sensors

Review
The Application of Wearable Sensors and Machine Learning
Algorithms in Rehabilitation Training: A Systematic Review
Suyao Wei 1 and Zhihui Wu 1,2, *

1 College of Furnishings and Industrial Design, Nanjing Forestry University, Nanjing 210037, China;
13675164200@163.com
2 Co-Innovation Center of Efficient Processing and Utilization of Forest Resources, Nanjing Forestry University,
Nanjing 210037, China
* Correspondence: wzh550@sina.com

Abstract: The integration of wearable sensor technology and machine learning algorithms has
significantly transformed the field of intelligent medical rehabilitation. These innovative technologies
enable the collection of valuable movement, muscle, or nerve data during the rehabilitation process,
empowering medical professionals to evaluate patient recovery and predict disease development
more efficiently. This systematic review aims to study the application of wearable sensor technology
and machine learning algorithms in different disease rehabilitation training programs, obtain the best
sensors and algorithms that meet different disease rehabilitation conditions, and provide ideas for
future research and development. A total of 1490 studies were retrieved from two databases, the
Web of Science and IEEE Xplore, and finally 32 articles were selected. In this review, the selected
papers employ different wearable sensors and machine learning algorithms to address different
disease rehabilitation problems. Our analysis focuses on the types of wearable sensors employed, the
application of machine learning algorithms, and the approach to rehabilitation training for different
medical conditions. It summarizes the usage of different sensors and compares different machine
learning algorithms. It can be observed that the combination of these two technologies can optimize
the disease rehabilitation process and provide more possibilities for future home rehabilitation
scenarios. Finally, the present limitations and suggestions for future developments are presented in
the study.

Citation: Wei, S.; Wu, Z. The Keywords: wearable sensor; machine learning; disease rehabilitation; rehabilitation training
Application of Wearable Sensors and
Machine Learning Algorithms in
Rehabilitation Training: A Systematic
Review. Sensors 2023, 23, 7667.
1. Introduction
https://doi.org/10.3390/s23187667
With the rapid development of information technology, traditional medical rehabil-
Academic Editor: Susanna Spinsante itation methods combined with various disciplines and technologies, such as wearable
Received: 5 August 2023
sensors and machine learning algorithms, are widely used in clinical diagnosis, rehabili-
Revised: 24 August 2023 tation medicine, and other fields [1,2]. Cervical spine diseases, musculoskeletal diseases,
Accepted: 4 September 2023 stroke, cerebral palsy, hand paralysis, lower-limb paralysis, Parkinson’s, and other diseases
Published: 5 September 2023 require long rehabilitation periods. Wearable sensors and machine learning technology can
assist clinicians in monitoring and predicting the prognosis and rehabilitation of patients.
For example, Vijay placed the IMU (inertial measurement unit) on the chest and thighs of a
patient to collect data on walking activities, such as standing, climbing stairs, cycling, etc.,
Copyright: © 2023 by the authors. to complete the monitoring of the patient’s rehabilitation process [3]. Wearable sensors are
Licensee MDPI, Basel, Switzerland. an important technology for gait analysis, diagnosing walking disorders in patients with
This article is an open access article gait disorders, and gait analysis is very important for the clinical assessment of patient
distributed under the terms and rehabilitation [4]. Patients with hemiparesis, such as apoplexy, usually must observe and
conditions of the Creative Commons
evaluate hand-movement performance during the rehabilitation training period. Therefore,
Attribution (CC BY) license (https://
wearable sensors that do not affect limb movement can be worn for tracking and monitor-
creativecommons.org/licenses/by/
ing purposes. The feedback on joint movement information is crucial for the adjustment
4.0/).

Sensors 2023, 23, 7667. https://doi.org/10.3390/s23187667 https://www.mdpi.com/journal/sensors

Sensors 2023, 23, 7667 2 of 30

of and change in the rehabilitation treatment process [5]. Machine learning technology
can integrate and predict the data obtained by sensors used for disease rehabilitation,
thereby improving the accuracy of diagnoses of stroke and other diseases and assisting
rehabilitation personnel in predicting the patient’s disease recovery trajectory [6–8].
Wearable sensors first appeared in the mid-20th century. As a hardware device,
they can perform data interactions. According to different needs, users wear devices
with specific functions to collect behavior or health records [9]. Wearable devices include
a device body and sensor components, which are mechanically connected. They have
different functions, principles, and forms, and are widely used in the fields of medicine
and health [10]. Wearable sensors have the characteristics of convenience and a low
price, providing researchers with a variety of possibilities and solutions [11]. Wearable
sensors help rehabilitation patients to exercise at home, relieve travel pressure, and reduce
psychological burden [12,13]. A variety of sensing devices are used to monitor patients’ vital
signs and physiological responses, such as electromyography (EMG), electrocardiogram
(ECG), and electroencephalogram (EEG), which can monitor the patient’s physical condition
in real time. Electromyography (EMG) can determine the functional status of peripheral
nerves, neurons, and muscles by receiving electrical activity signals when the muscles are
at rest or contracting [14]. Electrocardiography (ECG) records the electrical activity of the
heart by detecting the potential activity between cardiomyocytes and is commonly used
to rapidly check for signs of arrhythmia [15]. An electroencephalogram (EEG) typically
involves placing electrodes on a person’s scalp to detect changes in biological potential
caused by brain activity. Brain waves contain a large amount of physiological and disease
information. Through the processing of brain waves, doctors can perform the rehabilitation
identification of patients’ brain diseases [16]. Gait analysis using wearable sensors, such as
inertial sensors, gyroscopes, accelerometers, pressure sensors, etc., is widely used in many
fields, such as neurorehabilitation and sports medicine. An inertial sensor is a sensor that
detects and measures acceleration, tilt, vibration frequency, rotation angle, and multiple
degrees of freedom (DOF) motion. They can convert motion signals into electrical signals,
which are amplified and processed by electronic circuits [17]. A gyroscope is an angular
motion-detection device that measures the angular velocity around multiple axes [18].
Accelerometers are sensors that measure changes in velocity in a single direction. Due
to their low cost and strong reliability, they are often used in combination with various
sensors [19]. A pressure sensor is generally composed of a pressure-sensitive element and
a signal processing unit. It is a device that can sense the pressure on an object and convert
the pressure signal into an electrical signal according to a certain rule. It is usually placed
on the sole of the foot in gait recognition systems to obtain pressure information during
movement [20].
Machine learning is a mechanism that uses computers to simulate human learning
activities, enabling machines to learn autonomously without explicit programming, or
researching how to effectively use information to obtain hidden and effective knowledge
from big data [21]. Machine learning algorithms have been applied in different fields,
such as finance, environmental protection, social media, and healthcare industries. In
the medical field, machine learning is continuously upgraded and optimized in terms
of disease analysis and data prediction [22–24]. With the advent of the era of big data,
machine learning technology can efficiently acquire knowledge, conduct an in-depth
analysis of complex and diverse data, and improve the accuracy of prediction results [25].
The commonly used algorithms of traditional machine learning mainly include the support
vector machine algorithm (SVM), decision tree algorithm (DT), random forest algorithm
(RF), artificial neural network algorithm (ANN), and so on. The support vector machine
(SVM) algorithm is a supervised learning method that can be widely used in statistical
classification and regression analysis [26]. Support vector machines are mainly used for face
detection, image classification, and biological data mining. It is unlike the traditional way
of thinking. It simplifies a problem by inputting the space and increasing the dimension, so
that the problem can be reduced to a linearly separable classic problem [27]. The decision
Sensors 2023, 23, 7667 3 of 30

tree (DT) algorithm is an important classification and regression method in data mining
technology, and its predictive analysis model is generally expressed in a tree structure [28].
The understandability of a decision tree model is affected by the size of the tree, its depth,
and the number of nodes in the leaves. Decision tree has the characteristics of small
levels of calculation and high accuracy [29]. The random forest algorithm (RF) integrates
multiple trees through the idea of ensemble learning. The output category is determined
by the mode of the output category of each tree and is mainly used for classification
predictions [30]. This algorithm has the advantages of high precision, wide applicability,
strong nonlinear data analysis ability, and overfitting difficulty [31]. The artificial neural
network algorithm (ANN) is an algorithmic model that imitates the structure and function
of biological neural networks [32]. Inspired by the neural organization of the human brain,
the algorithm designs computing nodes similar to neurons and connects them to form a
network. It transmits information rapidly and has strong generalization and nonlinear
mapping abilities [33].
The review of wearable sensors and the machine learning algorithms in the literature
mainly focuses on stroke rehabilitation [34], gait monitoring [35], fall prevention [36], and
lower-limb movement [37,38]. For example, Jourdan et al. [39] focused on researching the
application of commercial sensors, aiming at data collection of how sensors are applied,
and seldom elaborated on the data processing that requires the application of machine
learning technology. Usmani et al. [40] analyzed and compared the basic information of
the participants, data sets, machine learning algorithms, sensor types, and where on the
body they are worn and other parameters, and described the latest application of machine
learning in fall monitoring and prevention systems. Boukhennoufa et al. [41] summarized
the latest research progress in the field of stroke rehabilitation and compared the data
processing of wearable sensors and machine learning algorithms.
At present, some reviews have summarized the latest research progress of wearable
sensors and machine learning technology; however, a summary of disease rehabilitation
training is lacking in the research. Many studies in the literature discuss the application
of various sensors and machine learning techniques in the treatment and rehabilitation
of certain diseases. For example, force sensors and bending sensors are added to stroke
rehabilitation gloves to measure the grip strength and bending degree of the hand, and
use machine learning technology to recognize gestures to promote the completion of the
rehabilitation training process for patients [42]. Facciorusso et al. [43] used CiteSpace
6.1.R6 software to review the research status of sensor-based rehabilitation in neurological
diseases, and to conduct a visual analysis of the research hotspots, authors, and journals.
Yen et al. [34] reviewed the application trends of sensors in the remote monitoring and
rehabilitation of neurological diseases, and discussed the functional evaluation elements
that sensors should simulate. The abovementioned reviews are based on different perspec-
tives of neurological diseases. According to the survey, there is no review summarizing the
application and trend of the use of wearable sensors and ML technology in rehabilitation
training for different diseases, which prevents researchers from making horizontal and
vertical comparisons in this regard. Therefore, it is necessary to summarize the status
of the use of wearable sensors and machine learning technology at present in different
rehabilitation training scenarios for different kinds of diseases. The focus of the research
should be on sensor location, sensor type, etc., as well as comparing the types and accuracy
of machine learning algorithms to obtain the optimal algorithm. Sensors and machine
learning-related information should be visualized to provide references for scholars to
facilitate additional research. The research objectives of this review are as follows:
• It outlines the application of wearable sensors and machine learning technology in
rehabilitation training;
• It specifically analyzes the sensor type, sensor location, and feature extraction applied
in the recovery process of different diseases;
• It evaluates the type and accuracy of machine learning algorithms applied in different
rehabilitation exercises;
Sensors 2023, 23, 7667 4 of 30

• It discusses the limitations, trends, and directions of sensors and machine learning
algorithms in rehabilitation applications.
The purpose of this study is to review the application of wearable sensor technology
and machine learning algorithms in rehabilitation training for different diseases. The
research results include the best sensors and ML algorithms that meet the rehabilitation
conditions of different diseases, providing researchers with a choice of research directions
and ideas for future research and development purposes.

2. Methods
This review used the Preferred Reporting Items for Systematic Reviews and Meta-
Analyses (PRISMA) for the paper selection [44].

2.1. Search Method

The literature search used the Web of Science database and IEEE Xplore to retrieve all
the literature published during the ten-year period from 1 January 2013 to 4 July 2023.

2.2. Document Retrieval

Firstly, the basic keywords used for the literature search were “wearable sensor”,
“machine learning”, and “rehabilitation training”. Then, more relevant keywords were
selected. The search formats of the two databases are shown in Table 1.

Table 1. Keyword strings used in database searches.

Academic Library Search String

((TS = (wearable OR wearable sensor OR wearable device OR
wearable sensing device OR accelerometer)) AND TS = (machine
learning OR intelligent system OR deep learning OR SVM OR
Web of Science support vector machines OR random forest algorithms OR neural
network algorithms OR multilayer perceptron OR artificial neural
networks OR ANN)) AND TS = (rehabilitation OR recovery OR
rehabilitation training)
(“All Metadata”: wearable OR “All Metadata”: wearable sensor
OR “All Metadata”: wearable device OR “All Metadata”:
wearable sensing device OR “All Metadata”: accelerometer) AND
(“All Metadata”: machine learning OR “All Metadata”: intelligent
system OR “All Metadata”: deep learning OR “All Metadata”:
IEEE Xplore SVM OR “All Metadata”: support vector machines OR “All
Metadata”: random forest algorithms OR “All Metadata”: neural
network algorithms OR “All Metadata”: multi-layer perceptron
OR “All Metadata”: artificial neural networks OR “All Metadata”:
ANN) AND (“All Metadata”: rehabilitation OR “All Metadata”:
recovery OR “All Metadata”: rehabilitation training)

Through the search, potentially relevant articles published between 1 January 2013
and 4 July 2023 were identified. Figure 1 presents the number of potentially relevant
articles published per year between 1 January 2013 and 31 December 2022, after excluding
duplicates. It can clearly be observed in Figure 1 that the number of published papers is
clearly on the rise.
Sensors 2023, 23, x FOR PEER REVIEW 5 of 33
Sensors 2023, 23, 7667 5 of 30

Numberofofrelevant
Figure1.1.Number
Figure relevantarticles
articlesretrieved
retrievedbetween
betweenthe
theyears
years2013
2013and
and2022.
2022.

2.3. Screening Criteria

2.3. Screening Criteria
This review only included peer-reviewed journals or conference papers written in
Thisbetween
English review only included
1 January 2013 peer-reviewed
and 4 July 2023.journals or conference
The article papers
research content written
was in
required
English
to meetbetween
all of the1following
January 2013 and 4 July 2023. The article research content was required
criteria:
to meet all of the following criteria:
1. The paper should include research conducted on wearable sensors, machine learning,
1. The andpaper
diseaseshould include research conducted on wearable sensors, machine learn-
rehabilitation;
2. ing, and disease rehabilitation;
The paper should provide a detailed analysis of the performance characteristics, where
2. The the paper
sensor should
is worn,provide
and the aaccuracy
detailedofanalysis of thesensor;
the wearable performance characteristics,
3. where the sensor is worn, and the accuracy of the wearable
The paper should elaborate on the application of the machine sensor;learning algorithm
3. The paper in
involved should elaborate on the application of the machine learning algorithm in-
data processing;
4. volved in data
The paper processing;
should include research conducted on the treatment and rehabilitation of
4. The onepaper
or moreshould include research conducted on the treatment and rehabilitation of
diseases.
one or more diseases.
In addition, the abovementioned criteria should be followed for paper selection, the
In addition,
exclusion criteriathe
wereabovementioned
as follows: criteria should be followed for paper selection, the
exclusion
1. criteria were as follows:
Exclude all review papers, review articles, and papers that lack specific research results;
1.2. Exclude all review
If the research papers,
exists in bothreview
academic articles, andand
journals papers that lack
conference specific
papers, research
select re-
the former;
3. sults;
Exclude papers that briefly mention wearable sensors or machine learning or dis-
2. Ifease recovery.exists in both academic journals and conference papers, select the for-
the research
mer;
3.2.4. Exclude
Article Screening
papers thatProcess
briefly mention wearable sensors or machine learning or disease
The investigator (WSY) entered the data exported from the two databases into a table,
recovery.
which included information on the author, title, keywords, abstract, DOI number, etc.
After
2.4. excluding
Article duplicate
Screening Processrecords according to the title and DOI number of the article, the
investigator (WSY) then
The investigator excluded
(WSY) enteredthe
thepapers that did not
data exported frommeet
the the
tworequirements
databases into according
a table,
to the screening criteria based on the title, keywords, and abstract. Finally,
which included information on the author, title, keywords, abstract, DOI number, etc. Af- the investigators
(WSY
ter and WZH)
excluding checked
duplicate whether
records the specific
according to thecontent
title andof DOI
the paper
numbermetofall the
the screening
article, the
requirements. The final screening was conducted and the selected papers
investigator (WSY) then excluded the papers that did not meet the requirements accord- were summarized.
ing to the screening criteria based on the title, keywords, and abstract. Finally, the inves-
3. Results
tigators (WSY and WZH) checked whether the specific content of the paper met all the
A total
screening of 1490 documents
requirements. were
The final retrieved
screening wasfrom the database,
conducted and theincluding
selected 527 fromwere
papers Web
of Science and 963 from IEEE Xplore. First, after removing 111 duplicates, 1379 papers
summarized.
were retained. Then, 1064 articles were excluded according to the title, keywords, and
abstract. Then, 57 reviews were excluded. Subsequently, the full-text content was reviewed
 3. Results
A total of 1490 documents were retrieved from the database, including 527 from Web
Sensors 2023, 23, 7667 of Science and 963 from IEEE Xplore. First, after removing 111 duplicates, 1379 papers 6 of 30
were retained. Then, 1064 articles were excluded according to the title, keywords, and
abstract. Then, 57 reviews were excluded. Subsequently, the full-text content was re-
viewed according to the screening criteria, 226 articles were excluded, and finally 32 arti-
according to the screening criteria, 226 articles were excluded, and finally 32 articles were
cles were obtained. The whole process of document retrieval shown in Figure 2 was based
obtained. The whole process of document retrieval shown in Figure 2 was based on the
on the screening results at each stage obtained by the method steps of PRISMA.
screening results at each stage obtained by the method steps of PRISMA.

Figure 2. Literature screening process.

Figure 2. Literature screening process.

Sensors 2023, 23, x FOR PEER REVIEW We sorted and summarized the 32 selected papers by year, and the results are 7shown of 33
We sorted
in Figure and were
3. There summarized the 32inselected
fewer papers 2023 duepapers
to theby year, and
deadline forthe
theresults arethe
scope of shown
article
insearch
Figurebeing
3. There were
4 July fewer papers in 2023 due to the deadline for the scope of the article
2023.
search being 4 July 2023.

Yeardistribution
Figure3.3.Year
Figure distributionofofselected
selectedpapers.
papers.

3.1. Wearable Sensors

Wearable sensors have the characteristics of light weight, flexibility, stability, and
comfort. They are widely used for pulse and heartbeat monitoring purposes, gait analysis,
and other health monitoring systems, disease diagnosis, and rehabilitation fields [45]. In
Sensors 2023, 23, 7667 7 of 30

3.1. Wearable Sensors

Wearable sensors have the characteristics of light weight, flexibility, stability, and
comfort. They are widely used for pulse and heartbeat monitoring purposes, gait analysis,
and other health monitoring systems, disease diagnosis, and rehabilitation fields [45]. In
the process of the diagnosis and rehabilitation of different diseases, different sensors are
required to detect the physiological information required. In order to comprehensively
evaluate the recovery of human health, it is sometimes necessary to work with multiple
sensors. For example, a gait recognition system based on pressure and inertial sensors
can obtain pressure information from the soles of the feet during exercise; inertial sensors
can obtain dynamic information, such as acceleration and angular velocity, from different
positions, such as on the thighs and ankles [20]. Based on the detected information, the
recovery status of stroke patients can be evaluated and subsequent interventions can be
performed.
Table 2 summarizes the specific results of the 32 screened documents on sensor type,
wearable sensor location, sampling frequency, exercise, disease types, and other information.

Table 2. Wearable sensors used for rehabilitation training in selected papers.

Wearable Sensor
References Participants Feature Sampling Rate Exercise Disease Type Methods
Sensors Type Location
10 post-stroke Mean
Hand to lumbar Upper-limb
hemiplegic- value/standard
deviation/root spine/shoulder evaluation
Nine-axis sensor simulated Post-stroke
[46] Wrist square mean 100 Hz flexion 90 de- method in the
(non-invasive) subjects hemiplegia
grees/forearm Fugl–Meyer
Male: 7 value of motion
tasks pronation scale
Female: 3
Framework that
Time domain uses a
combination of
feature:
machine
mean/mean
learning
absolute devia-
tion/peaks; Wheelchair models and
Two propul- wearable
Accelerometer individuals frequency Spinal cord
[47] Wrist/ankle 32 Hz sion/walking/ sensors to
(non-invasive) without spinal domain injury capture and
cord injuries features: total walking using
crutches track assistive
power between
technology-
a band of
based mobility
frequen-
cies/energy/entropy and function in
individuals
with SCI
Rigidity
features: mean
and standard
deviation of the
calculated
torque/standard
deviation of the
joint angle and
23 Parkinson’s angular
disease patients velocities, etc. Establish a new
Bradykinesia PDD model
Male: 12 Upper and evaluate it
MMG/IMU/force Female: 11 features: root Pronation
100 Hz supination Parkinson’s using Unified
[48] sensor 10 healthy arm/forearm/ mean square of
movements disease
(non-invasive) wrist/hand prona- Parkinson’s
subjects Disease Rating
Male: 8 tion/supination
motion speeds, Scale scores
Female: 2
etc.
Tremor features:
means and
standard
deviations of
processed
rates-of-turn
and
accelerations
Sensors 2023, 23, 7667 8 of 30

Table 2. Cont.

Wearable Sensor
References Participants Location Feature Sampling Rate Exercise Disease Type Methods
Sensors Type
By combining
force, angular
displacement,
Fifteen healthy, RMS of the and electromyo-
Force graphic signals
sensor/angular right-handed force Side-to-side
Hinge mecha-
male subjects sensor/RMS of reaching/back with torso
[49] displacement nism/trapezius the angular - Stroke constraints as
aged and forth/
sensor/sEMG muscle displacement up and down the main body,
(non-invasive) between 22 and
30 years old sensor automatic
detection of
compensated
motion is
achieved
Shoulder
anteflex-
ion/shoulder A novel remote
exten- quantitative
sion/forearm Fugl- Meyer
24 stroke pronation and
Accelerometer/flex patients evaluation
Shoulder/elbow/ AMP /MEAN 20 Hz supina-
[50] sensor Stroke (FMA)
(non-invasive) Male: 16 wrist/fingers /RMS/JERK/ApEn tion/lumbar framework that
Female: 8 touch/wrist maps sensor
flexion and ex- data to clinical
tension/lateral FMA scores
pinch/finger
touch
Long-term
13 young Means and center of
standard Alzheimer’s pressure
Pressure sensor participants Standing/ disease/Parkinson’s
[51] Plantar deviations of all 100 Hz monitoring
(non-invasive) Male: 7 the pressure walking/siting disease/chronic system in a
Female: 6 ankle instability
data smart-shoe
form
Hand
rehabilitation
8 subjects with system that
Force sensor/flex Knuckle/ MAV/ RMS/
normal hand WL/VAR supports both
[42] sensor motor functions finger- 200 Hz Finger flexion Hand paralysis
/standard mirror therapy
(non-invasive) Male: 5 tips/palm deviation and
Female: 3
task-oriented
therapy
Instrumented
18 healthy knee sleeve and
Piezoresistive Open-chain modeled using
sensor subjects Knee - 18.75 Hz
[52] Gonarthrosis
(non-invasive) Male: 9 knee flexion an adaptive
Female: 9 enhanced RFR
model
Time domain
features:
mean/root
Using a single
mean
square/standard inertial sensor
Shoulder and supervised
deviation, etc. 100 Hz
Frequency abduc- machine
Accelerometer/ 20 patients (accelerometer) learning
gyroscope/ domain 100 Hz tion/shoulder
Musculoskeletal
[53] magnetometer Male: 8 Shoulder features: (gyroscope) flexion/wall technology to
disorders
Female: 12 maximum 25 Hz slide/wall identify and
(non-invasive) press/shoulder
frequency com- (magnetometer) classify
rotation shoulder
ponent/mean
frequency com- rehabilitation
ponent/energy activities
spectral density,
etc.
Machine
Root mean learning
square/mean/standard algorithms and
Stroke/multiple
Accelerometer/ 48 patients devia- inertia signals
gyroscope/ tion/energy/spectral Elbow flexion sclero-
Dorsal side of 256 Hz
[54] magnetometer Male: 26 en- and extension sis/cerebral collected
the elbow
Female: 22 ergy/absolute movements palsy/spinal during passive
(non-invasive)
differ- cord injury stretching are
ence/variance/SMA/SV used to grade
spasms
Self-powered
IPMC sensor
that can
distinguish
IPMC sensor Raw voltage Cough/hum/ Oropharyngeal between the
[55] - Throat - different
(non-invasive) data nod/swallow dysphagia
pressures
exerted by
throat
movements
Sensors 2023, 23, 7667 9 of 30

Table 2. Cont.

Wearable Sensor
References Participants Location Feature Sampling Rate Exercise Disease Type Methods
Sensors Type
IMUs used to
recognize the
purposeful and
Mean value of non-purposeful
movement movements in
inten- ADLs for
10 healthy and identifying and
IMU sity/smoothness Arm
[56] 12 post-stroke Fingertip/hand 100 Hz Hemiparesis promoting the
(non-invasive) of MI/average movements
volunteers use of the
acceleration
and rotation impaired limb
energy, etc. during daily
life in people
affected by
stroke
Novel method
for automatic
assessment of
SL/GD/PSP/MH/ the gait task in
25 PD patients
IMU RL/RSZ/RSY/ Parkinson’s UPDRS based
[57] and 28 healthy Ankle/shank 100 Hz Walk on only two
(non-invasive) RSX/MPV/MVV/ disease
subjects MSV/MHD shank-mounted
IMUs and 12 m
straight
walking test
System that
Heel
slide/seated provides
IMU/accelerometer/ 44 clinical Mean/median/ knee exten- patients with
gyroscope and 10 healthy standard 102.4 Hz sion/inner automatic
[58] Shin Knee disorders
deviation/ range quadri- feedback on
(non-invasive) subjects variance, etc.
ceps/straight knee
rehabilitation
leg raise
exercises
New approach
for spastic
detection in
EMG 4 healthy male Stroke/multiple hemiplegia-
[59] Lower leg EMG data - Walk affected EMG
(non-invasive) subjects sclerosis
data using the
IPANEMA BSN
in combination
with SVM

Mean Using wearable

frequency/the sensors and ML,
first 5 DFT real-time step
coefficients/the detection can be
Accelerometer combined with
36 pediatric Trunk/ first 5 maxima Idiopathic toe
[60] /gyroscope 75 Hz Walk assistive
patients sacrum/shank of DFT walking
(noninvasive) coefficients and devices for
their intervention
corresponding and motor
frequencies rehabilitation
purposes
Mean absolute Pattern
value/standard recognition of
deviation/ thumb and
variance/root index finger
IMU/EMG mean 50 Hz (IMU) Thumb and
[61] - Arm index finger Stroke gestures using
(non-invasive) square/waveform 200 Hz (EMG)
movements EMG signal
length/
zero cross- recording
ing/integrated obtained from
EMG Myo armband

An off-line
classification
approach for
EMG Mean/variance Musculoskeletal the 26
Forearm 1000 Hz Hand
[62] (non-invasive) 22 subjects of EMG/MAV, disorders or upper-limb
movement
etc. injuries ADLs included
in the
KIN-MUS UJI
dataset
A neck motion
detector
comprising a
self-powered
Triboelectric sensor Neck Cervical spine triboelectric
[63] - Neck - -
(non-invasive) movement diseases sensor set and a
deep learning
module to
recognize neck
motion
Sensors 2023, 23, 7667 10 of 30

Table 2. Cont.

Wearable Sensor
References Participants Location Feature Sampling Rate Exercise Disease Type Methods
Sensors Type
Place hands
behind the head
with ten fingers Multi-path
crossed/push convolutional
the elbows back neural network
Accelerometer 49 healthy to the (MP-CNN)
[64] -
Shoulders/back/elbows/forehead 32 Hz body/stretch Joint disease based on sensor
(non-invasive) volunteers
data for
both hands up
rehabilitation
with ten fingers training
crossed/bend recognition
over to the
left/right
Home-based
Bilateral rehabilitation
shoulder (HBR) system
17 participants
flexion with that identifies
IMU/accelerometer/ in the HBR both hands and records the
gyroscope group and 6 interlocked/ type and
[65] Wrist - 10 Hz Chronic stroke
participants in wall push
(non-invasive) frequency of
the control /move the
group rehabilitation
scapula /towel exercises
slide performed by
the user
Method based
on the artificial
neural network
Diabetes/ to classify
Pressure sensor 12 healthy walking speed
[66] Foot - 300 Hz Walk peripheral
(non-invasive) subjects
arterial disease and walking
time by using
plantar
pressure images
Device
consisting of a
single-board
computer (SBC)
Accelerometer Walk/walk
21 healthy male upstairs/walk Mobility and a six-axis
[67] /gyroscope Waist - 50 Hz sensor that
volunteers down- disorder recognizes
(non-invasive) stairs/sit/stand/lay activities
through deep
learning
algorithm
Method for
controlling a 3D
prosthetic hand
using elec-
Hand
EMG/muscle open/hand tromyographic
5 healthy male Stroke/absence
[68] sensor Arm - - data of basic
subjects. close/pinch/pointing of hand
(non-invasive) gestures and
finger
manipulating
the prosthetic
hand using
classified data
Method for
interpretable
Gait BRS-L
line/regional evaluation of
20 hemiplegic Bilateral pressure/gait lower extremity
IMU/pressure patients and 10 feet/bilateral phase/ motion data
[69] sensor 200 Hz Walk Stroke and plantar
healthy calves/bilateral accelera-
(non-invasive)
individuals thighs/waist tion/step pressure data
length/joint collected using
angle IMUs and
pressure
sensors
Arm
rehabilitation
monitor system
using an IMU
sensor placed
12 stroke Mean of the on a single
IMU patients signal/variance wrist to acquire
[70] Wrist 20 Hz Arm movement Stroke arm motion
(non-invasive) Male: 7 of the signal/
Female: 5 RMS, etc. information
and process the
data using a
machine
learning
classifier
Sensors 2023, 23, 7667 11 of 30

Table 2. Cont.

Wearable Sensor
References Participants Location Feature Sampling Rate Exercise Disease Type Methods
Sensors Type
Method used to
monitor the
progress of
12 patients rehabilitation
Foot/lower using kinematic
IMU with hip leg/upper
[71] - 60 Hz Walk Hip disorder data obtained
(non-invasive) unilateral leg/lower back from a wearable
arthroplasty sensor system
and a deep
convolutional
neural network
Flexible
5 healthy cable-driven
EMG Wrist/elbow/ full-hand
subjects - 13.33 Hz
[72] (non-invasive) Hand/arm shoulder Stroke exoskeleton to
Male: 4 flexions aid the
Female: 1 rehabilitation of
stroke patients
Online
segmentation
Short-arc method for
Chest/thigh exercise knee OA
(close to the Angle of shank (SAE)/straight rehabilitation
10 subjects leg raise monitoring that
IMU knee)/shank for SAE and Knee
[73] Male: 5 - can provide
(non-invasive) (close to ankle) QSM/angle of (SLR)/quadriceps osteoarthritis
Female: 5
of the thigh for SLR strengthening real-time
working leg mini-squats feedback to
(QSM) patients and
physical
therapists
Method used
for applying the
machine
learning
Vector of all 0 algorithm to
values, electromyo-
Specific facial
Articulatory except for 1 in
sEMG Laryngectomee expres- Absence of graphic signals
[74] (non-invasive) muscles on elements where 250 Hz
volunteer the target sions/palpating larynx of joint muscles
hemiface
face to identify
sEMG feature is
represented silent speech in
patients
undergoing a
total
laryngectomy
Method using
data collected
from a
Extension and wristband, a
flexion of the wireless
fore- three-axis
4 healthy Standard arm/rotation of accelerometer,
Accelerometer/
deviation/ the forearm and a three-axis
gyroscope/ subjects
[75] magnetometer Wrist/arm RMS/ 50 Hz about the Stroke rate gyroscope
and 4 stroke information elbow/
(non-invasive) patients combined with
entropy, etc. rotation of the partial k-means
wrist about
long axis of clustering to
forearm identify basic
movements of
the upper body
in everyday life
Sensor system
MDF/power of capable of
the spectrum/ monitoring
peak fre- Walk/run both knee motion
12 healthy and classifying
quency/maximum indoors and
IMU subjects with no outdoors/ Knee
[76] Right knee spectral ampli- 122 Hz aspects of daily
(non-invasive) reported knee travel up and osteoarthritis
tude/output living activities
pain down the stairs
range of the to aid in the
signal in the rehabilitation of
time domain patients with
knee OA

Abbreviations used in table: MMG (mechanomyography), IMU (inertial measurement unit), EMG (electromyo-
graphy), sEMG (surface electromyography), IPMC (ion-exchange polymer metal composite), IMMU (inertial
and magnetic measurement unit), RMS (root mean square), SMA (signal magnitude area), SV (signal vector
magnitude), MI (movement intensity), SL (stride length), GD (gait cycle duration), PSP (percentage swing phase),
MH (max ankle height), RL (range of lateral displacement), RSZ (range of shank Z-axis rotation), RSY (range of
shank Y-axis rotation), RSX (range of shank X-axis rotation), MPV (maximum progressive ankle), MVV (maximum
ankle vertical velocity), MSV (maximum shank Z-axis angular), MHD (ankle displacement at MH), DFT (discrete
Fourier transform), AMP (amplitude of sensor data), MEAN (mean value of sensor data), RMS (root mean square
value of sensor data), JERK (root mean square value of the derivative of sensor data), ApEn (approximate entropy
of sensor data), MAV (mean absolute value), RMS (root mean square), WL (waveform length), VAR (variance).
 Sensors 2023, 23, 7667 12 of 30

3.2. Wearable Sensor Type

Figure 4 summarizes the types of sensors used in the selected 32 papers. Muscle,
IPMC, piezoresistive sensors, etc., have less applications and appear only once in all the
articles. The top three applications were IMU, accelerometer, and EMG and gyroscope.
Among them, IMU was the most widely used and appeared in 11 documents, accounting
for more than 34%. An IMU is an inertial sensor composed of an accelerometer, gyroscope,
and magnetometer. It can collect different types of data, such as acceleration and angular
velocity values during motion to obtain more accurate motion measurement values. The
researchers then evaluated and analyzed the motion processes based on the data [77].
An accelerometer presented the second highest number of applications and appeared in
10 studies. The accelerometer can detect the linear acceleration of the carrier and the
direction of the acceleration, and it can be worn on the wrist to detect the activity of the
arm. In the same way, when worn on the leg, it can detect the condition of the leg during
walking or running movements, and the obtained data can reflect the use of the limb when
playing sports [78]. The third most used sensor was the EMG and gyroscope, appearing
Sensors 2023, 23, x FOR PEER REVIEW 14 of 33
in seven studies. EMG can be used to monitor and record myoelectric signals generated
by skeletal muscle activity [79]. Gyroscopes can monitor angular velocity changes that
occur during motion for motion posture analysis [80]. A force sensor, magnetometer, and
sensor are sensor
pressure less frequently cited in the
are less frequently literature
cited in the and were and
literature onlywere
mentioned in three papers.
only mentioned in three
The flex sensor
papers. only
The flex appeared
sensor in two papers.
only appeared in two papers.

Quantityofofeach
Figure4.4.Quantity
Figure eachsensor.
sensor.

3.3. Disease Types

3.3. Disease Types
Among the 32 selected papers we examined, a total of 24 diseases were addressed. As
Among the 32 selected papers we examined, a total of 24 diseases were addressed.
shown in Figure 5, the highest proportion is stroke, and 10 articles address this, accounting
As shown in Figure 5, the highest proportion is stroke, and 10 articles address this, ac-
for more than 25%. This was followed by Parkinson’s disease, there are three articles that
counting for more than 25%. This was followed by Parkinson’s disease, there are three
addressed rehabilitation treatment for this disease. Two articles were related to spinal cord
articles that addressed rehabilitation treatment for this disease. Two articles were related
injury, musculoskeletal disorder, multiple sclerosis, and knee osteoarthritis, respectively.
toThe
spinal cord injury, musculoskeletal disorder, multiple sclerosis, and knee osteoarthritis,
remaining 18 diseases, including chronic ankle instability, cerebral palsy, and knee
respectively. Theaddressed
disorders, were remainingby18 diseases,
fewer including
studies, chronic
all of which ankle
only had oneinstability, cerebral
article published on
palsy, and knee
the subject. disorders, were addressed by fewer studies, all of which only had one
article published on the subject.
Sensors 2023, 23, x FOR PEER REVIEW 15 of 33

Sensors 2023, 23, 7667 13 of 30

Figure 5. The proportion of various diseases.

3.4. Sensor
Figure 5. TheLocation
proportion of various diseases.
It can be observed in Table 2 that, among the locations where the sensors are worn,
3.4.
theSensor
wrists,Location
arms, and legs are the body parts that are used the most frequently. It can
It can be in
be observed observed
Figure 6inthat
Table 2 that, among
wearable sensorsthe arelocations
mostly worn whereonthe
thesensors are worn,
wrist, with seven
articles
the mentioning
wrists, arms, andthis legsinare
thethe
research. In second
body parts that areplace is the
used thearm,
mostwhich is mentioned
frequently. It can bein
six articles.
observed Moreover,
in Figure the
6 that hand is sensors
wearable also oneare of the
mostlycommon
worn placement
on the wrist, areas
withfor wearable
seven arti-
sensors, and five articles were published on this. Then, four articles address
cles mentioning this in the research. In second place is the arm, which is mentioned in six sensors located
on the shoulder,
articles. Moreover,shank,the hand andisfoot
alsoof
onetheofpatient,
the commonrespectively.
placement Three articles
areas mention
for wearable the
sen-
placement
sors, and five of articles
the sensors
wereon the elbow,
published onthigh, and lower
this. Then, four leg, andaddress
articles two articles concern
sensors locatedthe
knee,
on the back, and shank,
shoulder, head. Only
and footone of
article mentions
the patient, the throat, Three
respectively. chest, articles
sacrum,mention
trunk, neck,
the
and left side
placement of the
of the waist.
sensors onItthe
can be observed
elbow, thigh, and from the results
lower leg, andthat
twowearable sensorsthe
articles concern are
more commonly placed on the wrists, hands, legs, and shoulders; therefore,
knee, back, and head. Only one article mentions the throat, chest, sacrum, trunk, neck, and they are less
frequently
left
Sensors 2023, 23, x FOR PEER REVIEW side of theplaced
waist.onItthe
canthroat, chest, sacrum,
be observed from the trunk,
resultsneck, and waist.
that wearable IMUsare
sensors are most
16 more
of 33
commonly used for monitoring the motion and acquiring the data
commonly placed on the wrists, hands, legs, and shoulders; therefore, they are less fre- of the wrist, leg, and
arm so that
quently placedmotion
on thein throat,
these areas can
chest, be quantified
sacrum, [81]. and waist. IMUs are most com-
trunk, neck,
monly used for monitoring the motion and acquiring the data of the wrist, leg, and arm
so that motion in these areas can be quantified [81].

Figure 6. Locations where sensors are worn.

Figure 6. Locations where sensors are worn.

3.5. Rehabilitation Exercise

In the process of disease rehabilitation training, different training actions are adopted
to achieve the effect of assisting the recovery of different diseases in patients. Rehabilita-
 Sensors 2023, 23, 7667 14 of 30

3.5. Rehabilitation Exercise

In the process of disease rehabilitation training, different training actions are adopted
to achieve the effect of assisting the recovery of different diseases in patients. Rehabilitation
training actions are determined by the type of disease and the site requiring rehabilitation.
The analysis of the training movements in the rehabilitation process collected from the
selected papers helps us to understand the research hotspots of disease rehabilitation.
Among them, 19 articles addressed upper-limb movement, accounting for approximately
60% of the total; 13 articles addressed lower-limb movement, accounting for approximately
40%. The studies conducted on upper-limb movement mainly addressed the movements of
hands (fingers, wrists, and palms), arms, shoulders, and other parts; the studies conducted
on lower-limb movement primarily concerned the movements of the thighs, calves, and feet.
Taking strokes as an example, different parts of the brain can cause different degrees of
limb dysfunction, such as hemiplegia, impaired mobility, and a loss of hand function [82].
For information on rehabilitation training following a stroke, refer to Figure 7. Out of all
the papers addressing stroke rehabilitation, hand and arm movements accounted for the
33 percent, which was the highest result, followed by shoulder movements at 20 percent
Sensors 2023, 23, x FOR PEER REVIEW 17 of 33
and leg movements at 13 percent. It can be observed that the research performed on stroke
rehabilitation, based on wearable sensors and machine learning technology, mainly focused
on upper-limb movement; therefore, less research exists on lower-limb movement.

Figure 7. Rehabilitation movements for stroke.

Figure 7. Rehabilitation movements for stroke.
3.6. Feature Extraction
Feature
3.6. Feature engineering includes feature construction, extraction, and selection; the gener-
Extraction
ation of features
Feature engineeringcan includes
be used as the input
feature data for machine
construction, extraction,learning algorithms.
and selection; Feature
the gen-
eration of features can be used as the input data for machine learning algorithms. Featurefrom
extraction is the optimization of a subset of features used to extract new features
the original
extraction features as the
is the optimization ofinput
a subset[81].
ofThe features
features usedare
to further divided
extract new into time-domain
features from the
and frequency-domain features. Time-domain features can include data,
original features as the input [81]. The features are further divided into time-domain such as the
andmean
value, mean absolute
frequency-domain deviation,
features. root mean
Time-domain square,
features and
can peak value.
include The frequency-domain
data, such as the mean
feature can contain data, such as the frequency band and total power
value, mean absolute deviation, root mean square, and peak value. The frequency-domain between energy and
entropy factors [47]. In this review, feature engineering was addressed and used in 23 of the
feature can contain data, such as the frequency band and total power between energy and
selected articles. For different types of diseases, because of the different wearable sensors
entropy factors [47]. In this review, feature engineering was addressed and used in 23 of
used and the different collected data types, the selection of features was also very different.
the selected articles. For different types of diseases, because of the different wearable sen-
In an article written on studying the walking gait of patients, the mean and standard
sors used and the different collected data types, the selection of features was also very
deviation values of all the pressure data received by plantar pressure sensors were used as
different. In an article written on studying the walking gait of patients, the mean and
features [51]. There are also gait lines, regional pressures, gait phases, accelerations, step
standard deviation values of all the pressure data received by plantar pressure sensors
lengths, and joint angles that combine leg and plantar wearable sensor data as features [69].
were used as features [51]. There are also gait lines, regional pressures, gait phases, accel-
erations, step lengths, and joint angles that combine leg and plantar wearable sensor data
as features [69]. In the remaining nine articles that did not mention or use feature engi-
neering, the applied machine learning methods were neural network algorithms such as
ANN, CNN, and NN. They did not require additional feature engineering. CNNs can self-
Sensors 2023, 23, 7667 15 of 30

In the remaining nine articles that did not mention or use feature engineering, the applied
machine learning methods were neural network algorithms such as ANN, CNN, and NN.
They did not require additional feature engineering. CNNs can self-learn and efficiently
learn representative features obtained from large amounts of data by applying convolution
operations to raw input data [71].

3.7. Machine Learning Methods

The machine learning algorithm is an application of artificial intelligence. It is based
on data-trained algorithms that can automatically learn to continuously improve, make
predictions, or act on data without being explicitly programmed [25,41]. Machine learning
is divided into supervised learning and unsupervised learning behaviors. Supervised
machine learning is a process of using samples of known categories to adjust the parameters
of the classifier to achieve target performance. Unsupervised machine learning mainly
involves discovering methods to solve various problems concerning pattern recognition
from unknowns [83]. The field of application of machine learning is very broad. In the
field of medical rehabilitation, the machine learning algorithm can predict the possibility
of health outcomes by analyzing data and assist medical staff to take effective preventive
measures [84].
Table 3 summarizes the relevant information about machine learning in the selected
32 papers, including the ML algorithm, accuracy, description, and limitation.

Table 3. Machine learning algorithms in the selected papers.

References ML Algorithm Accuracy Description Limitation

The five-fold cross-validation
method was used to divide the It can recognize
84.95% (KNN) feature data and action labels
88.12% (RF) into five groups, four groups upper-limb movements. It
[46] KNN/RF/BC/SVM 85.05% (BC) were used to train, and the cannot identify lower-limb
97.79% (SVM) remaining group was used to movements.
validate the accuracy.
Classification accuracy was
assessed using multiple Lack of evaluation of a
assessments, including high number of
[47] SVM/RF/NB/DTW 87.4% to 97.6% 10-fold-stratified individuals with varying
cross-validation and 50% degrees of spinal cord
cross-validation (50% for injuries.
training, 50% for testing).
A voting classification model
85.1% (KNN(K = 1)) was established by combining A larger dataset needs to
83.0% (AB)
[48] KNN/AB/NN/RF 81.9% (NN) three basic classifiers, and a be established to reduce
73.6% (KNN(K = 3)) soft voting algorithm was errors and improve the
72.4% (RF) used to select the final UPDRS accuracy of the model.
score.
The nine extracted features
were supplied to the LDA, It Is necessary to conduct
KNN, and SVM. A five-fold actual clinical trials on
cross-validation method patients to further verify
97.58 ± 3% (SVM) divided the feature data and the universality of
[49] KNN/SVM/LDA 95.68% (KNN) action labels into five equal detection equipment and
92.38% (LDA) groups. Four groups were prediction methods in
used to train the classifiers and identifying the abnormal
the other group was used to movement patterns of
verify the accuracy of the patients.
classifiers.
Five characteristics were
extracted for each exercise.
Each exercise had 240 data The ceiling effect makes it
[50] ELM - samples, of which 200 samples difficult for doctors to
served as the training set and accurately assess the
the remaining 40 samples patient’s motor functions.
served as the test set.
Sensors 2023, 23, 7667 16 of 30

Table 3. Cont.

References ML Algorithm Accuracy Description Limitation

The input features were the
means and standard
deviations of the pressure data Only simple-activity
97.9% (SVM)
97.9% (RF) of the sensor, and then the testing was conducted,
[51] SVM/RF/GBC/NN 97.2% (GBC) features were transferred to
the multi-class support vector lacking complex-activity
98.6% (NN) detection.
machine (SVM) with the radial
basis function (RBF) core as
the classifier.
The optimal feature subset
was selected from the original
99.65% (SVM) features and each feature was Similar actions are easily
[42] KNN/SVM/DT 96.27% (KNN) tested independently to
81.73% (DT) mistaken.
evaluate the combination of
different features using the
10 × 10 cross-validation.
Each model trains 90% of the
data and tests the remaining It is not possible to fully
10% of the data. The capture the peaks and
hyperparameters of the troughs of all knee joint
[52] RFR - multivariate machine learning flexions or the magnitude
regression were optimized of internal/external
using grid search and rotational degrees of
multivariate Bayesian freedom.
optimization methods.
Training used a subset of
high-level features; two
different validation methods
were used to evaluate the Some actions are
90.9% (DT) prediction performance. misclassified as junk
[53] DT/SVM/KNN/RF 95.7% (KNN) Ten-fold cross-validation activities, and similar
97.2% (SVM) distributed all labeled data activities are easily
96.4% (RF) segments randomly and
confused.
evenly across ten sections. The
data contained in the nine
folds were trained and the
remaining data were tested.
The performance of the
76.6% (DT) classifier was tested by Due to the limited sample
91.8% (RF) leave-one cross-validation. size, it is not guaranteed to
[54] DT/RF/SVM/LDA/MLP 71.8% (SVM) Each classifier was tested
80.6% (LDA) perform well on larger
under four different
82.6% (MLP) conditions to determine an datasets.
optimal classifier.
The training data set with the
kernel function was used to
train the SVM model, and the When the cough is not
test data set was input into the strong enough, it is
model to check the accuracy. impossible to measure the
[55] SVM 95.0% The model was optimized by amplitude in the signal,
punishing parameter C and which can lead to an
gamma parameter g, which incorrect judgment.
could test the probability of
misclassifications.
Purposeful events were
randomly selected to evaluate
the generalization ability of There is a lack of data on
81.20% (SVM) the machine learning model, other fingers and an age
[56] SVM/ANN 97.06% (ANN) and then the classifier was mismatch among
trained using all the participants in this study.
parameters. The ten-fold
cross-validation method was
used to train and test the data.
Sensors 2023, 23, 7667 17 of 30

Table 3. Cont.

References ML Algorithm Accuracy Description Limitation

The evaluation and data
collection are not
Recursive feature elimination synchronized, which may
was performed on each model
lead to errors. The dataset
73.6% (SVM) to study the relationship is small and unevenly
[57] SVM/NBC/MLR 73.6% (NBC) between the number of distributed, making it
66.0% (MLR) features and the accuracy, and prone to overfitting and
to find the optimal feature resulting in errors. The
selection. selection of gait features is
not comprehensive.
Each model was evaluated
through a five-step
cross-validation process. To
SKE:86.05% (LR) avoid overfitting, the folds Partial actions obtained
96.70% (SVM) were generated by dividing the less satisfying results
[58] LR/SVM/AB/RF/DT 94.13% (AB) the date set by the patient. In
93.11% (RF) in the
91.75% (DT) the classification process, the laboratory dataset.
data set contained only
duplicates that were correctly
segmented.
The set of training vectors was
created based on EMG signal
data collected from two The results cannot
[59] SVM - different patients, and then an represent all individual
SVM was trained, and the diseases.
resulting structure could be
stored by significant settings.
The data were randomly
divided into two parts,
85.8% (SVM) training and testing, with a
74.4% (DT) More datasets are needed
82.8% (RF) ratio of four to one. The data
[60] SVM/DT/RF/KNN/MLP/GP 92.9% (KNN) were normalized to between 0 to achieve a better
85.8% (MLP) and 1 using min–max scaling. classification performance.
86.8% (GP) Five cross-validations were
used for each classifier’s
training dataset.
88.42% (SVM) The ratio of training to testing
80.09% (KNN) was 4:1, and the test set The placement position of
73.04% (NB) accuracy was displayed as the
[61] SVM/KNN/NB/ECOC/ 84.34% (ECOC) the armband has a
DA/DT/ensemble average accuracy of 10 trials. significant impact on
81.73% (DA) In order to achieve the best
82.60% (DT) signal recognition.
result, the linear kernel
85.65% (Ensemble) function was used.
The classifier was trained and
tested using TD and FD
65.4% (SVM) features. The integrated
57% (RF) approach was built with the Similar movements with
[62] SVM/RF/XGBoost/CNN/GRU 47.7% (XGBoost) four models with the best both hands can easily lead
83.6% (CNN) training performance to to confusion.
79% (GRU) evaluate methods that could
improve the performance of
individual models.
The leave-one session-out
(LOSO) policy was adopted.
The data obtained from one
session were used as the test
dataset, and the data collected
from the remaining three
[63] CNN 92.63% -
sessions were used as the
training dataset. This
procedure was repeated four
times until the data for each
session were considered as one
test dataset.
Sensors 2023, 23, 7667 18 of 30

Table 3. Cont.

References ML Algorithm Accuracy Description Limitation

Depending on the number of
layers in the middle path, the
correlation of the output of the
[64] MP-CNN 90.63% last pooling layer was More action data needs to
be collected.
captured, and the accuracy
was highest when D-CNN and
S-CNN were combined.
Cross-validation was
performed on different input There is some degree of
[65] CNN 85.6~100% and sensor data, and the
model with the most accurate data loss.
data was determined.
Flatten layer was used to
convert the image of the
plantar region into a The data set used is not
one-dimensional value comprehensive and the
[66] ANN 94% sequence. The sequence was
plantar features are not
then used as input data for the detailed.
ANN model. Hidden layers to
propagate training
mechanisms.
A feature fusion model
containing nuclei of different
sizes was used. After signal
normalization and conversion When the data
into a fixed format, the inertial characteristics of two
[67] CNN 97.49% data were divided into three actions are similar,
partitions, which were classification errors are
prone to occur.
composed of three convolution
layers and one flattened layer,
respectively.
For ANN training and testing,
a 3:1 ratio was used. The No wrist motion and no
[68] ANN 91% training and verification errors force control.
were reduced in a certain
number of iterations.
A cross-validation approach
was used to evaluate the
80.07% (RF) predictive performance of the
[69] RF/KNN/SVM/NB 94.20% (KNN) -
75.35% (SVM) classification model. The
82.43% (NB) leave-one-subject-out strategy
was used to divide the data
into training and test sets.
A validation dataset was
generated by separating 20%
of the continuous portion of
Home-Home: 77.1% (RF) the training dataset obtained Datasets are small and
[70] RF/CNN 76.6% (CNN) from each participant in a unrepresentative.
random location. The results
of the validation dataset were
used to tune the classifier
hyperparameters.
Training, validation, and test
data were randomly divided
into 70%, 15%, and 15%,
respectively. The adaptive Lack of more detailed
analysis of DCNN input
moment estimation method
[71] DCNN 98% was used for optimization. data and gait kinematics
The stop-loss criterion was data during rehabilitation
process.
applied to the training
progress by evaluating the
validation loss.
Sensors 2023, 23, 7667 19 of 30

Table 3. Cont.

References ML Algorithm Accuracy Description Limitation

The Bayesian regularization
algorithm was used to train The data set is small and
[72] NN - the neural network to the system is not a fully
minimize the internal
parameters and model errors closed loop.
and avoid overfitting.
10x cross-validation was used
to validate the data. A total of
10 rounds were performed, Patient movements cannot
90.6% (layer 1) be fully simulated, and the
[73] SVM with 1 subset of the 10 subjects
92.7% (layer 2) data are not accurate
selected for each round as the enough.
training data and the other 9
subsets as the test data.
Need to improve silent
The feature data consisted of speech recognition
vectors representing all zeros
algorithm to realize the
[74] XGBoost 86.4% in the elements that
characterized the target translation of silent speech
surface EMG signal. into personalized
synthetic speech.
The clustering was formed
using a sorted list of features; The effects of sensor
HS: 88% (DOA) therefore, a combination of fusion and other
K-means 83% (DOG) 2–30 features was selected in attachment positions need
[75] SP: 70% (DOA) turn in each iteration, and to be observed in larger
66% (DOG) 10-fold cross-validations were
performed on the selected sample populations.
feature vector ten times.
The random forest algorithm
Test data should include
was a collection of 10 other activities of daily
classification decision trees,
with 90% of the data randomly living to allow for a more
[76] RF 93%
selected for building the tree comprehensive
and 10% for testing the classification of activities
of daily living.
algorithm.
Abbreviations used in the table: KNN (K-nearest neighbors), RF (random forest), BC (Bayesian classifier), SVM
(support vector machine), DTW (dynamic time warping), NB (naive Bayes), LDA (linear discriminant analysis),
ELM (extreme learning machine), GBC (gradient boosting classifier), DT (decision tree), RFR (random forest
regressors), MLP (multilayer perceptron), ANN (artificial neural network), MLR (multiple linear regression),
LR (logistic regression), AB (adaptive boosting), SKE (seated knee extension), GP (gaussian process), ECOCs
(error correcting output codes), DA (discriminant analysis), MP-CNN (multipath convolutional neural network),
CNN (convolutional neural network), DCNN (deep convolutional neural network), HS (healthy subject), DOA
(data of accelerometer), DOG (data of gyroscope), SP (stroke patients), XGBoost (extreme gradient boosting),
UPDRS (unified Parkinson’s disease rating scale), D-CNN (dynamic convolutional neural network), S-CNN (state
transition probability convolutional neural network). The bold font in the table represents the machine learning
algorithm with the highest accuracy.

We summarized the machine learning algorithms that could be obtained from each
article from the selected 32 documents and created statistics on all types of algorithms;
the results are presented in Figure 8. The most widely used algorithm was SVM, which
was used in 17 articles, accounting for more than 53% of the 32 articles. Followed by RF,
12 articles used this method. Then there was KNN, which was used in eight articles. Seven
articles mentioned using the CNN method. Six articles mentioned using the DT method.
Four articles mentioned using the NB method. ANN and NN, respectively were mentioned
in three articles using this method. MLP, LDA, AB, and XGBoost were each mentioned
in two articles using this method. The remaining 14 machine learning algorithms were
only used in one document. From the abovementioned results, it can be concluded that
SVM is favored by researchers. SVM has the characteristics of relatively easy training data
and high accuracy; however, its shortcomings are also very obvious, such as slow learning
speed and long training time [85].
 tioned in three articles using this method. MLP, LDA, AB, and XGBoost were each men-
tioned in two articles using this method. The remaining 14 machine learning algorithms
were only used in one document. From the abovementioned results, it can be concluded
that SVM is favored by researchers. SVM has the characteristics of relatively easy training
Sensors 2023, 23, 7667 data and high accuracy; however, its shortcomings are also very obvious, such as slow 20 of 30
learning speed and long training time [85].

Numberofofdifferent
Figure8.8.Number
Figure differentmachine
machinelearning
learningalgorithms.
algorithms.

4. Discussion
4. Discussion
This systematic review included 32 papers based on wearable sensors and machine
This systematic
learning algorithms review
used to included
assess the 32 papers
degree based onofwearable
of recovery sensors
patients and assistand machine
rehabilitation
learning algorithms used to assess the degree of recovery of patients
training. On the one hand, this review summarized the relevant research results and and assist rehabilita-
tion training. that
determined On the one hand,
wearable this and
sensors review summarized
machine learningthe relevant research
algorithms results
can be better and
applied
determined that wearable sensors and machine learning algorithms can
in the course of disease rehabilitation, helping doctors to keep abreast of patients’ recoverybe better applied
instatus
the course of disease
and relieve rehabilitation,
social helping
medical pressure. Ondoctors to keep
the other hand, abreast
for theofpatients
patients’themselves,
recovery
status and relieveofsocial
the application medical
wearable pressure.
sensors On the
facilitated other
their hand, at
recovery forhome,
the patients
which themselves,
could greatly
the application
reduce of wearable
the factor sensors facilitated
of psychological burden. their recovery
Therefore, it isatnecessary
home, which could greatly
to summarize the
reduce
application results of wearable sensors and machine learning algorithms in thethe
the factor of psychological burden. Therefore, it is necessary to summarize ap-of
field
plication results of wearable
disease rehabilitation, exploresensors and machine
the limitations of thelearning
research,algorithms
and propose in the
the field of dis-of
possibility
ease rehabilitation,
future studies. explore the limitations of the research, and propose the possibility of
futureWe studies.
searched and screened papers for our analysis using the IEEE and the Web of
We searched
Science and screened
core databases. paperssection
The following for ourdiscusses
analysis(1) using the IEEEofand
the selection the Web
wearable of
sensor
Science
types in core databases. The
rehabilitation following
training; section
(2) the discusses
application (1) the selection
of machine learningofalgorithms;
wearable sensor(3) the
types in rehabilitation
analysis training;training
of the rehabilitation (2) the application
process; andof(4) machine learning algorithms;
the suggestions (3) the
for future research.
analysis of the rehabilitation training process; and (4) the suggestions for future research.
4.1. Wearable Sensor Type Selection
4.1. Wearable Sensordetermined
This review Type Selection
that wearable sensors are more frequently used for upper-limb
thanThis
lower-limb rehabilitation
review determined that purposes.
wearableDuring the rehabilitation
sensors are more frequently processusedof stroke patients,
for upper-
the recovery speed of the upper limbs was slower than that of the
limb than lower-limb rehabilitation purposes. During the rehabilitation process of strokelower limbs. During the
recovery process of the upper limbs, certain complications, such as
patients, the recovery speed of the upper limbs was slower than that of the lower limbs. shoulder pain, shoulder–
hand syndrome, and upper-limb flexor spasms often occurred. Therefore, additional studies
in the field are focusing on the upper-limb recovery of stroke patients [86]. He et al. [46] used
a nine-axis sensor, including a three-axis accelerometer and a high-sensitivity three-axis
gyroscope, in order to avoid the “drift phenomenon” caused by the lack of a magnetometer
in the upper-limb rehabilitation evaluation of stroke patients. In this way, more accurate
data can be obtained. If there is a long-term compensatory dependence on certain areas,
such as the limbs and trunk, it affects the patient’s rehabilitation outcomes [87]. Xu et al. [49]
used three different types of sensors, namely force sensor, angular displacement sensor, and
sEMG, to realize the automatic detection of compensatory motion during the rehabilitation
Sensors 2023, 23, 7667 21 of 30

process of stroke patients. This method not only predicts the movement of the patient’s
limbs, but also restricts the trunk from making relatively large compensatory movements,
improving the safety and effectiveness of the patient during rehabilitation training. In a
study performed on hand rehabilitation training for stroke patients, Yu et al. [50] used two
acceleration sensors and seven bending sensors to monitor the motor functions of the arm,
wrist, and fingers. This study comprehensively covered the upper limbs and provided
a good understanding of the overall recovery of the upper limbs. For the rehabilitation
of stroke patients’ fine hand movements, Chen et al. [42] used gloves integrating both
force and flex sensors. Compared with gloves using biomedical signals, the gloves not
only improved the signal quality, but also did not need to pay attention to the precision of
electrode placement, thereby promoting the recovery of fine motor movements in stroke
patients. Such hand rehabilitation systems can facilitate the development of IoT healthcare
in the field of home rehabilitation. Kim et al. [54] considered patients in remote areas;
therefore, they proposed in their study a wearable device equipped with a minimum
number of IMUs to collect the characteristics of spastic movements, effectively improving
the utilization rate of the device. Shradha et al. [61] improved the wearable device according
to the use conditions of EMG, installed an IMU and EMG in the armband, and did not
require the wearer to shave the hair in the area where the sensor is worn; therefore, it
was more convenient to use. The device can also be designed in the form of a wristband.
Biswas et al. [75] tracked the arm movements of stroke patients around the clock with a
wristband inertial sensor to comprehensively assess the progress of rehabilitation. The
research conducted on the rehabilitation of the lower limbs of stroke patients is often
related to gait research, and gait research requires the cooperation of multiple sensors.
Chen et al. [69] combined the plantar pressure sensor and IMU to obtain stable walking
rehabilitation data through the combination of multi-directional data.
Parkinson’s disease is a common neurodegenerative disorder characterized by tremors,
stiffness, and slowness of movement [88]. For the assessment of upper extremity symptoms
in Parkinson’s disease patients, Huo et al. [48] designed the Parkinson’s diagnostic device
(PDD) system, which can simultaneously assess three main symptoms. The PDD system
is mainly composed of IMU and MMG sensors. Combining MMG signals can effectively
improve the accuracy of symptom classification. For the gait research of Parkinson’s
disease patients, Guo et al. [51] used plantar pressure sensors to efficiently collect patient’s
plantar pressure data, and the selection of low-power sensors can effectively extend the
daily monitoring time. From the perspective of users, Han et al. [57] selected a lighter
IMU, which could reduce the patient’s exercise burden and ensure the completion of
rehabilitation training.
It has become a trend in the research to apply wearable sensors for neck disease detec-
tion and rehabilitation purposes. For oropharyngeal dysphagia, Lee et al. [55] designed
a self-powered IPMC sensor to detect throat muscle movements, which could more accu-
rately identify actions such as coughing and swallowing. An [63] et al. designed a wearable
neck device consisting of four silicone rubber triboelectric sensors and a silicone rubber
collar. This device was highly flexibility, saved energy, and was cost-effective; therefore, it
could be better used in the rehabilitation of neck diseases. Rameau [74] placed sEMGs on
five joint muscles on one side of the face of laryngectomy volunteers who did not undergo
radiotherapy. This method can realize silent speech recognition through surface muscle
signals and help patients who have undergone laryngectomy and patients with impaired
speech functions to perform speech rehabilitation techniques.
Most of the problems targeted by lower limb rehabilitation focus on lower limb
dysfunction caused by spinal cord injuries, and diseases of the knee, hip, and other joints.
In their study, Amir et al. [47] installed accelerometers on both the patient and assistive
devices (crutches, wheelchairs, etc.). The information collected by the sensors placed on the
assistive device presented a unique perspective, which combined the different perspectives
of the patient and assistive device for the motion analysis. In order to relieve the pressure
of patients with knee joint disease during the rehabilitation process, Antonio et al. [58]
Sensors 2023, 23, 7667 22 of 30

placed an IMU on the patient’s tibia to make the patient feel relaxed, and this was a labor-
saving step employed during the rehabilitation training process. Moreover, Chen et al. [73]
achieved the same effect by using a miniature inertial sensor with a lighter weight. Javier
et al. [71] placed IMUs on the pelvis, thigh, calf, and foot of patients to collect different
signals in the lower-limb gait study of hip joint disease rehabilitation training, so as to
generate a comprehensive dataset for their analysis.
According to the statistical results of this review, 11 of the 32 papers used IMUs, which
was the most frequently used sensor. IMU sensors not only have good wearability features
and can be worn on any part of the wrist, arm, shoulder, and leg, but also collect kinematic
parameters, such as body position, acceleration, and speed of motion with higher-accuracy
results [89]. Therefore, they are favored by many researchers.

4.2. Application Analysis of Machine Learning Algorithms

In Section 3, it was observed that the data analysis of the machine learning algorithm
in the process of disease rehabilitation can help predict the disease and help doctors and
patients conduct more scientific and effective rehabilitation training techniques. The use of
different machine learning algorithms enabled the comparison of analysis results based on
the data set, the features extracted as input variables, and the complexity of the different
models employed in the study.
A total of 17 papers used SVM to perform the disease rehabilitation evaluation, rehabil-
itation exercise classification, rehabilitation action recognition, and disease prediction steps.
As a result of their high-accuracy characteristics, SVMs have a wide range of applications,
and the number of SVMs used in the past decade in the field has been high compared to
other ML algorithms [90]. He et al. [46] observed that the accuracy of SVMs was higher
than that of k-nearest neighbor, RF, and Bayesian classifier algorithms for the upper-limb
rehabilitation evaluation of stroke patients. It can be observed that the SVM is better than
other machine learning classifiers for the classification of rehabilitation sports data. Due
to the limited muscle strength available during a rehabilitation exercise, compensatory
exercises inevitably occurred. Xu et al. [49] aimed at the detection of compensatory motion
in rehabilitation exercises, and observed that the automatic compensation detection of SVM
performed better than other algorithms in a normal rehabilitation exercise state. Chen
et al. [34] combined features and used high-quality classification signals in their study, and
observed that the average accuracy of the SVM algorithm was the highest. It is not difficult
to observe that features such as input signals have a greater impact on the accuracy of
machine learning algorithms. Lee et al. [55] proposed an optimized SVM algorithm based
on SVM, which can produce high-accuracy results even when the sample size is small. This
allows the SVM algorithm to cope with more diverse situations. Chen et al. [73] used a
multi-layer support vector machine model capable of online segmentation, first through
learning to extract the features that matched the target motion, and then accurately segment
and classify the motion data online.
Deep learning methods also belong to machine learning methods, and their appli-
cations in the field of disease rehabilitation are gradually increasing in the field. CNN is
an important neural network in the field of deep learning, which can be applied to many
different scenarios and has an excellent learning ability [91]. Chae et al. [65] selected the
CNN algorithm for home rehabilitation exercise detection. CNN has a high accuracy rate
for human activity recognition, and does not require special feature extraction methods, and
its classification is more streamlined than other algorithm steps. Zhu and Yen et al. [64,67]
combined multiple CNN models and compared them with a single model and observed
that the accuracy of the combined model was higher. It can be concluded that combining
CNN models is a method that can effectively improve overall accuracy. An et al. [63]
trained the model by adding data recorded under different conditions, which can also
effectively improve the accuracy of the model. Guo et al. [51] applied the collected data
features to a variety of machine learning algorithms, and the accuracy rates were generally
Sensors 2023, 23, 7667 23 of 30

similar; however, the accuracy of the neural network model was the highest, indicating
that the amount of information collected by sensors can affect the accuracy of the model.
In addition to SVM and CNN algorithms, there are a variety of machine learning
algorithms applied in the field of disease rehabilitation. Amir et al. [47] studied the two dif-
ferent perspectives of the patient and the mobility aid; therefore, algorithms, such as SVM,
Bayesian, and DT, were used to detect physical activity results. We combined the Bayesian
algorithm with joint classification algorithms, such as DTW, to detect activity patterns
while using assistive devices. Rameau et al. [74] applied training data samples to different
machine learning models, and then used the XGBoost model with the highest accuracy
rate together with validation samples to create a predictive model for language recognition
purposes. Although the abovementioned machine learning algorithms are rarely used in
the research, their advantages are obvious under certain conditions. Therefore, no fixed
machine learning algorithm is always better than other algorithms.

4.3. Rehabilitation
This review aimed to study the application of wearable sensors and machine learning
algorithms in the field of disease rehabilitation. It was necessary to discuss the training
required for various disease rehabilitation techniques.
According to the research, it can be observed that wearable sensors are most widely
used in the rehabilitation of stroke diseases. The purpose of stroke rehabilitation training is
to improve the patient’s ability to control their muscles, enhance the coordination of muscle
groups, and improve the coordination ability for daily activities and body balance [92].
The most common symptoms of stroke are a limited movement of different parts of the
body and gait disturbance. Patients require long-term intensive rehabilitation training to
help them recover effectively [93]. Different scholars have conducted targeted research on
different parts of the body of stroke patients with limited movement. For example, for the
upper-limb rehabilitation of stroke patients, He et al. [46] used three movements: hand
to lumbar spine, shoulder flexion, and forearm pronation. These three actions effectively
covered the locations of all wearable sensors, which could help them accurately evaluate the
rehabilitation of the upper limbs. Chen et al. [42] conducted research on the fine-grained
training of hand rehabilitation for stroke patients. The purpose of the training was to
improve the coordination functions of single and multiple fingers. During the rehabilitation
training process, Kim et al. [54] arranged rehabilitation trainers to guide the patients to
maintain correct movements and postures and improve the effect of rehabilitation training.
Burns et al. [72] used a full-hand exoskeleton worn on the patient’s hand to assist the patient
in grasping small items in daily life. Lower-extremity training after a stroke affects the
future mobility of patients and is also of great importance. Chen et al. [69] provided visual
feedback to patients during their rehabilitation training based on the gait characteristics
collected by sensors, visualized lower-limb movements, stimulated patients’ awareness
of gait correction autonomously, and effectively improved the quality of rehabilitation
actions. In addition, Xu et al. [49] combined torso restraints with appropriate sensors
for compensatory movements during the rehabilitation of stroke patients. The device
effectively suppresses the compensatory movement that may occur during the rehabilitation
training of the patient, and at the same time detects the movement trend of the patient
during the training process to evaluate the accuracy of their rehabilitation actions.
Patients with spinal cord injuries must experience a long-term rehabilitation phase,
which has a considerable impact on body motor functions [94]. Amir et al. [47] used various
assistive mobility devices to improve the mobility of patients with spinal cord injuries
while collecting information from the assistive devices and wearable sensors placed on the
patient. The method provides ideas for helping researchers and healthcare professionals
analyze the complex movements of patients during their rehabilitation. Guo et al. [51]
aimed at the rehabilitation of Alzheimer’s, Parkinson’s, and other diseases, because the
main rehabilitation training for such diseases lies in daily walking activity; therefore, a
smart insole was used to monitor patients’ everyday walking activity. The design of such
Sensors 2023, 23, 7667 24 of 30

insoles has good development prospects for the rehabilitation of patients at home and in
the community. Bavan et al. [53] applied five conventional rehabilitation movements for
shoulder rehabilitation: shoulder abduction, shoulder flexion, wall sliding, wall pressing,
and shoulder rotation. Among them, the four movements of shoulder abduction, shoulder
flexion, wall sliding, and wall pressing were performed in a sitting position, the purpose
of which was to reduce the compensatory movements of the other muscles during the
rehabilitation process. Soangra et al. [60] focused on children’s idiopathic toe walking
(ITW), reducing the size of the sensor and wearing it directly on the upper body. Not
only did this not limit the walking rehabilitation movement, but it also helped parents
monitor the child’s walking status in real time and presented abnormal gait occurrence.
Javier et al. [71] used gait training as the basis for hip rehabilitation training, and strictly
required patients to perform rehabilitation training once a day. For the rehabilitation of
knee osteoarthritis, Enrica et al. [76] arranged rehabilitation training for different occasions,
simulating both indoor and outdoor situations, to ensure the authenticity of the patient’s
rehabilitation data.

4.4. Propositions for Future Studies

Based on the analysis of the selected articles, this review summarized some possible
future development directions and some limitations of previous studies.

4.4.1. Participants
In terms of the selection of experimental subjects, the experimental participants pre-
sented in some papers [48,50,53,54,56–58,60,65,69–71,73–75] selected disease patients or a
combination of disease patients and healthy participants for the experimental research.
Another approach [42,46,47,49,51,55,59,62,66–68,72,76] was to recruit disease-simulated
subjects to imitate patients for exercise experiments. There was a certain gap between the
information collected by simulated subjects and the real data of patients, and it was difficult
to guarantee the authenticity and validity of the research results.

4.4.2. Multiple Sensors and Special Patients

From the types of wearable sensors summarized in Table 2, it can be observed that the
simultaneous use of multiple sensors has been a research trend in recent years [42,46,48–
50,53,54,56–58,60,61,65,67–71,73,75,76]. However, sensors worn in different positions on the
body pose a considerable challenge to data integration due to their different sampling
frequencies. A variety of sensors can be combined to monitor the different movement
trajectories of patients and comprehensively evaluate functional rehabilitation and daily-life
activities At present, gait research systems based on pressure sensors are widely used in the
field of rehabilitation training; however, insufficient attention has been paid to special foot-
type rehabilitation research [4,95,96]. Therefore, in future research, special patients with
diseases can be the key research objects. For example, in the gait research of Parkinson’s
disease patients, a new rehabilitation training system can be established for patients with
flat feet.

4.4.3. Robot-Assisted Rehabilitation System

A robot-assisted rehabilitation system is an emerging intelligent rehabilitation training
system in the field. The main function of the robot is to help patients train by simulating
normal activities. It can also be worn on the patient to force them to perform various rehabil-
itation exercises, continuously stimulating their brains, improving the ability of their motor
organs, and achieving an early recovery [97]. Robot-assisted rehabilitation systems mostly
monitor hand extension, flexion, and wrist movements in the field of stroke rehabilitation,
and there are few studies on fine movements, such as finger coordination activities.
Sensors 2023, 23, 7667 25 of 30

4.4.4. Sensor Durability

In the study of joint rehabilitation training such as those on the knee joint, wearing the
sensor at the joint position affects the accuracy of motion detection during the movement
of the joint [98,99]. The durability of the sensor is an issue worthy of consideration in the
research [100]. The repeated bending of the flexible sensor leads to a decrease in durability
of and damage to sensor function. Therefore, how to improve the durability of wearable
sensors for long-term wear is also one of the focuses of the research in the future. In
response to such problems, researchers have proposed that stretchable and flexible sensors
can be placed in joints together to reduce the bending loss of flexible sensors installed in
the joints [72]. Therefore, how to solve the durability problem of wearable sensors in other
ways is also a focus of the research to be conducted in the future.

4.4.5. Virtual Reality

The combination of VR and disease rehabilitation training has become a development
trend in the field. From the perspective of interaction, increasingly more researchers have
proposed that future rehabilitation training can use virtual scenes to improve the interest
and autonomy of patients. At present, some studies in the literature combine VR and sen-
sors to assist patients in effective exercise rehabilitation techniques [101–103]. Combining
virtual reality gaming with a network of wearable sensors to monitor a patient’s recovery
is a promising form of technology. However, the existing research is not comprehensive
and more extensive and in-depth applications are necessary, and specific designs should be
created to be applied to rehabilitation training for different diseases.

4.4.6. Machine Learning Optimization and Deep Learning Methods

Machine learning, including deep learning, is developing very rapidly, and it involves
a wide range of applications [104,105]. For machine learning in the field of disease rehabili-
tation, if a smaller data set is selected, the coverage is reduced and cannot be extended to
more people; therefore, a greater amount of high-quality training data are needed [106].
Classification training cannot be generalized. For specific patients, a machine learning algo-
rithm trained separately should be used for classification purposes, because this can affect
the overall classification effect by adjusting a single specific feature space to maximize the
recognition performance [59]. Some studies report that deep learning methods outperform
classical machine learning algorithms. Model accuracy and generality can be improved by
obtaining larger sample sizes and applying deep learning techniques [107]. Future research
can focus on using novel machine learning techniques, such as CNNs, to bypass tedious
steps, such as feature extraction calculations.

5. Conclusions
This paper reviewed the research of wearable sensors and machine learning algo-
rithms in disease rehabilitation training. It can be observed that using machine learning
algorithms to process data obtained from wearable sensors is helpful for rehabilitation
training for different diseases. Based on the results obtained by this review, it is concluded
that IMUs are the most used sensors during rehabilitation. Most of the sensors used in
disease rehabilitation are non-invasive, and the research on sensors in the field of disease
rehabilitation should also pay more attention to other types of sensors. Machine learning
algorithms such as SVM have a good auxiliary effect on data analysis and prediction in
the process of disease recovery. In order to find the optimal solution, more algorithms
should be used in experiments. In the future, other approaches can be tested to compensate
for our deficiencies and complete a more comprehensive review of wearable sensors and
machine learning algorithms in the field of medical rehabilitation. In the future, with
the development of wearable sensor technology, characteristic data can be collected for
additional diseases, so as to facilitate the understanding of the recovery status of diseases.
At the same time, machine learning algorithms are transforming the field of healthcare.
Smarter machine learning algorithms are being developed to help healthcare professionals
Sensors 2023, 23, 7667 26 of 30

improve diagnostic accuracy, predict the progression of a patient’s disease, and make
personalized treatment recommendations. Combining the two methods increases the possi-
bility of remote disease diagnosis and home rehabilitation, which may change the shortage
of medical resources at present due to the aging population, to a certain extent. This review
may not have included some relevant papers as the data were only collected from the Web
of Science and IEEE Xplore. In addition, some recent high-quality papers may not have
received enough citations.

Author Contributions: Conceptualization, Z.W. and S.W.; methodology, S.W.; investigation, S.W.;
resources, S.W.; data curation, S.W.; writing—original draft preparation, S.W.; writing—review and
editing, Z.W.; visualization, S.W.; supervision, Z.W.; funding acquisition, Z.W. All authors have read
and agreed to the published version of the manuscript.
Funding: This research was funded by the National Key Research and Development Program of
China (2018YFD0600304). This project was from the Ministry of Science and Technology of the
People’s Republic of China.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: On behalf of all the authors, the corresponding author states that there are no
conflicts of interest.

References
1. Xie, S.; Zhang, J. Sensor-Based Exercise Rehabilitation Robot Training Method. J. Sens. 2023, 2023, 7881084. [CrossRef]
2. Qiu, S.; Zhao, H.; Jiang, N.; Wang, Z.; Liu, L.; An, Y.; Zhao, H.; Miao, X.; Liu, R.; Fortino, G. Multi-Sensor Information Fusion
Based on Machine Learning for Real Applications in Human Activity Recognition: State-of-the-Art and Research Challenges. Inf.
Fusion 2022, 80, 241–265. [CrossRef]
3. Semwal, V.B.; Gupta, A.; Lalwani, P. An Optimized Hybrid Deep Learning Model Using Ensemble Learning Approach for Human
Walking Activities Recognition. J. Supercomput. 2021, 77, 12256–12279. [CrossRef]
4. Prasanth, H.; Caban, M.; Keller, U.; Courtine, G.; Ijspeert, A.; Vallery, H.; Von Zitzewitz, J. Wearable Sensor-Based Real-Time Gait
Detection: A Systematic Review. Sensors 2021, 21, 2727. [CrossRef] [PubMed]
5. Yao, S.; Vargas, L.; Hu, X.; Zhu, Y. A Novel Finger Kinematic Tracking Method Based on Skin-Like Wearable Strain Sensors. IEEE
Sens. J. 2018, 18, 3010–3015. [CrossRef]
6. Mainali, S.; Darsie, M.E.; Smetana, K.S. Machine Learning in Action: Stroke Diagnosis and Outcome Prediction. Front. Neurol.
2021, 12, 734345. [CrossRef]
7. Mennella, C.; Maniscalco, U.; De Pietro, G.; Esposito, M. The Role of Artificial Intelligence in Future Rehabilitation Services: A
Systematic Literature Review. IEEE Access 2023, 11, 11024–11043. [CrossRef]
8. Liao, Y.; Vakanski, A.; Xian, M.; Paul, D.; Baker, R. A Review of Computational Approaches for Evaluation of Rehabilitation
Exercises. Comput. Biol. Med. 2020, 119, 103687. [CrossRef]
9. Wang, Y.; Yang, B.; Hua, Z.; Zhang, J.; Guo, P.; Hao, D.; Gao, Y.; Huang, J. Recent Advancements in Flexible and Wearable Sensors
for Biomedical and Healthcare Applications. J. Phys. D Appl. Phys. 2022, 55, 134001. [CrossRef]
10. Cheng, Y.; Wang, K.; Xu, H.; Li, T.; Jin, Q.; Cui, D. Recent Developments in Sensors for Wearable Device Applications. Anal.
Bioanal. Chem. 2021, 413, 6037–6057. [CrossRef]
11. Park, Y.-G.; Lee, S.; Park, J.-U. Recent Progress in Wireless Sensors for Wearable Electronics. Sensors 2019, 19, 4353. [CrossRef]
[PubMed]
12. Stack, E.; Agarwal, V.; King, R.; Burnett, M.; Tahavori, F.; Janko, B.; Harwin, W.; Ashburn, A.; Kunkel, D. Identifying Balance
Impairments in People with Parkinson’s Disease Using Video and Wearable Sensors. Gait Posture 2018, 62, 321–326. [CrossRef]
[PubMed]
13. Kelly, D.; Esquivel, K.M.; Gillespie, J.; Condell, J.; Davies, R.; Karim, S.; Nevala, E.; Alamäki, A.; Jalovaara, J.; Barton, J.; et al.
Feasibility of Sensor Technology for Balance Assessment in Home Rehabilitation Settings. Sensors 2021, 21, 4438. [CrossRef]
[PubMed]
14. Kimoto, A.; Fujiyama, H.; Machida, M. A Wireless Multi-Layered EMG/MMG/NIRS Sensor for Muscular Activity Evaluation.
Sensors 2023, 23, 1539. [CrossRef] [PubMed]
15. Husain, K.; Mohd Zahid, M.S.; Ul Hassan, S.; Hasbullah, S.; Mandala, S. Advances of ECG Sensors from Hardware, Software and
Format Interoperability Perspectives. Electronics 2021, 10, 105. [CrossRef]
16. Dan, J.; Foged, M.T.; Vandendriessche, B.; Van Paesschen, W.; Bertrand, A. Sensor Selection and Miniaturization Limits for
Detection of Interictal Epileptiform Discharges with Wearable EEG. J. Neural Eng. 2023, 20, 016045. [CrossRef] [PubMed]
Sensors 2023, 23, 7667 27 of 30

17. Acosta-Vargas, P.; Flor, O.; Salvador-Acosta, B.; Suárez-Carreño, F.; Santórum, M.; Solorzano, S.; Salvador-Ullauri, L. Inertial
Sensors for Hip Arthroplasty Rehabilitation: A Scoping Review. Sensors 2023, 23, 5048. [CrossRef]
18. Gill, W.A.; Howard, I.; Mazhar, I.; McKee, K. A Review of MEMS Vibrating Gyroscopes and Their Reliability Issues in Harsh
Environments. Sensors 2022, 22, 7405. [CrossRef]
19. Javed, A.R.; Sarwar, M.U.; Khan, S.; Iwendi, C.; Mittal, M.; Kumar, N. Analyzing the Effectiveness and Contribution of Each Axis
of Tri-Axial Accelerometer Sensor for Accurate Activity Recognition. Sensors 2020, 20, 2216. [CrossRef]
20. Li, W.; Lu, W.; Sha, X.; Xing, H.; Lou, J.; Sun, H.; Zhao, Y. Wearable Gait Recognition Systems Based on MEMS Pressure and
Inertial Sensors: A Review. IEEE Sens. J. 2022, 22, 1092–1104. [CrossRef]
21. Sun, S.; Cao, Z.; Zhu, H.; Zhao, J. A Survey of Optimization Methods from a Machine Learning Perspective. IEEE Trans. Cybern.
2020, 50, 3668–3681. [CrossRef] [PubMed]
22. Doupe, P.; Faghmous, J.; Basu, S. Machine Learning for Health Services Researchers. Value Health 2019, 22, 808–815. [CrossRef]
[PubMed]
23. Shehab, M.; Abualigah, L.; Shambour, Q.; Abu-Hashem, M.A.; Shambour, M.K.Y.; Alsalibi, A.I.; Gandomi, A.H. Machine Learning
in Medical Applications: A Review of State-of-the-Art Methods. Comput. Biol. Med. 2022, 145, 105458. [CrossRef] [PubMed]
24. Manickam, P.; Mariappan, S.A.; Murugesan, S.M.; Hansda, S.; Kaushik, A.; Shinde, R.; Thipperudraswamy, S.P. Artificial
Intelligence (AI) and Internet of Medical Things (IoMT) Assisted Biomedical Systems for Intelligent Healthcare. Biosensors 2022,
12, 562. [CrossRef] [PubMed]
25. An, Q.; Rahman, S.; Zhou, J.; Kang, J.J. A Comprehensive Review on Machine Learning in Healthcare Industry: Classification,
Restrictions, Opportunities and Challenges. Sensors 2023, 23, 4178. [CrossRef] [PubMed]
26. Zhang, M.; Song, W.; Zhang, J. A Secure Clinical Diagnosis with Privacy-Preserving Multiclass Support Vector Machine in Clouds.
IEEE Syst. J. 2022, 16, 67–78. [CrossRef]
27. Zhang, Z. Big Data Analysis with Artificial Intelligence Technology Based on Machine Learning Algorithm. IFS 2020, 39,
6733–6740. [CrossRef]
28. Chen, K.-H.; Su, C.; Hakert, C.; Buschjäger, S.; Lee, C.-L.; Lee, J.-K.; Morik, K.; Chen, J.-J. Efficient Realization of Decision Trees for
Real-Time Inference. ACM Trans. Embed. Comput. Syst. 2022, 21, 1–26. [CrossRef]
29. Nanfack, G.; Temple, P.; Frénay, B. Constraint Enforcement on Decision Trees: A Survey. ACM Comput. Surv. 2022, 54, 1–36.
[CrossRef]
30. Luo, J.; Liu, Y.; Zhang, S.; Liang, J. Extreme Random Forest Method for Machine Fault Classification. Meas. Sci. Technol. 2021, 32,
114006. [CrossRef]
31. Schonlau, M.; Zou, R.Y. The Random Forest Algorithm for Statistical Learning. Stata J. 2020, 20, 3–29. [CrossRef]
32. Tang, K.; Luo, R.; Zhang, S. An Artificial Neural Network Algorithm for the Evaluation of Postoperative Rehabilitation of Patients.
J. Healthc. Eng. 2021, 2021, 3959844. [CrossRef] [PubMed]
33. Teslyuk, V.; Kazarian, A.; Kryvinska, N.; Tsmots, I. Optimal Artificial Neural Network Type Selection Method for Usage in Smart
House Systems. Sensors 2020, 21, 47. [CrossRef] [PubMed]
34. Yen, J.M.; Lim, J.H. A Clinical Perspective on Bespoke Sensing Mechanisms for Remote Monitoring and Rehabilitation of
Neurological Diseases: Scoping Review. Sensors 2023, 23, 536. [CrossRef]
35. Topham, L.K.; Khan, W.; Al-Jumeily, D.; Hussain, A. Human Body Pose Estimation for Gait Identification: A Comprehensive
Survey of Datasets and Models. ACM Comput. Surv. 2023, 55, 120. [CrossRef]
36. Newaz, N.T.; Hanada, E. The Methods of Fall Detection: A Literature Review. Sensors 2023, 23, 5212. [CrossRef]
37. Bhoir, A.A.; Mishra, T.A.; Narayan, J.; Dwivedy, S.K. Machine Learning Algorithms in Human Gait Analysis. In Encyclopedia of
Data Science and Machine Learning; IGI Global: Hershey, PA, USA, 2023; pp. 922–937. [CrossRef]
38. Xiang, L.; Wang, A.; Gu, Y.; Zhao, L.; Shim, V.; Fernandez, J. Recent Machine Learning Progress in Lower Limb Running
Biomechanics with Wearable Technology: A Systematic Review. Front. Neurorobotics 2022, 16, 913052. [CrossRef]
39. Jourdan, T.; Debs, N.; Frindel, C. The Contribution of Machine Learning in the Validation of Commercial Wearable Sensors for
Gait Monitoring in Patients: A Systematic Review. Sensors 2021, 21, 4808. [CrossRef]
40. Usmani, S.; Saboor, A.; Haris, M.; Khan, M.A.; Park, H. Latest Research Trends in Fall Detection and Prevention Using Machine
Learning: A Systematic Review. Sensors 2021, 21, 5134. [CrossRef]
41. Boukhennoufa, I.; Zhai, X.; Utti, V.; Jackson, J.; McDonald-Maier, K.D. Wearable Sensors and Machine Learning in Post-Stroke
Rehabilitation Assessment: A Systematic Review. Biomed. Signal Process. Control 2022, 71, 103197. [CrossRef]
42. Chen, X.; Gong, L.; Wei, L.; Yeh, S.-C.; Da Xu, L.; Zheng, L.; Zou, Z. A Wearable Hand Rehabilitation System with Soft Gloves.
IEEE Trans. Ind. Inf. 2021, 17, 943–952. [CrossRef]
43. Facciorusso, S.; Spina, S.; Reebye, R.; Turolla, A.; Calabrò, R.S.; Fiore, P.; Santamato, A. Sensor-Based Rehabilitation in Neurological
Diseases: A Bibliometric Analysis of Research Trends. Brain Sci. 2023, 13, 724. [CrossRef] [PubMed]
44. Page, M.J.; McKenzie, J.E.; Bossuyt, P.M.; Boutron, I.; Hoffmann, T.C.; Mulrow, C.D.; Shamseer, L.; Tetzlaff, J.M.; Moher, D.
Updating Guidance for Reporting Systematic Reviews: Development of the PRISMA 2020 Statement. J. Clin. Epidemiol. 2021, 134,
103–112. [CrossRef]
45. Jia, S.; Gao, H.; Xue, Z.; Meng, X. Recent Advances in Multifunctional Wearable Sensors and Systems: Design, Fabrication, and
Applications. Biosensors 2022, 12, 1057. [CrossRef] [PubMed]
Sensors 2023, 23, 7667 28 of 30

46. He, J.; Chen, S.; Guo, Z.; Pirbhulal, S.; Wu, W.; Feng, J.; Dan, G. A Comparative Study of Motion Recognition Methods for Efficacy
Assessment of Upper Limb Function. Adapt. Control Signal 2019, 33, 1248–1256. [CrossRef]
47. Amiri, A.M.; Shoaib, N.; Hiremath, S.V. A Framework to Enhance Assistive Technology Based Mobility Tracking in Individuals
with Spinal Cord Injury. In Proceedings of the 2017 IEEE Global Conference on Signal and Information Processing (GlobalSIP),
Montreal, QC, Canada, 14–16 November 2017; pp. 467–471. [CrossRef]
48. 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. [CrossRef]
49. Xu, P.; Xia, D.; Zheng, B.; Huang, L.; Xie, L. A Novel Compensatory Motion Detection Method Using Multiple Signals and
Machine Learning. IEEE Sens. J. 2022, 22, 17162–17172. [CrossRef]
50. Yu, L.; Xiong, D.; Guo, L.; Wang, J. A Remote Quantitative Fugl-Meyer Assessment Framework for Stroke Patients Based on
Wearable Sensor Networks. Comput. Methods Programs Biomed. 2016, 128, 100–110. [CrossRef]
51. Guo, R.; Cheng, X.; Hou, Z.-C.; Ma, J.-Z.; Zheng, W.-Q.; Wu, X.-M.; Jiang, D.; Pan, Y.; Ren, T.-L. A Shoe-Integrated Sensor System
for Long- Term Center of Pressure Evaluation. IEEE Sens. J. 2021, 21, 27037–27044. [CrossRef]
52. Wood, D.S.; Jensen, K.; Crane, A.; Lee, H.; Dennis, H.; Gladwell, J.; Shurtz, A.; Fullwood, D.T.; Seeley, M.K.; Mitchell, U.H.; et al.
Accurate Prediction of Knee Angles during Open-Chain Rehabilitation Exercises Using a Wearable Array of Nanocomposite
Stretch Sensors. Sensors 2022, 22, 2499. [CrossRef]
53. Bavan, L.; Surmacz, K.; Beard, D.; Mellon, S.; Rees, J. Adherence Monitoring of Rehabilitation Exercise with Inertial Sensors: A
Clinical Validation Study. Gait Posture 2019, 70, 211–217. [CrossRef] [PubMed]
54. Kim, J.-Y.; Park, G.; Lee, S.-A.; Nam, Y. Analysis of Machine Learning-Based Assessment for Elbow Spasticity Using Inertial
Sensors. Sensors 2020, 20, 1622. [CrossRef] [PubMed]
55. Lee, J.-H.; Chee, P.-S.; Lim, E.-H.; Tan, C.-H. Artificial Intelligence-Assisted Throat Sensor Using Ionic Polymer–Metal Composite
(IPMC) Material. Polymers 2021, 13, 3041. [CrossRef] [PubMed]
56. Butt, A.H.; Zambrana, C.; Idelsohn-Zielonka, S.; Claramunt-Molet, M.; Ugartemendia-Etxarri, A.; Rovini, E.; Moschetti, A.;
Molleja, C.; Martin, C.; Salleras, E.O.; et al. Assessment of Purposeful Movements for Post-Stroke Patients in Activites of Daily
Living with Wearable Sensor Device. In Proceedings of the 2019 IEEE Conference on Computational Intelligence in Bioinformatics
and Computational Biology (CIBCB), Siena, Italy, 9–11 July 2019; pp. 1–8. [CrossRef]
57. Han, Y.; Liu, X.; Zhang, N.; Zhang, X.; Zhang, B.; Wang, S.; Liu, T.; Yi, J. Automatic Assessments of Parkinsonian Gait with
Wearable Sensors for Human Assistive Systems. Sensors 2023, 23, 2104. [CrossRef] [PubMed]
58. Bevilacqua, A.; Huang, B.; Argent, R.; Caulfield, B.; Kechadi, T. Automatic Classification of Knee Rehabilitation Exercises Using a
Single Inertial Sensor: A Case Study. In Proceedings of the 2018 IEEE 15th International conference on wearable and implantable
Body Sensor Networks (BSN), Chicago, IL, USA, 19–22 May 2018; pp. 21–24. [CrossRef]
59. Lueken, M.J.; Misgeld, B.J.E.; Leonhardt, S. Classification of Spasticity Affected EMG-Signals. In Proceedings of the 2015 IEEE
12th International Conference on Wearable and Implantable Body Sensor Networks (BSN), Cambridge, MA, USA, 9–12 June 2015;
pp. 1–6. [CrossRef]
60. Soangra, R.; Wen, Y.; Yang, H.; Grant-Beuttler, M. Classifying Toe Walking Gait Patterns Among Children Diagnosed With
Idiopathic Toe Walking Using Wearable Sensors and Machine Learning Algorithms. IEEE Access 2022, 10, 77054–77067. [CrossRef]
61. Singhvi, S.; Ren, H. Comparative Study of Motion Recognition with Temporal Modelling of Electromyography for Thumb and
Index Finger Movements Aiming for Wearable Robotic Finger Exercises. In Proceedings of the 2018 3rd International Conference
on Advanced Robotics and Mechatronics (ICARM), Singapore, 18–20 July 2018; pp. 509–514. [CrossRef]
62. Salinas, S.A.; Elgalhud, M.A.T.A.; Tambakis, L.; Salunke, S.V.; Patel, K.; Ghenniwa, H.; Ouda, A.; McIsaac, K.; Grolinger, K.; Trejos,
A.L. Comparison of Machine Learning Techniques for Activities of Daily Living Classification with Electromyographic Data. In
Proceedings of the 2022 International Conference on Rehabilitation Robotics (ICORR), Rotterdam, The Netherlands, 25–29 July
2022; pp. 1–6. [CrossRef]
63. An, S.; Pu, X.; Zhou, S.; Wu, Y.; Li, G.; Xing, P.; Zhang, Y.; Hu, C. Deep Learning Enabled Neck Motion Detection Using a
Triboelectric Nanogenerator. ACS Nano 2022, 16, 9359–9367. [CrossRef]
64. Zhu, Z.-A.; Lu, Y.-C.; You, C.-H.; Chiang, C.-K. Deep Learning for Sensor-Based Rehabilitation Exercise Recognition and
Evaluation. Sensors 2019, 19, 887. [CrossRef]
65. Chae, S.H.; Kim, Y.; Lee, K.-S.; Park, H.-S. Development and Clinical Evaluation of a Web-Based Upper Limb Home Rehabilitation
System Using a Smartwatch and Machine Learning Model for Chronic Stroke Survivors: Prospective Comparative Study. JMIR
Mhealth Uhealth 2020, 8, e17216. [CrossRef]
66. Chen, H.-C.; Sunardi; Liau, B.-Y.; Lin, C.-Y.; Akbari, V.B.H.; Lung, C.-W.; Jan, Y.-K. Estimation of Various Walking Intensities
Based on Wearable Plantar Pressure Sensors Using Artificial Neural Networks. Sensors 2021, 21, 6513. [CrossRef]
67. Yen, C.-T.; Liao, J.-X.; Huang, Y.-K. Feature Fusion of a Deep-Learning Algorithm into Wearable Sensor Devices for Human
Activity Recognition. Sensors 2021, 21, 8294. [CrossRef]
68. Maheen, A.; Bin Shahzad, M.; Asif, M.U.; Ahmad, S.F.; Zafar, S.; Maqbool, H.F.; Usman, M.; Arshad, H. Human Hand
Gesture Recognition System Using Body Sensor Network. In Proceedings of the 2021 International Conference on Robotics and
Automation in Industry (ICRAI), Xi’an, China, 30 May–5 June 2021; pp. 1–5. [CrossRef]
69. Chen, X.; Hu, D.; Zhang, R.; Pan, Z.; Chen, Y.; Xie, L.; Luo, J.; Zhu, Y. Interpretable Evaluation for the Brunnstrom Recovery Stage
of the Lower Limb Based on Wearable Sensors. Front. Neuroinform. 2022, 16, 1006494. [CrossRef] [PubMed]
Sensors 2023, 23, 7667 29 of 30

70. Gomez-Arrunategui, J.P.; Eng, J.J.; Hodgson, A.J. Monitoring Arm Movements Post-Stroke for Applications in Rehabilitation and
Home Settings. IEEE Trans. Neural Syst. Rehabil. Eng. 2022, 30, 2312–2321. [CrossRef] [PubMed]
71. Alcaraz, J.C.; Moghaddamnia, S.; Penner, M.; Peissig, J. Monitoring the Rehabilitation Progress Using a DCNN and Kinematic
Data for Digital Healthcare. In Proceedings of the 2020 28th European Signal Processing Conference (EUSIPCO), Amsterdam,
The Netherlands, 18–21 January 2021; pp. 1333–1337. [CrossRef]
72. Burns, M.K.; Pei, D.; Vinjamuri, R. Myoelectric Control of a Soft Hand Exoskeleton Using Kinematic Synergies. IEEE Trans.
Biomed. Circuits Syst. 2019, 13, 1351–1361. [CrossRef] [PubMed]
73. Chen, H.-C.; Liu, K.-C.; Chan, C.-T. Online Segmentation with Multi-Layer SVM for Knee Osteoarthritis Rehabilitation Monitoring.
In Proceedings of the 2016 IEEE 13th International Conference on Wearable and Implantable Body Sensor Networks (BSN), San
Francisco, CA, USA, 14–17 June 2016; pp. 55–60. [CrossRef]
74. Rameau, A. Pilot Study for a Novel and Personalized Voice Restoration Device for Patients with Laryngectomy. Head Neck 2020,
42, 839–845. [CrossRef] [PubMed]
75. Biswas, D.; Cranny, A.; Gupta, N.; Maharatna, K.; Achner, J.; Klemke, J.; Jöbges, M.; Ortmann, S. Recognizing Upper Limb
Movements with Wrist Worn Inertial Sensors Using K-Means Clustering Classification. Hum. Mov. Sci. 2015, 40, 59–76. [CrossRef]
[PubMed]
76. Papi, E.; Spulber, I.; Kotti, M.; Georgiou, P.; McGregor, A.H. Smart Sensing System for Combined Activity Classification and
Estimation of Knee Range of Motion. IEEE Sens. J. 2015, 15, 5535–5544. [CrossRef]
77. Zhang, M.; Xu, X.; Chen, Y.; Li, M. A Lightweight and Accurate Localization Algorithm Using Multiple Inertial Measurement
Units. IEEE Robot. Autom. Lett. 2020, 5, 1508–1515. [CrossRef]
78. Regterschot, G.R.H.; Selles, R.W.; Ribbers, G.M.; Bussmann, J.B.J. Whole-Body Movements Increase Arm Use Outcomes of
Wrist-Worn Accelerometers in Stroke Patients. Sensors 2021, 21, 4353. [CrossRef]
79. Zhao, S.; Liu, J.; Gong, Z.; Lei, Y.; OuYang, X.; Chan, C.C.; Ruan, S. Wearable Physiological Monitoring System Based on
Electrocardiography and Electromyography for Upper Limb Rehabilitation Training. Sensors 2020, 20, 4861. [CrossRef]
80. Lim, S.; D’Souza, C. Measuring Effects of Two-Handed Side and Anterior Load Carriage on Thoracic-Pelvic Coordination Using
Wearable Gyroscopes. Sensors 2020, 20, 5206. [CrossRef]
81. Chen, M.; Wang, H.; Yu, L.; Yeung, E.H.K.; Luo, J.; Tsui, K.-L.; Zhao, Y. A Systematic Review of Wearable Sensor-Based
Technologies for Fall Risk Assessment in Older Adults. Sensors 2022, 22, 6752. [CrossRef] [PubMed]
82. Miao, S.; Shen, C.; Feng, X.; Zhu, Q.; Shorfuzzaman, M.; Lv, Z. Upper Limb Rehabilitation System for Stroke Survivors Based on
Multi-Modal Sensors and Machine Learning. IEEE Access 2021, 9, 30283–30291. [CrossRef]
83. Gomes, M.A.S.; Kovaleski, J.L.; Pagani, R.N.; Da Silva, V.L. Machine Learning Applied to Healthcare: A Conceptual Review. J.
Med. Eng. Technol. 2022, 46, 608–616. [CrossRef] [PubMed]
84. Kumar, Y.; Koul, A.; Singla, R.; Ijaz, M.F. Artificial Intelligence in Disease Diagnosis: A Systematic Literature Review, Synthesizing
Framework and Future Research Agenda. J. Ambient. Intell. Hum. Comput. 2023, 14, 8459–8486. [CrossRef] [PubMed]
85. Wang, P.; Fan, E.; Wang, P. Comparative Analysis of Image Classification Algorithms Based on Traditional Machine Learning and
Deep Learning. Pattern Recognit. Lett. 2021, 141, 61–67. [CrossRef]
86. saacs-Itua, A.; Wong, S. Stroke Rehabilitation and Recovery. Br. J. Hosp. Med. 2021, 82, 1–7. [CrossRef]
87. Luo, Z.; Lim, A.E.-P.; Durairaj, P.; Tan, K.K.; Verawaty, V. Development of a Compensation-Aware Virtual Rehabilitation System for
Upper Extremity Rehabilitation in Community-Dwelling Older Adults with Stroke. J. Neuroeng. Rehabil. 2023, 20, 56. [CrossRef]
88. Pérennou, D.; Dai, S.; Gastaldi, R.; Fraix, V.; Leroux, N.; Clarac, E.; Davoine, P.; Piscicelli, C.; Krack, P. Retropulsion with Tilted
Postural Vertical Causing Backward Falls in an Individual with Parkinson’s Disease: Improvement by Specific Rehabilitation.
Ann. Phys. Rehabil. Med. 2023, 66, 101728. [CrossRef]
89. Brognara, L.; Palumbo, P.; Grimm, B.; Palmerini, L. Assessing Gait in Parkinson’s Disease Using Wearable Motion Sensors: A
Systematic Review. Diseases 2019, 7, 18. [CrossRef]
90. Roy, A.; Chakraborty, S. Support Vector Machine in Structural Reliability Analysis: A Review. Reliab. Eng. Syst. Saf. 2023, 233,
109126. [CrossRef]
91. Li, Z.; Liu, F.; Yang, W.; Peng, S.; Zhou, J. A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects.
IEEE Trans. Neural Netw. Learning Syst. 2022, 33, 6999–7019. [CrossRef]
92. Xuefang, L.; Guihua, W.; Fengru, M. The Effect of Early Cognitive Training and Rehabilitation for Patients with Cognitive
Dysfunction in Stroke. Int. J. Methods Psych. Res. 2021, 30, e1882. [CrossRef] [PubMed]
93. Tavares, E.; Coelho, J.; Rogado, P.; Correia, R.; Castro, C.; Fernandes, J.B. Barriers to Gait Training among Stroke Survivors: An
Integrative Review. JFMK 2022, 7, 85. [CrossRef] [PubMed]
94. Rao, M.; Li, Y.; Liu, H.; Wang, I.; Ren, Y. Influencing Factors Analysis of Rehabilitation for Patients with Spinal Cord Injury. Intell.
Autom. Soft Comput. 2022, 34, 455–466. [CrossRef]
95. Chen, Y.-J.; Wu, C.-M.; Chen, P.-C.; See, A.R.; Chen, S.-C. Pressure-Sensor-Based Gait Analysis for Disabled People. Sens. Mater.
2022, 34, 225. [CrossRef]
96. Pan, Z.; Gao, H.; Chen, Y.; Xie, Z.; Xie, L. Evaluation of Hemiplegic Gait Based on Plantar Pressure and Inertial Sensors. IEEE
Sens. J. 2023, 23, 12008–12017. [CrossRef]
97. Wu, J.; Liu, Y.; Zhao, J.; Zang, X.; Guan, Y. Research on Theory and a Performance Analysis of an Innovative Rehabilitation Robot.
Sensors 2022, 22, 3929. [CrossRef]
Sensors 2023, 23, 7667 30 of 30

98. Di Tocco, J.; Carnevale, A.; Presti, D.L.; Bravi, M.; Bressi, F.; Miccinilli, S.; Sterzi, S.; Longo, U.G.; Denaro, V.; Schena, E.; et al.
Wearable Device Based on a Flexible Conductive Textile for Knee Joint Movements Monitoring. IEEE Sens. J. 2021, 21, 26655–26664.
[CrossRef]
99. Franco, T.; Sestrem, L.; Henriques, P.R.; Alves, P.; Varanda Pereira, M.J.; Brandão, D.; Leitão, P.; Silva, A. Motion Sensors for Knee
Angle Recognition in Muscle Rehabilitation Solutions. Sensors 2022, 22, 7605. [CrossRef]
100. Wang, S.; Deng, W.; Yang, T.; Tian, G.; Xiong, D.; Xiao, X.; Zhang, H.; Sun, Y.; Ao, Y.; Huang, J.; et al. Body-Area Sensor Network
Featuring Micropyramids for Sports Healthcare. Nano Res. 2023, 16, 1330–1337. [CrossRef]
101. Proffitt, R.; Ma, M.; Skubic, M. Novel Clinically-Relevant Assessment of Upper Extremity Movement Using Depth Sensors. Top.
Stroke Rehabil. 2023, 30, 11–20. [CrossRef] [PubMed]
102. Postolache, O.; Hemanth, D.J.; Alexandre, R.; Gupta, D.; Geman, O.; Khanna, A. Remote Monitoring of Physical Rehabilitation of
Stroke Patients Using IoT and Virtual Reality. IEEE J. Select. Areas Commun. 2021, 39, 562–573. [CrossRef]
103. Ramasamy, P.; Calderon-Sastre, E.; Renganathan, G.; Das, S.; Kurita, Y. Soft Actuators-Based Skill Training Wearables: A Review
on the Interaction Modes, Feedback Types, VR Scenarios, Sensors Utilization and Applications. Robomech J. 2023, 10, 1. [CrossRef]
104. Shafee, A.; Awaad, T.A. Privacy Attacks against Deep Learning Models and Their Countermeasures. J. Syst. Archit. 2021, 114,
101940. [CrossRef]
105. Qian, Y. School of Science, Hubei University of Technology, Wuhan 430068, China Exploration of Machine Algorithms Based on
Deep Learning Model and Feature Extraction. MBE 2021, 18, 7602–7618. [CrossRef] [PubMed]
106. Kang, J.; Ullah, Z.; Gwak, J. MRI-Based Brain Tumor Classification Using Ensemble of Deep Features and Machine Learning
Classifiers. Sensors 2021, 21, 2222. [CrossRef] [PubMed]
107. Chen, K.; Zhang, D.; Yao, L.; Guo, B.; Yu, Z.; Liu, Y. Deep Learning for Sensor-Based Human Activity Recognition: Overview,
Challenges, and Opportunities. ACM Comput. Surv. 2022, 54, 1–40. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.