MBE, 19(8): 7899–7919.

DOI: 10.3934/mbe.2022369
Received: 03 April 2022
Revised: 24 April 2022
Accepted: 25 April 2022
Published: 27 May 2022
http://www.aimspress.com/journal/MBE

Review

A scoping review on monitoring mental health using smart wearable

devices

Nannan Long1,2, Yongxiang Lei3, Lianhua Peng2,4, Ping Xu5 and Ping Mao1,6,*

1
Department of Nursing, The Third Xiangya Hospital, Central South University, Changsha 410013,
China
2
Xiangya Nursing School, Central South University, Changsha 410031, China
3
Department of Mechanical Engineering, Politecnico di Milano, Milan 10056, Italy
4
Affiliated Hospital of Jinggangshan University, Jianggangshan 343100, China
5
ZiBo Hospital of Traditional Chinese and Western Medicine, Zibo 255020, China
6
Hunan Key Laboratory of Nursing, Changsha 410013, China

* Correspondence author: Email: ping.mao@csu.edu.cn.

Abstract: With the continuous development of the times, social competition is becoming increasingly
fierce, people are facing enormous pressure and mental health problems have become common. Long-
term and persistent mental health problems can lead to severe mental disorders and even death in
individuals. The real-time and accurate prediction of individual mental health has become an effective
method to prevent the occurrence of mental health disorders. In recent years, smart wearable devices
have been widely used for monitoring mental health and have played an important role. This paper
provides a comprehensive review of the application fields, application mechanisms, common signals,
common techniques and results of smart wearable devices for the detection of mental health problems,
aiming to achieve more efficient and accurate prediction for individual mental health, and to achieve
early identification, early prevention and early intervention to provide a reference for improving the
level of individual mental health.

Keywords: artificial intelligence; mental health; psychological prediction; signal indicator

Abbreviation: ACC: Accelerometer; ANN: Artificial neural network; BP: Blood pressure; BVP:
Blood volume pulse; CNN: Convolutional neural network; DNN: Deep neural network; ECG:
 7900

Electrocardiograph; EDA: Electrodermal activity; EEG: Electroencephalography; EMG:

Electromyogram; GSR: Galvanic skin response; HR: Heart rate; HRV: Heart rate variability; IBI: Inter-
beat interval; KNN: K-nearest neighbor: LASSO: Least absolute shrinkage selection operator; PCA:
Principal component analysis; PPG: Photoplethysmography; RF: Random forest; RR: Respiration rate;
RBF: Radial basis function kernel classifier; ST: Skin temperature; SVM: Support vector machine

1. Introduction

Mental health problems are also known as mental health disorders, and they have become an
important factor affecting individual growth and social stability [1]. According to the “Global Burden
of Disease Study”, the global prevalence of mental health disorders has always been above 10%, and
about 450 million people have mental health disorders [2,3]. Long-term and persistent mental health
problems increase the risk of cardiovascular disease, which is the leading cause of increased mortality
worldwide [4]. Mental health disorders not only damage the physical and mental health of individuals,
but they also increase the overall medical burden on society [5]. It is predicted that the total global
medical expenditure for mental health problems will reach 16 trillion US dollars by 2030 [6]. Therefore,
it is imperative to detect the mental health condition of individuals at an early stage and prevent the
occurrence of mental health problems.
Smart sensors can accurately detect subtle changes (such as insomnia, headache, rapid heartbeat
and muscle tension) in the body caused by stress in order to prevent and reduce the adverse effects of
mental health problems [7,8]. With the increasing need for mental health care, the application of
portable smart devices for monitoring mental health has received widespread attention and has been
welcomed by most people [9]. At present, there are many studies on the use of wearable devices for
mental health monitoring. Previous reviews have integrated the use of wearable sensors to detect
stress levels, and there is a lack of exhaustive reviews on detection topics such as mental health problems
or mental illness [8,10–13]. However, although some of the studies covered comprehensive information,
the classification of detected signals were not described in detail in these reviews [14–17]. To fill this
gap, this paper analyzes and integrates a large number of studies, covering the types of equipment used
for data collection, common locations of sensors, common signals, advantages and disadvantages of
technology, types of mental health states, etc. In this section, we discuss some existing studies.
Abbreviations of nomenclatures used in this review paper are enumerated at the beginning of the article.
First, smart wearable devices have been widely used in the study of classifying and identifying
individuals’ daily stress and mental fatigue. Betti et al. [18] measured the electroencephalography
(EEG), electrocardiography (ECG) and electrodermal activity (EDA) of healthy adults using wearable
devices and a support vector machine (SVM) classification algorithm for statistical analysis; the results
show that the accuracy of the classification algorithm based on salient features can reach 86% for
detecting individual stress levels. Goumopoulos et al. [19] proposed a method to detect mental fatigue;
they used a wearable device to collect the heart rate variability (HRV) of healthy subjects, and then
selected an appropriate machine learning model to predict mental fatigue with high accuracy.
Additionally, the role of wearable devices in the prediction and classification of mental health
assessment and emotional states cannot be ignored. Sano et al. [20] explored the relationships between
the academic performance, sleep quality, stress perception and mental health of college students using
wearable devices and smartphones; they found that college students’ academic performance, sleep
quality and stress perception can significantly predict their mental health level, and that the accuracy

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7901

rate of the algorithm is as high as 92%. Costa et al. [21] found that wearable devices can acquire and
analyze social signals and identify the current negative emotions of individuals, and that, when it
identifies a negative emotion, the device activates the emotional regulation system to help the
individual stabilize their emotions. Overall, wearable devices have been widely used in the field of
mental health with strong practicability. The main contributions of this paper are given as follows:
1) Recent artificial intelligence (AI)-based wearable devices methods and techniques are
reviewed in this paper. This study comprehensively explores the application of AI-based wearable
devices in the field of mental health, focusing on stress levels, mental health problems and mental
illness.
2) This study covers comprehensive and detailed information, including types and application
mechanisms for AI-based wearable devices, common locations of sensors, common signals, the
advantages and disadvantages of technology and types of mental health states.
3) The articles included in this paper are from 2017 to 2022, and the articles are relatively new
and cutting-edge.
The remainder of this paper is organized as follows. Section 2 introduces the review of body
signals and mental health. Section 3 summarizes the methodology and Section 4 gives the literature
survey. Finally, application, discussion and conclusion are given in Sections 5, 6 and 7.

2. Body signals and mental health

The autonomic nervous system has two branches: the sympathetic nervous system (SNS) and
parasympathetic nervous system (PNS). The SNS can rapidly stimulate some vital organs (such as the
heart, lungs, and sweat glands), and the PNS can inhibit this excitation; they complement the regulation
of important bodily functions and maintenance of physiological homeostasis [22]. When an individual
is stressed, the individual will activate the protective mechanism; the SNS will be activated, the heart
rate will increase, the breathing will increase, sweat will be secreted and other physiological systems
will also be affected [23,24]. Additionally, the hypothalamus axis is a hormonal system that is initiated
in response to stressors and drives a cascade of endocrine responses, culminating in the secretion of
cortisol [22]. Therefore, individual health, especially mental health, can be known by observing the
activity of the SNS and hypothalamus axis. Finally, language signals can also reflect an individual's
mental health condition. Language is the basis for inferring individual thought processes; it provides a
special perspective for thinking observations [25]. When an individual is in a state of tension, the
respiratory rate increases and the glottis pressure increases, prompting an increase in the pitch of the
vocalized part [25]. Therefore, clinicians mainly diagnose and treat mental illness through language and
pronunciation [26]. As mentioned above, subtle changes in the body caused by stress can be detected by
sensors connected to the body and hardware devices to accurately predict mental illness [27]. Therefore,
AI-based wearable devices can collect one or more signals of the individual to identify and analyze
the mental health level.

3. Methodology

3.1. Search strategy

This review incorporated searches for relevant papers in the databases of Web of Science, Embase,

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7902

MEDLINE and Engineering Village. The selected papers ranged from 2017 and 2022, which covers
the recent developments in this area. The basic sets of keywords were “Psychological assessment”,
“psychological prediction”, “psychological monitoring”, “mental health”, “psychological health”,
“mental illness”, “psychological disorders” and “wearable device”. A total of 1537 papers were
retrieved by using all keywords, and 25 papers were finally included. The search methodology for this
article was designed to provide a detailed review of published research papers on wearable sensor-
based mental health detection. Figure 1 shows a flowchart describing the search strategy used for this
review article.

3.2. Inclusion criteria

1) The date of publication of the selected papers was between 2017 and 2022. 2) Conference
papers and journal papers were included. 3) The research direction for the paper was mental health
detection (focusing on stress, mental health problems, mental illness, etc.) and AI-based wearable
devices for data collection.

3.3. Exclusion criteria

1) Research published in a language other than English was excluded. 2) Studies structured as
editorials, press releases, research highlights or letters, review articles, reviews and technical papers were
excluded. 3) Duplicate papers, related review papers and papers not related to the topic were excluded.

Figure 1. Flow diagram of the systematic search strategy.

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7903

4. Literature survey

4.1. Wearable devices for mental health monitoring

We use smart/embedded wearable devices, such as smartwatches, that collect signals from the
body [28]. This data can be processed by sensors on smart/embedded devices. WiFi is used to send the
collected data to cloud servers [7]. In cloud computing, advanced AI models can be applied to mental
health monitoring big data to generate simple knowledge through the use of expensive processing
and advanced AI algorithms [29], such as Welch's algorithm [30], SVM algorithms [31] and period
graphs [32], for the analysis of data. Using these AI algorithms, then by analyzing the information, we
will be able to extract sensor data with the appropriate classification characteristics and then classify
their mental health states [33]. Artificial intelligence algorithms such as information and SVMs are
used again. The main function of an analog classifier is to classify different classifiers [34]. By ensuring
the rapidity of calculation, feature patterns are transformed into related classes and the classes most
suitable for a single feature are predicted. During the classification process, each feature will be
analyzed to find the most suitable feature [35]. They will be analyzed separately to predict stresses.
Once the classification is complete, the smart device displays their mental health rating on the screen.
Therefore, people can improve their mental health through self-management (as shown in Figure 2).
The virtual reality-augmented reality (VR-AR) based digital twin is the critical element for the
intelligent development of AI-based wearable devices [36]. Finally, all of the elements are integrated
into a smart device’s ecology for the future direction framework.

Figure 2. Intelligent ecology flowchart of AI-based wearable devices with data storage,
transaction, interaction and communication networks.

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7904

Mental health detection systems based on wearable devices mainly consist of sensors, data
acquisition, data pre-processing, feature extraction, machine learning algorithms and mental health
condition classification. The general scheme of mental health prediction is shown in Figure 3. Section 4
will systematically introduce the current related research work according to the workflow of the wearable
mental health detection system, as well as classify, compare and statistically analyze the existing
research results.

Figure 3. General scheme for the analysis of mental health.

4.2. Data acquisition devices

Currently, the commonly used signals in mental health detection are HRV, EEG, galvanic skin
response (GSR), skin temperature (ST), blood pressure (BP), sleep, salivary cortisol and blood oxygen
saturation (SpO2). In most studies, researchers use multiple indicators to accurately reflect an
individual's mental health. At the same time, the wearable device can also collect indicators such as
steps, activity category, posture and activity trajectory. According to the needs of users and the
functions achieved by the device, the device will be worn on different parts of the body. Common
wearing positions are shown in Figure 4.

Figure 4. Common positions of AI-based wearable devices.

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7905

4.3. Pre-processing and feature extraction

To solve the noise phenomenon in the data acquisition module, data pre-processing is required to
reduce the impact of noise on the performance of mental health detection. Common filtering algorithms
in mental health detection systems are wavelet transforms, the Kalman filter, etc. These filters can filter
out high-frequency noise signals in the data [37,38].
The pre-processed data need to extract corresponding features for training the mental health
detection model. Among them, the common feature extraction method is to use the sliding window
algorithm to extract the features in the corresponding window, and then put the trained features into
the trainer for training to form a mental health detection model. Table 1 summarizes the features
extracted by the feature extraction module of an existing detection system. The commonly used feature
values for feature extraction are Mean, Standard Deviation (StDev), Skew, Root Mean Square (RMS),
Kurtosis, Quart1, Quart3, Inter-Quartile Range (IRQ), Sum, Frequency, etc.

Table 1. Statistical features commonly used in mental health.

References Feature Description

[19,39–51] Mean
Mean   si2 / n
i
N
sd   xi    / ( N  1)
2
[19,39–47,49–51] StDev
i 1
3
n x x
n  
[40,45] Skew G1   i
(n  1)(n  2) i 1 s 
[19,40,43] RMS rmΔ( s )  s 2j / x
i

[39,40,45,51] Kurtosis K   /
[39,40,45] Quart1 Median of lower 25% of the values
[39,40,45] Quart3 Median of up 25% of the values
[39,40,45,51] IRQ Difference between Quart1 and Quart3
N

[41,42,44] Sum S   wi
i 1
n
sd 2  1/ n  xi   
2
[45] Variance
i 1

[43] Coefficient of variation CV  sd / 

Inverse coefficient of
[45] ICV   / sd
variation
Peak frequency, low-frequency power (LF), high-frequency
power (HF), very-low-frequency power (VLF), prevalent
[18,19,41,43,46,47, 50–
Frequency low frequency (pLF), prevalent high frequency (pHF), the
52]
ratio of LF to HF (LF/HF), (From Lomb–Scargle) LF, HF
and LF/HF

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7906

Table 2. Comparison analysis summary of several existing machine learning techniques.

Machine learning
Definition Advantages Disadvantages
algorithms
Works well with any complex
Sensitive to missing data; It
In order to better find the problem, can scale high-
is difficult to implement
hyperplane that can divide the dimensional data, risk of
Support Vector large-scale training
feature space, the low- overfitting is lower; It shows
Machine (SVM) samples; There are
dimensional data are mapped to high robustness and strong
difficulties in solving the
the high-dimensional plane generation ability with low
multi-classification problem
overfitting
Simple thought, mature theory;
In the feature space, if k samples
Can be used for multiple
K-Nearest Neighbor near a sample belong to a Poor generalization ability
classification, high accuracy, no
(KNN) category, that sample also and slow prediction speed
assumptions on the data, not
belongs to that category
sensitive to outliers
It can handle high-dimension
In noisy classification or
data, has strong model
regression problems,
generalization ability, and is
A classifier for multiple decision overfitting and attribute
Random Forests (RF) fast in training. It can balance
trees weights produced based on
errors in unbalanced data sets,
data of different attribute
maintain accuracy if a large
values are not credible
part of features is lost
It has a high accuracy of Many parameters are
classification, strong parallel required, which are not
A mathematical model for
distribution processing ability, suitable for small samples,
Artificial Neural distributed parallel information
strong distribution storage and high calculation cost, long
Network (ANN) processing that mimics the
learning ability, adaptability, learning process, and
features of human neural network
strong robustness and fault intermediate results cannot
tolerance to noise nerve be observed
The computational cost is low, It is easy to underfit, and
The result of linear regression is
the implementation is easy to the performance effect is
mapped between 0 and 1 by a
understand, the robustness to not good when the
Logical Regression sigmoid function. The result of
small noise in the data is good, classification accuracy is
(LR) the mapping is the probability
and it is not particularly not high, the data features
that the data sample point
affected by slight are missing or the feature
belongs to a certain category
multicollinearity space is large
Easy to understand and explain,
visualization of tree structure,
A prediction model is a mapping Overfitting, poor
less data required for training,
Decision Tree (DT) between object attributes and generalization ability and
fast analysis speed, and strong
object values unstable calculation results
ability to process samples with
missing attributes

Continued on next page

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7907

Machine learning
Definition Advantages Disadvantages
algorithms
The process is simple and fast, Sample attributes do not
Bayesian Network Graphical networks based on efficient for multi-classification work well when they are
(BN) probabilistic inference problems, and requires a associated and are sensitive
smaller sample size to input forms
The labeled data is projected to a
lower dimensional space so that
It can reflect the differences
the points after projection are
between samples and use the
Linear Discriminant distinguished by category. Points Easy to fit, restricted by
prior knowledge experience of
Analysis (LDA) of the same category are closer to sample type
categories in dimension
each other in the space after
reduction
projection, and the differences
are as far as possible
The amount of information
A method of multivariate must be kept at a high level
statistical analysis in which Eliminate the mutual influence after dimensionality
Principal Component
several variables are transformed between indicators, small reduction, and the
Analysis (PCA)
linearly to select a smaller workload interpretation of principal
number of important variables components is usually
fuzzy

4.4. Machine learning techniques

Machine learning technology refers to the method of using some special formulas to guide the
computer to use the known data to build a model that can make predictions for new situations. Common
machine learning algorithms for mental health detection include SVMs [18,19,39,43,44,48,51–55] k-
nearest neighbor (KNN) algorithms [19,48,51,53], random forests (RF) [40,45,51], artificial neural
networks (ANNs) [39], logistic regression (LR) [19,42,56,50,51], decision trees (DTs), Bayesian
networks (BNs) [48], linear discriminant analysis (LDA) [51] and principal component analysis
(PCA) [47,51]. The SVM is the most used algorithm for mental health detection among them. This
paper summarizes the artificial intelligence algorithms involved in mental health detection and briefly
describes their definitions, advantages and disadvantages, as shown in Table 2.

5. Application

Mental health problems are often accompanied by obvious body responses, and wearable sensors
play an important role in mental health and its related applications. They can collect an individual's
EEG, heart rate (HR), HRV, GSR, BP, body temperature, respiratory rate, etc., to monitor mental health.
This section summarizes and discusses the signals commonly used in mental health detection, and
briefly summarizes the relevant research on AI-based wearable devices in mental health detection over
the past five years. The main contents include mental health disease types, parameters, results, etc., as
shown in Table 3.

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7908

Table 3. Mental health applied research model and summary of results.

Study Type Device Signals Model Result

End
Samsung Gear S, S2 Diastolic PCA, LDA,
Can et al., The accuracy of the multi-layer
Stress and Empatica E4 Velocity, SVM, KNN,
2019 [51] perceptron algorithm is 92.19%.
wristband HRV and RF and LR
ACC
HRV, EDA,
The classification algorithm based
Betti et al., Mobile EEG headset, EEG and
Stress SVM on salient features has an accuracy
2018 [18] Chest belt Salivary
of 86%.
cortisol
Egilmez et al., The pressure detection F-value of
Stress LG smartwatch GSR RF
2017 [40] the system is 88.8%.
Experiments with 55 days of real
HR, BVP,
Gjoreski et Empatica device’s data showed that the method
Stress IBI, ST and SVM
al., 2019 [39] wristband detected stress events with 95%
ACC
accuracy.
The accuracy of the classification
of pressure states using SVM
Ahn et al., Head-mounted
Stress ECG, EEG SVM technology and the 5-fold cross
2019 [52] electrode
validation of EEG and HRV
features was 87.5%.
Mental stress index is sensitive to
acute stress and can predict the
Wu et al., Cortisol,
Stress Smart shirt LR level of association between
2019 [50] HRV
normal individuals and the stress
group with about 97% accuracy.
Jesmin et al., HRV, GSR Multi-sensor data fusion can better
Stress Smartwatch SVM, KNN
2020 [53] and SpO2 monitor individual stress levels.
Using machine learning
technology, the accuracy of Class
Chest-worn device, III and Class II classification
ACC, ECG,
RespiBAN Machine problems reached 81.65 and
Bobade et al., BVP, ST,
Stress Professional and a Learning 93.20%, respectively, and using
2020 [41] RR, EMG
wrist-worn device deep learning technology, the
and EDA
Empatica E4 accuracy of Class III and Class II
classification problems reached
84.32 and 95.21%, respectively.
Several parameters of the HRV
Silva et al.,
Stress Smartwatch HRV ANN under pressure were significantly
2020 [49]
different from baseline.

Continued on next page

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7909

Study Type Device Signals Model Result

The overall classification accuracy
Kim et al., Empatica E4 of the model was 85.3%, and the
Stress GSR LR
2020 [42] wristband cross-validation accuracy was
83.2%.
The accuracy of 10-fold cross-
Smart wrist band, validation was 94.55%, while the
Han et al., Empatica E4 ECG, PPG KNN, SVM, accuracy of subject cross-
Stress
2020 [48] wristband and a and GSR and BN validation was 85.71%. In daily
Shimmer ECG device settings, the system had 81.82%
accuracy in assessing stress.
HR increased with both sensory
Tonacci et Shimmer™,Shimmer Pan Tompkins (olfactory) and cognitive
Stress GSR, ECG
al.,2019 [62] 3™ algorithm stimulation, and there was a
similar decrease in HRV.
Overall, HRV characteristics
Stress and OMsignal smart-shirt calculated from wearable vest data
Tiwari et
anxiety and a Fitbit Charge 2 HRV SVM (with higher temporal resolution)
al.,2019 [43]
smart-bracelet proved to be more effective in
predicting stress and anxiety.
Method had an average accuracy
ECG, GSR,
Patlar Stress and PCA, Feed- of 92 and 89% in distinguishing
Wrist-worn wearable ST, SpO2,
Akbulut et anxiety forward neural the stress levels of the
device Blood sugar
al., 2020 [47] network experimental group and the
and BP
control group, respectively.
The sensors were 78.3% accurate
Stress and SC, ST, ACC in classifying students as high or
Sano et al., Q-sensor and Motion SVM, LASSO
mental health and Motion low stress groups, and 87%
2018 [54] logger and RBF
condition logger accurate in classifying students as
high or low mental health groups.
There was significant variability in
Fraiwan et Mental health HRV
Smart glove ST and HRV skin resistance and HRV under
al.,2018 [65] condition algorithm
relaxation and stimulation.
The use of wearable long-term
speech-to-social data sets to
Chen et al., Mental health Transfer
Smartwatch Speech categorize various sound scenes in
2018 [68] condition learning
unconstrained and natural
environments yielded good results.
The MP150 multi- The classifier achieved 81.82%
Wen et al., HR (ECG
Anxiety channel physiological SVM accuracy in detecting a high
2020 [44] electrodes)
recorder anxiety state.

Continued on next page

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7910

Study Type Device Signals Model Result

General physical activity was
O’Brien et Depression significantly reduced in depressed
Wrist-mounted ACC LR
al., 2017[56] older adults compared with
device
healthy controls.
Participants' ecological transient
assessment self-report and passive
perception (sensor data) were
Motion Machine highly correlated between physical
Narziev et Depression
Smartwatch logger and learning activity, mood and sleep levels;
al.,2020 [67]
HR Depression detector verified the
feasibility of grouping and
classification with an accuracy
rate of 96.00% (StDev = 2.76).
Patients with depression were
Zanella- accurately detected in 86.7% of
Depression Motion
Calzada et Actiwatch RF the cases, while those without
Logger
al.,2019 [45] depression were found in 91.9% of
the cases.
After comparing the HRV features
HRV, of the emotional states before and
Gentili et Bipolar
Sensorized t-shirt respiration SVM after, the classification accuracy of
al.,2017 [55] disorder
and ACC the SVM in the target state was
significantly improved.
The modal fusion performance of
the multi-mode deep neural
Head/wearable eye- EEG and
Emotion Multi-mode network was significantly
Zheng et tracking glasses, Eye-
recognition deep neural improved, and the average
al.,2019 [46] EmotionMeter movement
network model accuracy of the four emotions
hardware signal
(happiness, sadness, fear and
neutral) reached 85.11%.
Mental fatigue caused by
Goumopoulos Zephyr HXM-BT,
SVM，KNN cognitive load can be detected by
et al., 2022 Mental fatigue Chest-strap built-in HRV
and LR wearable commercial devices and
[19] ECG sensor
a single biomarker.
Through the fatigue evaluation
Wavelet experiment of 10 healthy subjects,
Qiu et al., Head-mounted ECG, EEG transform, it was verified that the device can
Mental fatigue
2020 [38] devices and ST Kalman reliably detect the mental fatigue
filtering state by monitoring and analyzing
EEG, ECG and proximal ST data.

5.1. Mental health detection using EEG

EEG is a detection method that records the electrical activity of the cerebral cortex by connecting
multiple electrodes to the scalp [57] to reflect the emotional changes of individuals to specific stimuli.

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7911

The stress state of the human body is mainly identified by the changes in the alpha and delta bands
of the EEG. When people are under high levels of stress, alpha decreases and delta increases in the
EEG, whereas, when they are relaxed or inactive, alpha increases and delta decreases [58]. Ahn et
al. [52] used a binaural hanging device that combined EEG and ECG signals and found that the
accuracy of EEG and ECG signals in continuous stress assessment was 87.5%. Additionally, based on
the multi-modal emotion recognition framework of EmotionMeter, some scholars found that the
individual's happiness, sadness, fear and neutral emotions can be distinguished and identified by EEG
signals [46]. Finally, EEG combined with ECG signals, body temperature and other data can reliably
detect the mental fatigue state of an individual, which provides support for the popularization and
specific implementation of mental health assessments [38]. Table 3 shows the studies and their details
on the use of EEG for mental health detection. EEG is one of the most effective physiological signals
for detecting psychological stress, and the combination of ECG signals and machine learning
applications can improve the accuracy of psychological stress detection.

5.2. Mental health detection using ECG

Stressors induce changes in myocardial activity, promote blood distribution to vital organs and
cause changes in the HR and HRV [59]. ECG is a reliable technique that uses electrodes attached to the
body to record the electrical potential produced by heart muscle cells to observe changes in HRV [59].
Studies have shown that HR and HRV signals are closely related to mental health problems (such as
anxiety and depression) [60], and they have an important role in mental health detection. Tiwari et
al. [43] found that HRV data obtained through smart t-shirts and smart bracelets can more accurately
assess the degree of anxiety and depression of individuals. Second, HRV characteristics are closely
associated with an individual’s clinical course and mood changes. Gentili et al. [55] showed that the
HRV characteristics (including time domain, frequency domain, nonlinearity) of patients with bipolar
disorder can significantly predict their clinical course and emotional changes. Additionally, Akbulut et
al. [47] believe that the stress state of an individual can be accurately identified by signals such as the
HRV obtained by wearable devices. Table 3 presents an overview and the details of mental health
testing using ECG.

5.3. Mental health detection using GSR

Sweat is a biological fluid produced by sweat glands. Sweat glands are distributed almost all over
the body, with high-density distribution in the palms, soles and forehead; they are only stimulated by
the sympathetic nerves, which are activated under tension [14], so the level of sweat secretion is the
body’s response to stress. Since sweat contains ions, small molecules and even macromolecular
substances such as proteins, it has electrical conductivity [61]. Therefore, the EDA signal is obtained
by connecting two electrodes on the skin surface, and the voltage between them is measured to monitor
the stress. Sriamprakash et al. [33] collected the GSR, HRV and other indicators, extracted individual
stress characteristics by using Welch's algorithm and periodograms, etc., and used SVM and KNN
algorithms to classify levels of mental health. Tonacci et al. [62] captured the ECG and skin
conductance (SC) signals of subjects by using wearable sensors and compared the effects of cognitive
stress and olfactory stimuli on human mental health. Additionally, EDA and GSR also change with
different emotions; this can effectively reflect the state of emotional calm or emotional arousal [63].

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7912

Evidence shows that the emotional state of patients with metabolic syndrome can be analyzed by using
indicators such as the GSR, HRV and body temperature detected by wearable devices [47]. The above
studies have confirmed that combining GSR with HRV, body temperature and other parameters can
improve the prediction rate for individual stress, and accurately provide and detect information about
individual emotional states.

5.4. Mental health detection using ST

High levels of stress can cause muscle tension, increase the HR and increase the work of the vital
organs; additionally, the blood will be shunted from the limbs to directly act on the vital organs. In just
a few minutes, the body surface temperature can change [64]. Most of the studies involved the
collection of body temperature data to detect mental health conditions. Fraiwan et al. [65] designed a
mobile mental health monitoring system for continuous monitoring of changes in ST, skin electrical
impedance and HRV; they found that, when the body was exposed to relaxation or stimulation, the skin
electrical resistance value and HRV exhibited significant variability. Engelniederhammer et al. [66]
monitored the EDA and ST in subjects wearing a smart wristband (Bodymonitor); the results showed
that being in a crowded environment triggers aversive emotional responses (i.e., stress, anger and fear)
in subjects. Sano et al. [54] had 201 college students wear two sensors on each wrist: a Q-sensor (for
measuring SC, ST, triaxial acceleration (ACC)) and motion recorder (for measuring acceleration and
ambient light data); the study showed that wearable sensor eigenvalues are more accurate than cell
phones for classifying poor mental health (87%) and stress (78.3%) when using machine learning
techniques. In conclusion, the ST, combined with parameters such as EDA, is more accurate for
detecting emotional changes and stress levels.

5.5. Mental health detection using speech and behavior

With the development of artificial intelligence, the natural language processing program of the
computer can more accurately identify the subtle barriers in language before the onset of mental illness,
achieving the effect of accurately predicting the mental illness [27]. Corcoran et al. [27] confirmed that
the automatic learning of speech classifiers with reduced semantic consistency, increased semantic
consistency differences and reduced use of case pronouns can predict psychotic episodes with an
accuracy rate of 83%. Language signals help to improve the accuracy of mental disease prediction and
have become a powerful tool for psychiatric diagnosis and treatment.
In addition, mental illness is associated with changes in an individual’s behavior, especially physical
movement and social habits. AI-based wearable devices (accelerometers, global position system sensors)
can detect these abnormal behaviors and effectively reduce the incidence of individual mental illness.
Zanella-Calzada et al. [45] obtained data on motor activity recorded by wearable devices from a
depression database, extracted the statistical features and then fed them to an RF classifier to detect
subjects with depression; the results showed that those subjects with the presence of depression had
an 86.7% likelihood of being correctly classified, while the specificity showed that those subjects with
the absence of depression had a 91.9% likelihood of being classified with a correct response. Additionally,
Narziev et al. [67] recruited 20 participants from four different levels of depression groups and
established a machine learning model to automatically classify the categories of depression in a short
period of time; the results of this pilot study revealed high correlations between participants’ ecological

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7913

momentary assessment self-reports and passive sensing (sensor data) in terms of physical activity, mood
and sleep levels. All of the above studies have shown that the role of speech signals and behavioral
signals in the detection of mental health states such as depression, mood disorders or schizophrenia
cannot be ignored.

6. Discussion

Multiple studies have integrated the use of wearable sensors to detect stress levels, and there is a
lack of exhaustive reviews on topics such as mental health or illness detection. However, some reviews
comprehensively cover information, but do not provide details such as machine learning methods,
types of mental health states and the classification of detection signals. To fill this gap, this paper
presents an extensive review of mental health detection using wearable sensors.
The significant observations from this review are listed below.
 Most wearable sensors are placed on the wrist (such as smartwatches and smart wristbands),
chest (smart t-shirts, chest straps, etc.) and head (head-mounted devices, etc.) for data collection.
 The Empatica, EmotionMeter and SHIMMER platforms have been popularly used for data
collection.
 HRV, EEG, GSR and ST are the most distinctive signals for detecting mental health conditions.
 SVM, LR and KNN have been widely used in the classification of mental health conditions.
 Most of the disease types studied focused on psychological stress levels and mental health
problems (anxiety and depression).
The identified challenges are described below.
 Compliance: The collection of subjects' physiological indicators and daily activity data is
affected by individual living habits, which makes the detection process more complicated and thus
degrades the performance of the mental health detection system. Therefore, the mental health
assessment based on wearable devices can only do a rough screening and cannot directly assess the
disease accurately.
 Singleness of indicators: most studies involve the collection of only a single category of
signals, such as physiology, speech and behavior, while fewer collect a combination. Since different
categories of data and signals reflect different aspects of mental health, the integrated collection of all
types of signals is essential for the accurate screening of mental health problems.

7. Conclusions

Globally, mental health problems are still serious social public health problems, and mental health
has always been detected via subjective and time-consuming methods. Wearable sensors have great
potential as a continuous monitoring system for mental health. As the needs for mental health care
services increase, so does the production of AI-based wearable devices. This paper reviewed the types
of equipment, common locations of sensors, common signals, advantages and disadvantages of
technology and types of mental health states in previous studies. First of all, the most common types
of states in mental health testing are psychological stress and mental health problems (anxiety and
depression), and most wearable sensors are placed on the wrist, chest or head. Additionally, it was
found that there are many AI-based wearable devices on the market that mainly collect physiological
signals for mental health detection, and that there are few studies on detecting mental health with

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7914

speech signals and behavioral signals. Among the physiological signals, the HRV, EEG, ST and GSR
signals are the most effective parameter indicators for detecting the mental health condition and this
conclusion is helpful for researchers to create devices to improve people's mental health level. Finally,
this paper integrated important information, such as the technical advantages and limitations of previous
studies; it was found that the most common classifiers are SVM, KNN and LR algorithms, and that these
algorithms provide the most reliable results for mental health detection. In the future, with the continuous
innovation of science and technology, deterministic annealing neural networks [69,70], convolutional
neural networks [71–73], the logarithmic descent direction algorithm [74], Laplacian-Hessian
regularization semi-supervised extreme learning machine [75,76] and back-propagation neural
networks [77] are also expected to be applied and explored in the field of mental health detection. In
a word, the goal of mental health testing is to develop an efficient and affordable model with high
accuracy. This review will help future researchers design optimal wearable devices to achieve their
goals for mental health detection.

Acknowledgement

This work is supported by the Changsha Natural Science Foundation No. kq2202422 and the
Chinese Scholarship Council (CSC) under grant No. 202006370101.

Conflict of interest

All authors declare no conflicts of interest in this paper.

References

1. L. Kesner, J. Horacek, Global adversities, the media, and mental health, Front. Psychiatry, 12
(2021), 809239. https://doi.org/10.3389/fpsyt.2021.809239
2. E. D. Vecchia, M. M. Costa, E. Lau, Urgent mental health issues in adolescents, Lancet Child
Adolesc. Health, 3 (2019), 218–219. https://doi.org/10.1016/S2352-4642(19)30069-0
3. M. L. Wainberg, P. Scorza, J. M. Shultz, L. Helpman, J. J. Mootz, K. A. Johnson, et al., Challenges
and opportunities in global mental health: A research-to-practice perspective, Curr. Psychiatry
Rep., 19 (2017), 28. https://doi.org/10.1007/s11920-017-0780-z
4. J. Vogel, A. Auinger, R. Riedl, H. Kindermann, M. Helfert, H. Ocenasek, Digitally enhanced
recovery: Investigating the use of digital self-tracking for monitoring leisure time physical activity
of cardiovascular disease (CVD) patients undergoing cardiac rehabilitation, PloS One, 12 (2017),
e0186261. https://doi.org/10.1371/journal.pone.0186261
5. F. B. Oliveira, M. R. C. Moreira, J. F. Lima, D. C. Nascimento, F. M. S. Silva, J. Antunes,
Articulation of mental health public policies and solidarity economy-Generating initiatives of
labour and income, Holos, 34 (2018), 130–140. https://doi.org/10.15628/holos.2018.5233
6. Report: Mental illness will cost the world $16 trillion (USD) by 2030, Mental Health Wkly., 28
(2018), 7–8. https://doi.org/10.1002/mhw.31630
7. E. Jovanov, A. O’Donnell Lords, D. Raskovic, P. G. Cox, R. Adhami, F. Andrasik, Stress
monitoring using a distributed wireless intelligent sensor system, IEEE Eng. Med. Biol. Mag., 22
(2003), 49–55. https://doi.org/10.1109/MEMB.2003.1213626

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7915

8. S. Gedam, S. Paul, A review on mental stress detection using wearable sensors and machine
learning techniques, IEEE Access, 9 (2021), 84045–84066.
https://doi.org/10.1109/ACCESS.2021.3085502
9. B. Pardamean, H. Soeparno, A. Budiarto, B, Mahesworo, J. Baurley, Quantified self-using
consumer wearable device: Predicting physical and mental health, Healthcare Inf. Res., 26 (2020),
83–92. https://doi.org/10.4258/hir.2020.26.2.83
10. M. Elgendi, C. Menon, Assessing anxiety disorders using wearable devices: Challenges and future
directions, Brain Sci., 9 (2019), 50. https://doi.org/10.3390/brainsci9030050
11. S. Lee, H. Kim, M. Park, H. J. Jeon, Current advances in wearable devices and their sensors in
patients with depression, Front. Psychiatry, 12 (2021), 672347–672347.
https://doi.org/10.3389/fpsyt.2021.672347
12. H. Hunkin, D. L. King, I. T. Zajac, Wearable devices as adjuncts in the treatment of anxiety related
symptoms: A narrative review of five device modalities and implications for clinical practice, Clin.
Psychol., 26 (2019), e12290. https://doi.org/10.1111/cpsp.12290
13. L. Sequeira, S. Perrotta, J. LaGrassa, K. Merikangas, D. Kreindler, D. Kundur, et al., Mobile and
wearable technology for monitoring depressive symptoms in children and adolescents: A scoping
review, J. Affective Disord., 265 (2020), 314–324. https://doi.org/10.1016/j.jad.2019.11.156
14. M. Kang, K. Chai, Wearable sensing systems for monitoring mental health, Sensors, 22 (2022),
994. https://doi.org/10.3390/s22030994
15. B. A. Hickey, T. Chalmers, P. Newton, C. Lin, D. Sibbritt, C. S. McLachlan, et al., Smart devices
and wearable technologies to detect and monitor mental health conditions and stress: A systematic
review, Sensors, 21 (2021), 10. https://doi.org/10.3390/s21103461
16. C. Su, Z. Xu, J. Pathak, F. Wang, Deep learning in mental health outcome research: A scoping
review, Trans. Psychiatry, 10 (2020), 116. https://doi.org/10.1038/s41398-020-0780-3
17. S. Patel, K. E. A. Saunders, Apps and wearables in the monitoring of mental health disorders, Br.
J. Hosp. Med., 79 (2018), 672–675. https://doi.org/10.12968/hmed.2018.79.12.672
18. S. Betti, R. M. Lova, E. Rovini, G. Acerbi, L. Santarelli, M. Cabiati, et al., Evaluation of an
integrated system of wearable physiological sensors for stress monitoring in working
environments by using biological markers, IEEE Trans. Biomed. Eng., 65 (2018), 1748–1758.
https://doi.org/10.1109/TBME.2017.2764507
19. C. Goumopoulos, N. Potha, Mental fatigue detection using a wearable commodity device and
machine learning. Journal of Ambient Intelligence and Humanized Computing, 261 (2022), 1–19.
https://doi.org/10.1007/s12652-021-03674-z
20. A. Sano, A. Phillips, A. Z. Yu, A. W. McHill, S. Taylor, N. Jaques, et al., Recognizing academic
performance, sleep quality, stress level, and mental health using personality traits, wearable
sensors and mobile phones, in International Conference on Wearable and Implantable Body
Sensor Networks, (2015), 1–6. https://doi.org/10.1109/BSN.2015.7299420
21. J. Costa, M. F. Jung, M. Czerwinski, F. Guimbretière, T. Le, T. Choudhury, Regulating feelings
during interpersonal conflicts by changing voice self-perception, in Proceedings of the 2018 CHI
Conference on Human Factors in Computing Systems, (2018), 1–13.
https://doi.org/10.1145/3173574.3174205
22. S. Rotenberg, J. J. McGrath, Inter-relation between autonomic and HPA axis activity in children
and adolescents, Biol. Psychol., 117 (2016), 16–25.
https://doi.org/10.1016/j.biopsycho.2016.01.015

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7916

23. A. E. Draghici, J. A. Taylor, The physiological basis and measurement of heart rate variability in
humans, J. Physiol. Anthropol., 35 (2016), 22. https://doi.org/10.1186/s40101-016-0113-7
24. J. M. Karemaker, An introduction into autonomic nervous function, Physiol. Meas., 38 (2017),
R89. https://doi.org/10.1088/1361-6579/aa6782
25. P. Rajasekaran, G. Doddington, J. Picone, Recognition of speech under stress and in noise, in
IEEE International Conference on Acoustics, Speech, and Signal Processing, (1986), 733–736.
https://doi.org/10.1109/ICASSP.1986.1169207
26. N. C. Andreasen, W. M. Grove, Thought, language, and communication in schizophrenia:
Diagnosis and prognosis, Schizophr. Bull., 12 (1986), 348–359.
https://doi.org/10.1093/schbul/12.3.348
27. C. M. Corcoran, F. Carrillo, D. Fernández-Slezak, G. Bedi, C. Klim, D. C. Javitt, et al., Prediction
of psychosis across protocols and risk cohorts using automated language analysis, World
Psychiatry, 17 (2018), 67–75. https://doi.org/10.1002/wps.20491
28. V. J. Madhuri, M. R. Mohan, R. Kaavya, Stress management using artificial intelligence, in 2013
Third International Conference on Advances in Computing and Communications, (2013), 54–57.
https://doi.org/10.1109/ICACC.2013.97
29. S. Graham, C. Depp, E. E. Lee, C. Nebeker, X. Tu, H. Kim, et al., Artificial intelligence for mental
health and mental illnesses: An overview, Curr. Psychiatry Rep., 21 (2019), 116.
https://doi.org/10.1007/s11920-019-1094-0
30. P. Welch, The use of fast Fourier transform for the estimation of power spectra: A method based
on time averaging over short, modified periodograms, IEEE Trans. Audio Electroacoust., 15
(1967), 70–73. https://doi.org/10.1109/TAU.1967.1161901
31. Q. Shi, H. Zhang, Fault diagnosis of an autonomous vehicle with an improved SVM algorithm
subject to unbalanced datasets, IEEE Trans. Ind. Electron., 68 (2021), 6248–6256.
https://doi.org/10.1109/TIE.2020.2994868
32. P. Kokoszka, T. Mikosch, The periodogram at the Fourier frequencies, Stochastic Processes Their
Appl., 86 (2000), 49–79. https://doi.org/10.1016/S0304-4149(99)00086-1
33. S. Sriramprakash, V. D. Prasanna, O. V. R. Murthy, Stress detection in working people, Procedia
Comput. Sci., 115 (2017), 359–366. https://doi.org/10.1016/j.procs.2017.09.090
34. J. R. Parker, Evaluating classifier combination using simulated classifiers, Department of
Computer Science, University of Calgary, Res. Rep., 2000 (2000).
http://dx.doi.org/10.11575/PRISM/30938
35. B. Moore, Principal component analysis in linear systems: Controllability, observability, and
model reduction, IEEE Trans. Autom. Control, 26 (1981), 17–23.
https://doi.org/10.1109/TAC.1981.1102568
36. A. Fernández-Caballero, E. Navarro, P. Fernández-Sotos, P. González, J. J. Ricarte, J. Latorre, et
al., Human-avatar symbiosis for the treatment of auditory verbal hallucinations in schizophrenia
through virtual/augmented reality and brain-computer interfaces, Front. Neuroinform., 11 (2017),
64. https://doi.org/10.3389/fninf.2017.00064
37. Y. Lei, H. R. Karimi, L. Cen, X. Chen, Y. Xie, Processes soft modeling based on stacked
autoencoders and wavelet extreme learning machine for aluminum plant-wide application,
Control Eng. Prac., 108 (2021), 104706. https://doi.org/10.1016/j.conengprac.2020.104706

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7917

38. X. Qiu, F. Tian, Q. Shi, Q. Zhao, B. Hu, Designing and application of wearable fatigue detection
system based on multimodal physiological signals, in 2020 IEEE International Conference on
Bioinformatics and Biomedicine (BIBM), (2020), 716–722.
https://doi.org/10.1109/BIBM49941.2020.9313129
39. M. Gjoreski, M. Luštrek, M. Gams, H. Gjoreski, Monitoring stress with a wrist device using
context, J. Biomed. Inf., 73 (2017), 159–170. https://doi.org/10.1016/j.jbi.2017.08.006
40. B. Egilmez, E. Poyraz, W. Zhou, G. Memik, P. Dinda, N. Alshurafa, UStress: Understanding
college student subjective stress using wrist-based passive sensing, in 2017 IEEE International
Conference on Pervasive Computing and Communications Workshops (PerCom Workshops),
(2017), 673–678. https://doi.org/10.1109/PERCOMW.2017.7917644
41. P. Bobade, M. Vani, Stress detection with machine learning and deep learning using multimodal
physiological data, in 2020 Second International Conference on Inventive Research in Computing
Applications (ICIRCA), (2020), 51–57. https://doi.org/10.1109/ICIRCA48905.2020.9183244
42. J. Kim, J. Park, J. Park, Development of a statistical model to classify driving stress levels using
galvanic skin responses, Hum. Factors Ergon. Manuf. Serv. Ind., 30 (2020), 321–328.
https://doi.org/10.1002/hfm.20843
43. A. Tiwari, R. Cassani, S. Narayanan, T. H. Falk, A comparative study of stress and anxiety
estimation in ecological settings using a smart-shirt and a smart-bracelet, in 2019 41st Annual
International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC),
(2019), 2213. https://doi.org/10.1109/EMBC.2019.8857890.
44. W. Wen, G. Liu, Z. Mao, W. Huang, X. Zhang, H. Hu, et al., Toward constructing a real-time
social anxiety evaluation system: Exploring effective heart rate features, IEEE Trans. Affect.
Comput., 11 (2020), 100–110. https://doi.org/10.1109/TAFFC.2018.2792000
45. L. A. Zanella-Calzada, C. E. Galvan-Tejada, N. M. Chavez-Lamas, M. del Carmen Gracia-Cortes,
R. Magallanes-Quintanar, J. M. Celaya-Padilla, et al., Feature extraction in motor activity signal:
Towards a depression episodes detection in unipolar and bipolar patients, Diagnostics, 9 (2019),
8. https://doi.org/10.3390/diagnostics9010008
46. W. Zheng, W. Liu, Y. Lu, B. Lu, A. Cichocki, EmotionMeter: A multimodal framework for
recognizing human emotions, IEEE Trans. Cybern., 49 (2019), 1110–1122.
https://doi.org/10.1109/TCYB.2018.2797176
47. F. P. Akbulut, B. Ikitimur, A. Akan, Wearable sensor-based evaluation of psychosocial stress in
patients with metabolic syndrome, Artif. Intell. Med., 104 (2020), 101824.
https://doi.org/10.1016/j.artmed.2020.101824
48. H. J. Han, S. Labbaf, J. L. Borelli, N. Dutt, A. M. Rahmani, Objective stress monitoring based on
wearable sensors in everyday settings, J. Med. Eng. Technol., 44 (2020), 177–189.
https://doi.org/10.1080/03091902.2020.1759707
49. E. Silva, J. Aguiar, L. P. Reis, J. O. Sá, J. Gonçalves, V. Carvalho, Stress among Portuguese
medical students: The EuStress solution, J. Med. Syst., 44 (2020), 45–45.
https://doi.org/10.1007/s10916-019-1520-1
50. W. Wu, S. Pirbhulal, H. Zhang, S. Mukhopadhyay, Quantitative assessment for self-tracking of
acute stress based on triangulation principle in a wearable sensor system, IEEE J. Biomed. Health
Inf., 23 (2019), 703–713. https://doi.org/10.1109/JBHI.2018.2832069

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7918

51. Y. S. Can, N. Chalabianloo, D. Ekiz, C. Ersoy, Continuous stress detection using wearable sensors
in real life: Algorithmic programming contest case study, Sensors, 19 (2019), 1849.
https://doi.org/10.3390/s19081849
52. J. W. Ahn, Y. Ku, H. C. Kim, A novel wearable EEG and ECG recording system for stress
assessment, Sensors, 19 (2019), 1991. https://doi.org/10.3390/s19091991
53. S. Jesmin, M. S. Kaiser, M. Mahmud, Towards artificial intelligence driven stress monitoring for
mental wellbeing tracking during COVID-19, in 2020 IEEE/WIC/ACM International Joint
Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT), (2020), 845–851.
https://doi.org/10.1109/WIIAT50758.2020.00130
54. A. Sano, S. Taylor, A. W. McHill, A. J. K. Phillips, L. K. Barger, E. Klerman, et al., Identifying
objective physiological markers and modifiable behaviors for self-reported stress and mental
health status using wearable sensors and mobile phones: Observational study, J. Med. Int. Res.,
20 (2018), e9410. https://doi.org/10.2196/jmir.9410
55. C. Gentili, G. Valenza, M. Nardelli, A. Lanatà, G. Bertschy, L. Weiner, et al., Longitudinal
monitoring of heartbeat dynamics predicts mood changes in bipolar patients: A pilot study, J.
Affect. Disord., 209 (2017), 30–38. https://doi.org/10.1016/j.jad.2016.11.008
56. J. T. O’Brien, P. Gallagher, D. Stow, N. Hammerla, T. Ploetz, M. Firbank, et al., A study of wrist-
worn activity measurement as a potential real-world biomarker for late-life depression, Psychol.
Med., 47 (2017), 93–102. https://doi.org/10.1017/S0033291716002166
57. U. Rashid, I. K. Niazi, N. Signal, D. Taylor, An EEG experimental study evaluating the
performance of Texas Instruments ADS1299, Sensors, 18 (2018), 3721.
https://doi.org/10.3390/s18113721
58. M. M. Sani, H. Norhazman, H. A. Omar, N. Zaini, S. A. Ghani, Support vector machine for
classification of stress subjects using EEG signals, in 2014 IEEE Conference on Systems, Process
and Control, (2014), 127–131. https://doi.org/10.1109/SPC.2014.7086243
59. R. Costin, C. Rotariu, A. Pasarica, Mental stress detection using heart rate variability and
morphologic variability of EeG signals, in 2012 International Conference and Exposition on
Electrical and Power Engineering, (2012), 591–596.
https://doi.org/10.1109/ICEPE.2012.6463870
60. H. G. Kim, E. J. Cheon, D. Bai, Y. H. Lee, B. Koo, Stress and heart rate variability: A meta-
analysis and review of the literature, Psychiatry Investig., 15 (2018), 235–245.
https://doi.org/10.30773/pi.2017.08.17
61. W. Boucsein, D. C. Fowles, S. Grimnes, G. Shakhar, W. T. Roth, M. E. Dawson, et al., Publication
recommendations for electrodermal measurements, Psychophysiology, 49 (2012), 1017–1034.
https://doi.org/10.1111/j.1469-8986.2012.01384.x
62. A. Tonacci, L. Billeci, E. Burrai, F. Sansone, R. Conte, Comparative evaluation of the autonomic
response to cognitive and sensory stimulations through wearable sensors, Sensors, 19 (2019),
4661. https://doi.org/10.3390/s19214661
63. K. T. Johnson, R. W. Picard, Advancing neuroscience through wearable devices, Neuron, 108
(2020), 8–12. https://doi.org/10.1016/j.neuron.2020.09.030
64. L. Hardesty, Researchers amplify variations in video, making the invisible visible. Available from
http://web.mit.edu/newsoffice/2012/amplifying-invisible-video-0622.html.

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.

 7919

65. L. Fraiwan, T. Basmaji, O. Hassanin, A mobile mental health monitoring system: A smart glove,
in 2018 14th International Conference on Signal-Image Technology & Internet-Based Systems
(SITIS), (2018), 235–240. https://doi.org/10.1109/SITIS.2018.00043
66. A. Engelniederhammer, G. Papastefanou, L. Xiang, Crowding density in urban environment and
its effects on emotional responding of pedestrians: Using wearable device technology with sensors
capturing proximity and psychophysiological emotion responses while walking in the street, J.
Hum. Behav, Social Environ., 29 (2019), 630–646.
https://doi.org/10.1080/10911359.2019.1579149
67. N. Narziev, H. Goh, K. Toshnazarov, S. Lee, K. Chung, Y. Noh, STDD: Short-term depression
detection with passive sensing, Sensors, 20 (2020), 1396. https://doi.org/10.3390/s20051396
68. Y. Chen, B. Gao, L. Jiang, K. Yin, J. Gu, W. Woo, Transfer learning for wearable long-term social
speech evaluations, IEEE Access, 6 (2018), 61305–61316.
https://doi.org/10.1109/ACCESS.2018.2876122
69. Z. Wu, H. R. Karimi, C. Dang, A deterministic annealing neural network algorithm for the
minimum concave cost transportation problem, IEEE Trans. Neural Networks Learn. Syst., 31
(2020), 4354–4366. https://doi.org/10.1109/TNNLS.2019.2955137
70. Z. Wu, H. R. Karimi, C. Dang, An approximation algorithm for graph partitioning via
deterministic annealing neural network, Neural Networks, 117 (2019), 191–200.
https://doi.org/10.1016/j.neunet.2019.05.010
71. S. Lawrence, C. L. Giles, A. C. Tsoi, A. D. Back, Face recognition: A convolutional neural-
network approach, IEEE Trans. Neural Networks, 8 (1997), 98–113.
https://doi.org/10.1109/72.554195
72. T. Durand, N. Thome, M. Cord, Weldon: Weakly supervised learning of deep convolutional neural
networks, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), (2016), 4743–4752. https://doi.org/10.1109/CVPR.2016.513
73. Y. Lei, X. Chen, M. Min, Y. Xie, A semi-supervised Laplacian extreme learning machine and
feature fusion with CNN for industrial superheat identification, Neurocomputing, 381 (2020),
186–195. https://doi.org/10.1016/j.neucom.2019.11.012
74. Z. Wu, B. Jiang, H. R. Karimi, A logarithmic descent direction algorithm for the quadratic
knapsack problem, Appl. Math. Comput., 369 (2020), 124854.
https://doi.org/10.1016/j.amc.2019.124854
75. Y. Lei, L. Cen, X. Chen, Y. Xie, A hybrid regularization semi-supervised extreme learning
machine method and its application, IEEE Access, 7 (2019), 30102–30111.
https://doi.org/10.1109/ACCESS.2019.2900267
76. Y. Lei, F. Liu, H. R. Karimi, X. Chen, Manifold semi-supervised learning for aluminum
electrolysis temperature identification based on regularized hierarchical extreme learning machine,
Proc. Inst. Mech. Eng., 236 (2022), 1109–1118. https://doi.org/10.1177/09596518221082857
77. D. Rumelhart, G. Hinton, R. Williams, Learning representations by back-propagating errors,
Nature, 323 (1986), 533–536. https://doi.org/10.1038/323533a0

©2022 the Author(s), licensee AIMS Press. This is an open access

article distributed under the terms of the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/4.0)

Mathematical Biosciences and Engineering Volume 19, Issue 8, 7899–7919.