoustic sensor to be mounted on the chest. The 2 cm square-shaped sensor has a metal foil diaphragm. The sensor functionality was demonstrated by the ability to distinguish between sensor recordings from one young person who was asked to say hello, cough and make humming sounds to imitate wheezing.
A smartwatch-based system aimed to support the monitoring and control of paediatric asthma was reported upon by Buonocore et al. (2017) in [
13]. The system incorporates a wireless spirometer, two environmental sensors (for monitoring particulate matter and dust density), a smartwatch and a smartphone. The smartwatch includes a 6D IMU, a sensor measuring HR and GPS. The smartphone is used for collecting data and for administering ecological momentary assessment (EMA) questionnaires. The system is an extension to the BREATHEapplication, which did not feature EMA. Notifications on EMA questionnaires are both scheduled and random. Scheduled questions include asking how the wearer slept in the morning and how school was in the afternoon. EMA questionnaires can also be triggered after the spirometer is used, if the system detects air quality changes or when the energy expenditure is high and after physical activity. No study has been conducted yet, but a clinical trial is planned.
4.2. Cardiovascular Diseases
The works reported upon in this section include 24/7 systems for ECG monitoring [
18,
19,
20], a preliminary work on person identification aiming toward zero-effort monitoring of cardiovascular diseases [
21], three works aiming toward estimating BP continuously without wearing a BP cuff [
22,
23,
24] and work on a bipolar wireless ECG monitor [
25,
26]. The device was later certified as a medical device, and Reference [
27] reported on a number of studies conducted with the device. Furthermore, this section reports on two works focusing on extracting HRV data [
28,
29]: one work assessing the accuracy of the Apple watch for monitoring atrial fibrillation (AF) [
30] and two works focusing on the triage of patients with cardiovascular disease [
31,
32].
Table 6 shows that there are significant shortcomings in the retrieved articles reporting on participant demographics. Five studies [
18,
19,
22,
31,
32] did not provide any information on the participants’ age. Only five of the reported studies [
21,
25,
28,
29,
30] and one of the sub-studies in [
27] reported on the participants’ gender. Two articles [
21,
25] reported on studies conducted solely with healthy participants. The number of participants varied between studies, and few studies or sub-studies involved more than 60 participants. Shifting focus to research design, five works [
18,
20,
25,
31,
32] lacked this information. As shown in
Table 7, mainly ECG sensors with a varying number of leads were used in the studies. However, a few works reported on the use of other sensors.
An overview of the WE-CARE system that provides 24/7 health monitoring using a wearable mobile seven lead ECG device was presented by Huang et al. (2014) in [
19]. The system’s performance and validity were evaluated in a case-control study including both healthy users and patients diagnosed with a cardiovascular disease. Using several databases for validating the performance of the system’s R wave and T wave detection algorithms, the authors observed detection rates of 99.4% and 97.7%, respectively. The system also yielded a high anomaly detection rate. In another work [
18], Huang et al. (2014) evaluated WE-CARE collection of data using a five lead ECG. The study included 225 participants, but the information provided was insufficient for interpreting how the study was conducted.
Another 24/7 prototype of a telehealth system integrating a portable ECG sensor via the interface Alive ECG was presented by Raad et al. (2015) [
20]. Preliminary work collecting data from 30 students and two elderly patients with arrhythmia showed that data can be transmitted to a PC where HR and the QRScomplex of the ECG analysis could be retrieved from the filtered signal.
Starting from the standpoint that BCG, a measure of body vibrations caused by ejection of blood into aorta, can potentially be used for zero-effort monitoring of cardiovascular diseases, Javaid et al. (2018) [
21] conducted work aiming at identifying people using BCG in a home. They built a BCG sensor with four load cells into a glass tile, i.e., a floor tile. Sixty healthy young participants, who also wore a Shimmer 2r ECG sensor, were asked to stand still on the tile in an upright position for 60 s. They were then asked to perform a 15 s long Valsalva manoeuvre, i.e., close the mouth, pinch the nose and blow up a balloon. Thereafter, they were asked to rest in an upright position on the tile for 5 min. The task was repeated three times. Data from eight participants had to be excluded in the analysis: four participant’s data for physiological reasons, i.e., abnormal BP drop, preventricular contractions in the ECG or involuntary movement, and four for technical problems with the tile. The authors found that the BCG signal contains information that can be used for identifying a person.
Another work by Susič and Stanič (2015) focused on excluding or confirming arrhythmia [
25]. Data from 13 healthy volunteers were collected for a period of a few hours to a day using a prototype sensor with bipolar leads (PCARD, a system that can collect data for up to three days without recharging the battery). Participants found that the sensor was easy to wear and did not disturb them or their family. Data were also collected from five patients, but the information was insufficient for reporting on the data in this article. Information on four upcoming pilot studies for evaluating PCARD’s validity and impact at various stages of care in Slovenia was presented by Depolli et al. (2016) [
26].
Since then, PCARD has been certified as a class IIa device according to the Medical Devices Directive MDD 93/42/EEC. The device is available on the market under the name SavvyECG. Detailed information on the hardware and firmware design of Savvy ECG was provided by Rashkovska et al. (2020) in [
27], which also compared the device with three other CE-marked and/or FDA-approved devices for measuring ECGs (SEEQ™, ZIO
®XT and Philips’ wearable biosensor) with respect to design concepts. A main difference is the fact that Savvy ECG is rechargeable, while the others are for single-use scenarios. The battery time has been extended, and the Savvy ECG sensor can now collect data for up to 10 days between recharges. Hence, Savvy ECG allows for data collection over a longer period of time. Compared to a 12 lead ECG, the Savvy ECG can be positioned at different locations and in different orientations. Furthermore, Reference [
27] reported brief information from two pilot studies [
36,
37] that we presume correspond to the first two pilots outlined in [
26]. The aim of [
36] by Kocjančič and Avbelj (2018) was to obtain insights on the practical use of PCARD on patients in whom there is a suspected heart rhythm disorder. Rashkovska et al. (2020) [
27] reported that the physicians involved in [
36] decided to follow up on 63.5% of 100 patients at the Health Centre Ljubljana using the sensor. Out of these, eighteen-point-three percent were referred to a cardiologist. In addition, further testing and prescription of new medicine were the actions taken for 6.7% and 6.7%, respectively. It was concluded that the use of a personal ECG sensor such as PCARD can lead to new pathways for patients with cardiovascular diseases. The aim of [
37] by Ĉarman et al. (2018) was to assess PCARD’s ability to detect post-surgery AF in comparison to the currently established clinical protocols. In the study, forty-seven patients wore the ECG sensor continuously from Day 1 to 5 post-surgery. Out of these, thirteen patients developed a paroxysmal AF. All of these cases were detected by the sensor, but only nine of them were detected through the clinically established protocols.
Furthermore, Rashkovska et al. (2020) [
27] reported on two additional studies. It was found in Rashkovska and Avbelj (2017) [
38] that foetal ECG measured using Savvy was only sufficient for detecting heart rates. However, the study included only two pregnant women. Širaiy et al. (2019) [
39] studied the applicability of ECG measurements for monitoring heart rhythm during the conduction of intensive activity. Aiming at evaluating ECG distortion levels in relation to the sensor position and fixation method, twenty-three participants conducted a total of four exercise stress tests (EST) during four days, i.e., two while cycling on an ergometer and two while running on a treadmill. Resting of 5 min was allowed after four electrodes (of two types, namely, PREMIER T-60 and ELITE FS-VB01) were mounted and after the end of the test. The participants’ ECGs were also recorded during the resting period. During the first two days, the participants conducted the EST using the same device, i.e., the cycle ergometer or the treadmill, with sensors mounted on the LIand LSpositions. On the second day, the sensors were secured using self-adhesive tape. The process was repeated for the last two days on the other device. It was reported in [
39] that the signal quality obtained using both fixed and non-fixed sensors was good at the LI position. The signal quality of the data from the LS position was less acceptable than the signal quality of the data from the LI position due to the influence of shoulder movement, particularly during the treadmill EST.
Aiming toward a noninvasive and continuous method for estimating BP, Li et al. (2019) [
22] studied the phase difference between two pulse waves collected at the wrist. Data were collected from 16 participants wearing an electronic BP monitor on the upper arm and a wristband measuring PPG and DPV above the radial artery. While insufficient details were provided for interpreting how the experiments were conducted, the authors found that the phase difference is highly correlated with BP and blood flow fluctuation. However, contraction and relaxation of muscles can influence estimation accuracy.
In an additional attempt to estimate BP without wearing a BP cuff, Simjanoska et al. (2018) [
23] developed a method for estimating BP from ECG signals. Data were collected from 51 people with ages 16–83 using three different commercial ECG biosensors (Cooking hacks 3 lead, 180 eMotion FAROS 3 lead and Zephyr Bioharness 1 lead) and a dataset containing ECG recordings using medical equipment from seven people. Reference BP values were taken using a BP monitoring device. After filtering and segmenting the data, the authors applied a machine learning algorithm combining stacking-based classification and regression for predicting systolic BP (SBP), diastolic BP (DBP) and mean arterial pressure (MAP). The predictions were presented in three BP classes (normal, prehypertension and hypertension) or in numeric values. Using unobtrusive sensors, the authors achieved results that were comparable to prior work. An updated report on the work was provided in Simjanoska et al. (2020) [
24]. In this work, additional data were added from participants wearing the Savvy ECG. In addition, eleven participants from the third experiment reported upon in [
23] were excluded. We assume that the excluded data were collected from unhealthy participants and that the data from healthy participants were kept. In addition, Reference [
24] presented 17 performance metrics for evaluating the seven classification models used. Then, PROMETHEE2 was used for a pairwise comparison between all classifiers for each performance metric. PROMETHEE methods are used for decision making, and the classifiers were ranked by calculating positive and negative performance preferences in relation to the other classifiers. Additional information on the methodology was provided in [
24]. A similar approach was used to rank the regression models. Then, information of all regression models was fused. Predicted values of SDP, DBP and MAP were obtained. These were better than the values observed in [
23] when there were data available for training. However, Simjanoska et al. (2020) [
24] stated that further effort is needed before the approach is acceptable for clinical purposes.
Gonzalez et al. (2019) [
28] compared the impact of the inclusion of signal quality on different predictive models of RR intervals (i.e., HRV) in real-world settings. The analysed data were collected from one young male participant performing daily activities. The data represent a subset of a data collection. Data from IMUs (it is unclear whether these were accelerometers, 6D IMUs or 9D IMUs) were collected using a smartwatch and the wearable Bioharness from which ECG data were also retrieved. The data were collected during 21 sessions lasting for approximately 2 h each. All sessions took place during a period of approximately seven weeks. Information on the participant’s specific activity during a session was not provided. Starting from the fact that behaviours and environments impact noise, Reference [
28] used the MATLAB PhysioNet Cardiovascular Signal Toolbox [
40] to compute RR intervals and normalized signal quality at a sampling rate of 1 Hz to match the sampling rate of extracted and normalized time- and frequency-domain data from the IMUs. Three different prediction models were trained using the IMU features: linear regression (LR), a random forest (RF) and a long-short term memory network (LSTM). A four-fold cross-validation that retained intact individual sessions was performed on each model. This ensured that data from one session were not included in both training and validation datasets. The models were evaluated by calculating the root mean squared error (RMSE). Thereafter, all models were modified by also taking a signal quality index (SQI) as the input. The addition of the new feature had a marginal effect on all prediction models’ RMSE. Aiming at predicting the RR interval using the trends of the physiological responses rather than learning to predict incorrect RR intervals due to noise, Reference [
28] also removed data based on the SQI using two different approaches: first, by training the models using all of the data and then removing any noise data from the validation set based on an SQI threshold. The authors calculated the RMSE for each of the three models when varying the SQI threshold. The RF was found to outperform the two other models. The optimal SQI threshold resulting in the lowest RMSE for both the LR and RF was 0.95. This resulted in a removal of 30% of the data. In the second approach, data segments were removed from both training and validation data based on a varied SQI threshold, i.e., the model learned by using only data with a high signal quality. In this attempt, the LSTM’s performance was improved with an optimal SQI threshold value of 0.95. In addition, the LR and RF were positively impacted by the second approach. Notably, the RF was found to outperform the LR at lower SQI values while the LR outperformed the RF at higher SQI. Gonzalez et al. (2019) [
28] reasoned that simpler models may predict the RR intervals when the SQI is high, i.e., light activities and sedentary activities where the noise is lower. Further comparison against additional measurements of signal quality and inclusion of data from more participants are planned.
Aiming at predicting AF, Pérez-Valero et al. (2020) [
29] also used HRV as a basis in their algorithms. First, they provided information on their already published Recurrence Analysis to Detect Atrial Fibrillation (ReAD-AF) algorithm [
41], which can be employed on the time series of HRV data to distinguish between normal sinus (NS) and AF patients. The HRV extracted from an ECG signal was taken as the input. In ReAD-AF [
41], the signal is converted into symbols. Thereafter, a matrix containing the recurrences of all symbols observed is defined. The plots of these matrices from NS and AF patients are different. The distribution appears structured for an NS patient and unstructured for an AF patient. Thereafter, a logistic model estimating the probability of patient category is applied. The ReAD-AF algorithm was validated through a k-fold cross-validation procedure using a publicly available dataset provided by Physiobank [
42]. The calculated sensitivity and specificity were approximately 95% and found to increase with the selected window size. Further information on the algorithm can be found in [
41], and the pseudocode for the algorithm is presented in [
29]. In [
29], Pérez-Valero et al. (2020) presented information on and an initial validation of a fully functional prototype capable of extracting HRV data from PDF and JPEG files that can be of low resolution. The prototype consists of four modules: digitalization, signal processing, calibration and the application of the ReAD-AF algorithm. The ECG signal is digitalized using the MATLAB function
imread at a 600 dpi resolution and by selecting the ECG part of the file. Since ECG charts typically include a grid, an RGB threshold is applied to remove the background. Thereafter, the RGB colour images are converted into binary images. In the next module (signal processing), the signal is preprocessed by removing noise, smoothing and amplifying the QRS slope. Thereafter, the MATLAB function
findpeaks is used to obtain signal peaks. In the third module, data are calibrated against a 12 lead gold standard ECG (MAC800) using an LR model. Finally, the ReAD-AF algorithm presented in [
41] is applied for distinguishing between NS and AF patients. For validating the calibration procedure, HRV was obtained from 20 patients using a one lead ECG (KardiaMobile) and MAC800. Twenty samples from each patients were collected; however, it is unclear what the patients were doing. All patients had a normal NS, which leads us to believe that patients should be regarded as participants and not patients. To estimate the error after calibration, the mean squared error (MSE) was used. Two MSEs were calculated. First, the MSE between the uncalibrated signals using the one lead ECG and the MAC800 was obtained, yielding an MSE of 0.001971. Second, the MSE between the calibrated one lead ECG signal and the MAC800 was obtained, yielding an even lower MSE of 0.000355. For validating the classification model, a dataset containing 3658 ECG recordings using KardiaMobile was used [
34]. The dataset contained four different classes of ECGs (NS, AF, other rhythms and noisy recordings), and only those being classified as NS and AF were used for validation. The classification performance was analysed using both raw and calibrated data. It was found that the sensitivity, specificity and accuracy were rather similar when running ReAD-AF on raw and calibrated data. Regarding the raw data, the sensitivity, specificity and accuracy performance achieved was 0.8765, 0.9186 and 0.9133, respectively, i.e., the algorithm correctly detected AF in 87.65% of the cases and NS in 91.86% of the cases. Overall, the data were correctly classified in 91.33% of the cases. In addition, the results in [
29] were compared with results achieved with other methods in the literature. Using only nine features instead of 30 or more features in the comparative algorithms, the ReAD-AD algorithm performed good and sometimes better. It was reported in [
29] that future work includes extending the logistic model such that it can detect several types of arrhythmia and not only AF. Data from medical records should also be used in the logistic model.
Seshadri et al. (2020) [
30] reported on the results of a pilot study aiming at assessing the accuracy of the Apple Watch for monitoring patients with AF in comparison to telemetry. A total of 50 postoperative cardiac surgery patients wore a watch for a maximum of 5 min on a randomly decided wrist. An equal portion of them had AF and NS. A minimum of three assessments per day for at least two days were made. A total of five watches and five iPhone 8s were used. According to Lin’s concordance correlation coefficients (r
c), there was an overall agreement of 0.7 between the AW and telemetry assessments. It was also found that the HR was measured more accurately among the patients with AF (r
c = 0.86) than among patients with an NS (r
c = 0.64). The overall r
c value of 0.7 was found to be lower than prior results obtained by the authors’ team, and further studies are needed.
The two last works in this article category are notably different from the others reported. A.S. Albahri et al. (2019) [
31] presented a smart real-time health recommender framework for remote chronic heart services provision. The input was taken from wearable sensors (ECG, SpO
2 and BP) and text. Data from 500 patients with different symptoms were used to triage patients into the groups: normal, sick, urgent or at risk. Different healthcare service packages need to be provided to these patients, particularly during challenging scalability situations. Therefore, Reference [
31] ranked 12 hospitals in Baghdad based on what they offer to patients with a chronic disease. Three packages were provided in which the hospitals offering the first package had the most services to offer. Experts were used for ranking the hospitals and the importance of the different services. The triage indicated that 66 patients required the first healthcare service package, 151 required the second one and 260 required the third one. Further information on decision-making techniques and validation thereof was provided in [
31]. In addition, O.S. Albahri et al. (2019) [
32], i.e., the same research group, presented work in this area. Aiming for generalizability, they presented a fault-tolerant framework on mHealth in an IoT context. Triage is done locally, and the medical centre receives a warning if there is a failure related to a sensor. Then, hospital selection taking into account both available healthcare service packages and the factor of time of arrival at the hospital (TaH) is proposed. TaH is particularly important for urgent patients. Two datasets were used in [
32], one with 572 patients with chronic heart disease and one with the 12 Baghdad hospitals. The latter includes information on the maximum capacity per hospital. In addition, an assumed dataset with 12 hospitals in Kuala Lumpur was used. The results showed that the hospital ranking depends both on TaH and the availability of services. For further information, please refer to [
32].