Various data types were used in the studies (n=98) that used ML algorithms for stress and stress-related MDs. Studies used questionnaires (n=31, 32%); HRV (n=25, 26%); skin response (eg, skin temperature, skin conductance, etc; n=24, 24%); photoplethysmogram (PPG; n=21, 21%); electrocardiogram (ECG; n=19, 19%); HR (n=17, 17%); electroencephalogram (EEG; n=9, 9%); acceleration or body movement (n=8, 8%); text data (n=7, 7%); respiratory signals (n=7, 7%); electromyogram (EMG; n=3, 3%); eye tracking (n=3, 3%); speech signals (n=3, 3%); and others (n=4, 4%) including audio signals (n=2, 2%), blood pressure (BP; n=1, 1%), and hormones (n=1, 1%). Table 1 shows the distribution of the type of data used for stress detection using ML techniques.

Heart metrics are primarily used for stress detection and are typically gathered through 2 main methods: ECG and PPG. ECG is a noninvasive diagnostic test that records the heart’s electrical activity, while PPG is a noninvasive optical technique that detects changes in blood volume within the tissue’s microvascular bed. By using these methods, it is possible to measure various heart-related parameters, including HR, as well as time and frequency domain features of HRV and BP.

HRV (n=25, 26%) has been used to assess MH issues, such as stress, anxiety, and depression, due to its rich time and frequency domain features [53]. The blood volume pulse signal is another effective method for capturing HRV features, as it represents the heart’s beat-to-beat volume changes. From the blood volume pulse signal, time-domain measures such as the root mean square of successive RR interval differences (RMSSD), SD of neural network (NN) intervals, and SD of RR intervals can be derived. In addition, the frequency domain aspects of HRV, including total power (frequencies <0.4 Hz), low frequencies (LFs; ranging from 0.04 to 0.15 Hz), and high frequencies (HFs; between 0.15 and 0.4 Hz), reflect the autonomic nervous system’s dynamics during beat-to-beat measurements of the HR (Figure 2) [54,55]. These HRV measures, both in the time and frequency domains, provide a nuanced view of the physiological underpinnings associated with various MH conditions.

One of the most important indicators of stress is an abrupt increase in HR (n=17, 17%). Among the physiological signals, HR is among the top measures that explain stress in ML models, and it has been used in different studies with almost all ML algorithms [56-58].

BP (n=1, 1%) can be obtained by pulse transit time or by pressure cuffs [59]. Stressful conditions create an influx of hormones that increase HR and constrict blood vessels, leading to a temporary BP elevation [60]. In most cases, BP recovers to its prestress level after the stress response diminishes [61]. Schultebraucks et al [62] used systolic BP as one of the measures in predicting one’s level of susceptibility to PTSD.

EEG (n=9, 9%) detects brain electrical activity. Compared with other brain mapping techniques for stress detection, it is more practical due to several factors including affordability, noninvasiveness, nonintrusiveness, and, most importantly, its high temporal resolution [63]. The high temporal resolution of EEG makes it appropriate for real-time stress detection, as well as deep learning (DL) approaches, which require large data sets for training [63-67].

The most commonly used EEG features for the detection of stress are the power of different frequency bands (α, 8-13 Hz; β, 12.5-30.0 Hz; θ, 4.5-7.5 Hz; γ, 30-40 Hz), average and SD of a specific time window of EEG signal, and time-frequency features obtained by discrete wavelet transform algorithm [67-69]. It has also been shown that statistical features of EEG signals, such as kurtosis and entropy, are useful features in stress prediction using ML algorithms [66]. Moreover, power spectral density (PSD), correlation, divisional asymmetry, rational asymmetry, and power spectrum are other EEG features that have been used in different studies for stress detection [70].

Since EEG signals are collected from the scalp, they include excessive noise and have high uncertainty. Therefore, signal processing and feature selection or extraction are important steps while dealing with EEG data. Several well-developed methods are available for treating the EEG data. Among them, latent space derived from auto-encoders and signal reconstruction techniques such as artifact subspace reconstruction (ASR) are well-known methods that can be applied on EEG data to significantly reduce the artifacts [65]. These methods are also fast enough to make real-time detection feasible.

The amygdala and hippocampus are the parts of the brain that have the major responsibility for human reactions to stress [71]. Brain activity caused by stress in those regions would affect the prefrontal cortex. Studies collecting data from the prefrontal cortex have also verified that EEG data from this brain region can be used for stress detection [72]. EEG can be collected from the prefrontal cortex using off-the-shelf EEG recording products such as Muse and Neurosky Mindwave [66,69,70,72].

Eye-tracking features (n=3, 3%) can be indicators of stress. For example, to diagnose the level of stress, the changes in the striations of muscle material in the iris as a response to stress can be used as features for ML algorithms. In other words, pupil diameter, which is controlled by iris sphincter muscles, can be used as a feature [73]. Other eye-tracking features that have been used for stress detection are visual fixations, saccade movements, pupil size, microsaccades, and the number of eye blinks in a specific time window during a certain task [74-76].

A skin response (n=24, 24%) can be defined as a stimulus-regulated electrodermal response and is typically measured using electrodes placed on the fingertips or hands. Skin response is usually associated with an increase in sympathetic activity on inducing stress events [77]. The skin becomes a better conductor of electricity when it is stimulated either externally or internally by physiologically stimulating factors, including stressful conditions [78].

Mental stress can affect different respiratory cycle phases and breathing patterns (n=7, 7%) [79,80]. For example, it has been discovered that stress had no impact on overall breath duration (respiration rate) but that exhalation periods were longer and pause periods were shorter in the stress experiment than in the neutral condition [81].

On the basis of the findings of several studies, it can be concluded that respiratory signal is one of the top contributing factors in the explanation of stress in ML models. The most common time-domain respiratory signal features that are extracted for stress detection are root mean square, IQR, and mean square differences between adjacent elements of breathing rate and blood oxygenation levels. The most commonly used frequency domain features of the respiratory signal are the power of LFs (<2 Hz), the power of HFs (>2 Hz), and the ratio of the power of LFs over the power of HFs (LF and HF) [58,62,63,82-84].

EMG (n=3, 3%) detects the electrical activity of muscles at rest, during a modest contraction, and during a strong contraction [85]. Similar to acceleration data, several studies have shown that using EMG data can help increase the performance of ML models trained on ECG data. The action potential intrigued in the EMG during stress can reduce the variance for decision-making of classification models that use ECG [58,86,87].

It has been shown that stress can alter the levels of glucocorticoids, catecholamines, growth hormones, and prolactin in the bloodstream. Therefore, in ML models, levels of hormones such as cortisol, dehydroepiandrosterone sulfate, thyroid-stimulating hormone, free triiodothyronine, and free thyroxine can be used as predictors for the detection of stress-related disorders (n=1, 1%) [62].

Mental stress may cause a wide variety of behavioral and body movement symptoms such as shaking hands and feet, which can be measured by the acceleration data (n=8, 8%) [88]. Moreover, research has shown that people with a greater stress score had less variance in their activity level and body movements [89-91]. For example, in older adults, stressful life events can be related to a reduced rate of regular physical exercise [92]. Time and frequency features such as mean absolute deviation from the mean, the total power of acceleration, SD, the mean norm of acceleration, absolute integral, and peak frequency of each axis are the features of hand and body acceleration used for stress detection [31,57,93]. One practical characteristic of motion and acceleration data is that they can be used to identify sources of noise in other signals. For example, motion data can help distinguish stress from physical activity (eg, exercise) when other physiological measures, such as ECG, have uncertainty in prediction [94,95].

Using speech signals (n=3, 3%), it is feasible to diagnose and assess neurological disorders and MDs [96]. Moreover, studies have shown that like body acceleration and EMG, features of speech signals can make stress predictions of heart measurements more robust. The best explanatory parameters of speech signals are frequency domain parameters (eg, PSD, strongest frequency from fast Fourier transform) and time-frequency features such as Mel-frequency cepstral coefficient [56,97,98]. Since time-frequency measures are 2-dimensional measurements with a high number of samples, they make this signal suitable for use in convolutional NN (CNN) models of stress and depression detection [99].

For laboratory-based studies, audio signals (eg, beeping sounds; n=2, 2%) can be used to stimulate stress events in participants [100,101].

Social media content (n=7, 7%) is frequently subjected to reviews, opinions, and influence, as well as sentiment analysis. Natural language processing methods may be used to evaluate social networking posts and comments for mood and emotion to detect whether a user is stressed [102-108].

Different questionnaires (n=31, 32%) are used for the diagnosis of stress and MDs including anxiety and depression. The scores from different items on these questionnaires can be used as dependent and independent variables in ML studies. The questionnaires mentioned here were selected based on their prevalence in the literature as well as their relevance to the ML outcomes being predicted. For instance, some studies have successfully leveraged scores from multiple questionnaires, such as the Diagnostic and Statistical Manual of Mental Disorders, Depression Anxiety and Stress Scale, Edinburgh Perinatal/Postnatal Depression Scale, Center for Epidemiological Studies-Depression survey, Mean Opinion Score, Hamilton Depression Rating Scale, State-Trait Anxiety Inventory, Posttraumatic Stress Disorder Checklist for Diagnostic and Statistical Manual of Mental Disorders, Beck Depression Inventory, Beck Anxiety Inventory, Hospital Anxiety and Depression Scale, Goldberg’s Depression Scale, self-reports, and clinician reports [109-126].

In the detection of stress and its related MDs, the number of samples for the stress or MD class is usually significantly lower than the nonstress or non-MD class. This imbalance in the number of samples for each class leads to a bias in prediction (toward the majority class). To correct for data bias, it is possible to oversample the underrepresented group. In stress detection studies using ML models, the synthetic minority oversampling technique is one of the most common approaches to boost the minority class, which creates new samples by synthesizing those already available in the data (by combining their features; n=3, 3%) [31,110,127].

In ML models used for the prediction of stress with a multimodal approach, it has been shown that early fusion of multimodal data before feature extraction is more effective and achieves a better performance (n=1, 1%). This is because early modality fusion better catches the important characteristics that are coherent with each other. For example, a study showed that combining different measures including skin response, skin temperature, and body acceleration before feature extraction outperforms the approach that extracts the features for each measure separately and combines them afterward (Figure 3) [128].

In physiological signals for stress detection, usually the power of the signal changes during the moments of stress. PSD (n=13, 13%) explains the frequency-based power distribution of a time series and reveals the locations of strong and weak frequency variation. Welch’s method is one of the most common approaches for calculating PSD [65]. PSD is often used in studies that include frequency domain HRV features for stress detection such as total HF or LF power [82-84,101,129-136].

In the detection of MDs, such as depression and anxiety, using ML techniques, having the least error rate is significantly important so that the person can take further actions appropriately. In this matter, the data preprocessing step has an important role in minimizing the noise and bias toward the false prediction. Iliou et al [114] proposed ILIOU (n=1, 1%), a data mapping and transformation method that identifies useful information for the detection of MDs, especially for depression. This method outperforms common data preprocessing techniques such as principal component analysis (PCA), evolutionary search algorithm, and isomap for the detection of depression.

PCA (n=3, 3%) is a method for lowering the dimensionality of such data sets while maximizing interpretability and minimizing loss of information. It does this by generating new variables that are uncorrelated and progressively optimize variance [58,90,123].

Independent component analysis (ICA; n=4, 4%) is a computational and statistical method for uncovering hidden elements underlying random variables, observations, or signals. This method is mostly used for removing artifacts from stationary signal noises of the multichannel data. ICA optimizes higher-order statistics such as kurtosis, while PCA optimizes the covariance matrix of the data, which reflects second-order statistics. In stress detection using physiological signals that contain stationary noises (eg, eyeblink noise in EEG) it is recommended to remove noises using ICA [63-65,67].

ASR (n=1, 1%) is an adaptive approach for removing artifacts from signal recordings on the web or offline, mostly nonstationary signal noises. To identify artifacts based on their statistical qualities in the component subspace, a PCA on covariance matrices is repeatedly computed [137]. Since there are usually substantial nonstationary noises in the EEG data, in order to classify stress at multiple levels using EEG data, using ASR before classification is highly recommended [65].

Growth mixture modeling (n=1, 1%) is to discover numerous hidden subpopulations, describe longitudinal development within each hidden subpopulation, and investigate variation in hidden subpopulations’ rates of change. Latent growth mixture models are gaining popularity as a statistical tool for estimating individual development over time and for probing the presence of latent trajectories, in which people belong to the trajectories that are not directly observable [62,138,139].

It is common practice to transform data from 2 time series into vectors and then compute the Euclidean distance between the resulting points in vector space to determine the degree of similarity or dissimilarity between the series, regardless of whether they vary in time or velocity. Dynamic time warping (DTW) method (n=1, 1%) can be applied to find such similarities that may exist between people in terms of their mood series. As an example, one may compare time series to find whether they match for stress, depression, or anxiety. Moreover, it can be used to forecast the mental condition of persons with substantially comparable series patterns [130,140]. The difference between DTW and Euclidian matching is that unlike Euclidean matching, DTW considers the distance of each point in one sequence to every point in the other sequence to determine the similarity between them (Figure 4).

The Kalman filter (n=2, 2%) is a technique for making predictions about unknown variables (eg, missing data) based on observable data. Kalman filters include 2 iterative steps, predict and update, that are used to estimate states using linear dynamical systems in state-space format. Iterative cycles of predict and update are performed until convergence is achieved [143]. Kalman filter has been used to handle the missing data for stress detection in some studies [144,145].

Autoencoders (n=3, 3%) are a type of NNs that learn a representation of the data in lower dimensions than the original data (encoding) by regenerating the input from the encodings (decoding). For data with very high dimensionality, usually clustering is not optimized because of the noise present in the original data. Hence, it is an appropriate practice to use the encoded representation of the data, obtained by autoencoders, to have lower and more optimized dimensions for clustering [65,108,146].

In ML, a self-organizing map (n=3, 3%) produces a low-dimensional, typically 2-dimensional, representation of a high-dimensional data set while preserving its topology by creating clusters. Therefore, it is possible to visualize and analyze high-dimensional data more easily (Figure 5) [107,133,147].

Wrapper methods try to use a subset of features while training a model. Changes will be made to the feature subset on the basis of the performance of the prior model (Figure 6). Therefore, finding the best features using the wrapper method is a search problem. These methods often have high computing costs [148]. Some of the most common wrapper methods are naive search, sequential forward feature selection, sequential backward feature selection, and generalized sequential search [149]. Some studies used this approach as their feature selection technique [72,75].

In general, filter methods are used as a preprocessing step without regard to any ML algorithms. Statistic tests are used instead to select features based on their correlation with dependent variables (Figure 7). The filter feature selection methods used in the literature are mentioned in subsequent sections.

This test checks for independence between categorical features and the target variable. Features with high chi-square scores are selected, implying a strong association with the target variable, which may be valuable for the model (n=3, 3%) [56,135,150].

Pearson linear correlation coefficient (n=2, 2%) is a way to quantify how closely 2 sets of data are correlated linearly. It indicates how different measures are related to each other by a number between –1 and 1. Therefore, among highly correlated variables, some of them can be removed, as they do not add useful information to ML models [113,151].

The maximum relevance minimum redundancy technique (n=2, 2%) chooses characteristics having a high correlation to output (relevance) and a low correlation to one another (redundancy). The F statistic is used to determine the correlation between features and the output, while the Pearson correlation coefficient (for non–time series features) and DTW (for time-series features) may be used to calculate the correlation between features (Figure 8). The objective function, which is a function of relevance and redundancy, is then maximized by selecting features one at a time using a greedy search. Mutual information difference and mutual information quotient criteria are both frequently used objective functions that depict the difference or quotient between relevance and redundancy [152,153]. Using this feature selection method, Giannakakis et al [130] have ranked ECG measurements in the order of importance as mean HR, LF, NN50, SD of HR, pNN50, LF or HF, RMSSD, HF, and total power.

LR (n=26, 27%) is a supervised parametric ML technique in which multiple independent variables will be used to detect the occurrence of stress or normal conditions [72,117]. Some studies used the numerical independent variables (eg, HRV time-domain features: RMSSD, HR, and pNN50) [94,155] or categorical data (eg, answers to multiple choice questions) obtained from questionnaires [107,114,115].

NB algorithm (n=22, 22%) is a supervised, generally parametric, classification method that uses the Bayes Theorem as its foundation and has the naive assumption of predictor independence. In other words, the NB classifier assumes that the existence of a given independent variable to predict the dependent variable is independent of the presence of any other independent variable that predicts the dependent variable.

The DT (n=23, 23%) is a supervised nonparametric ML algorithm used in classification and regression applications. It comprises a root node, branches, internal nodes, and leaf nodes in a hierarchical, tree-like structure (Figure 9).

Boosting (n=22, 22%) is an ensemble learning for reducing training errors by combining a group of weak learners. When using the Boosting algorithm, models are fitted on random samples of data, and then models are trained repeatedly in a sequence. When each model starts being trained in that sequence, it attempts to make up for the flaws of the one that came before it. The most commonly used Boosting algorithms are adaptive boosting, gradient boosting, and XGBoost.

RF (n=37, 38%) is a supervised nonparametric ensemble learning algorithm that uses many DTs built during the training process. The RF algorithm is used for both classification and regression problems. When it comes to classification, the RF’s output is the class that most of the DTs choose. For regression purposes, an individual tree’s predicted mean or average is returned as the output. Using RFs, we can overcome the tendency of DTs to overfit their training data.

DA (n=6, 6%) is a supervised parametric classification algorithm that works with data including a dependent variable and independent variables and is mostly used to classify the observation into a certain group based on the independent variables in the data. Linear DA and quadratic DA are the 2 forms of DA.

KNN (n=22, 22%) is a nonparametric supervised ML algorithm that is used for both classification and regression purposes. In classification, the algorithm determines the label of a new sample not available in the training data by assigning the label of the majority of k-nearest training data points to that new sample (Figure 10). In regression, the output for each sample is the average of the values of KNNs to that sample (not including the sample itself). In this literature, KNN has only been used for classification.

SVM (n=48, 49%) is a parametric supervised ML algorithm used for both classification and regression problems. It can solve both linear and nonlinear problems using nonlinear kernels. For classification, the SVM algorithm finds a line (or a hyperplane for nonlinear kernels) between each pair of classes of the training data in a way that the margin distance of that line or hyperplane to the closest point of each of those 2 classes is maximized (Figure 11). This is repeated for all pairs of classes in the data set. Then, the obtained lines are used as boundaries for the classes. In regression, the SVM tries to find the line or hyperplane that within a very small margin of has the maximum number of data points. That line or hyperplane was used for regression.

K-means clustering (n=4, 4%) is an unsupervised ML algorithm that aims to arrange objects into groups based on their similarity. To find those similarities, it calculates the distance of data points into K random cluster centroids and assigns each data point to its closest centroid. The location of each centroid is then updated by the average value of all data points associated with that centroid. This process is repeated until there is no change in the location of the centroids. In ML models for stress detection, K-means clustering has been used in the literature for the personalization of the ML models [58,67], and for labeling the data set [146,156].

DL methods are a subset of ML methods, and NNs are at the heart of the DL algorithms. The NN (n=38, 39%) is a method for implementing ML that uses interconnected nodes or neurons arranged in a layered structure resembling the human brain. The different types of NNs have been explained in subsequent sections.

It is possible to think of a single perceptron (or neuron) as an abstract LR. In each layer of artificial neural networks (ANNs), a group of multiple perceptron or artificial neurons is used. Figure 12 shows an ANN with 1 layer and its working mechanism.

CNNs are a form of NN that is especially adept at handling data structures with a grid-like layout, such as images or objects. Classification and computer vision applications are common uses for CNNs (Figure 13).

A recurrent neural network (RNN) is a subset of ANNs designed specifically for use with time-series data and other sequence-based data. Long short-term memory (LSTM) networks are the most common type of RNNs. In RNNs, the attention mechanism is a method that simulates cognitive attention in NNs. The purpose of the impact is to encourage the network to give greater attention to the small but significant portions of the input data by enhancing some and reducing others. Since stress may alter a small portion of physiological data (eg, ECG), attention mechanisms can be used to detect stress using RNNs when large data sets are available [141].

Cong et al [102] introduced X-A-BiLSTM, which is a DL model that includes XGBoost (to filter data and handle imbalanced data) and attention Bi-LSTM (LSTM with forward and backward memory and attention mechanism) NN used for stress classification using text data.

The total number of studies in this category include (n=19, 19%) of studies.

The classification is decided based on weighted voting, which is determined by using a voting ensemble approach. The voting classifier allows for voting in which the final class labels are determined either by the class chosen most frequently by the classification models or by the average of the output probabilities from each classification model. In the literature, this method has been used for PTSD detection [127], stress, and stress-related MDs [84,95,118,156,157].

FCM is a clustering approach that assigns every data point to all the clusters with a certain probability instead of assigning each point to only 1 cluster. For instance, a data point that is near the cluster’s center will have a high degree of membership, while a data point that is distant from the cluster’s center will have a low degree of membership [158]. Since depression and anxiety are not discrete measures, some studies have used FCM as an alternative to other clustering techniques for the detection of these MDs [114,116].

The pivotal insights from this review underscore the potential of ML to redefine the approach to MH care, particularly in the diagnosis and management of stress-related conditions and MDs. As we have discerned, there is an expansive field ripe for further exploration, with research gaps suggesting a number of promising directions. Guided by these insights, we can now chart a course for future research that not only expands the boundaries of our scientific understanding but also translates into tangible improvements in clinical practice.

The scarcity of real-time studies in naturalistic settings has highlighted the importance of developing ML models that accurately reflect and respond to the complexities of real life. Future research must prioritize the creation of algorithms capable of operating amidst the unpredictability of daily life, providing immediate insights and adaptable interventions. These models hold the potential to transform practice by offering tools that can preemptively identify stress and MD symptoms, enabling clinicians to intervene before conditions worsen.

Our review illuminates the untapped potential of time-series data in capturing the evolution of stress and MDs. DL techniques, specifically designed to interpret complex, sequential data, could lead to breakthroughs in how we understand and predict MH trajectories. For practice, this means more sophisticated diagnostic tools that can provide a nuanced picture of a patient’s MH over time, enabling personalized treatment plans that are responsive to the patient’s changing condition.

The need for individualized care for MH cannot be overstated. The heterogeneity of stress responses and MD symptoms calls for personalized ML models tailored to individual physiological and behavioral patterns. Future research should focus on leveraging multitask learning to refine algorithms that adapt to individual baselines, enhancing the personalization of care. For clinicians, this means access to tools that can more accurately reflect and respond to the unique needs of each patient, reducing the risk of misdiagnosis and improving treatment efficacy.

Predictive analytics can be instrumental in identifying key factors that contribute to misdiagnosis and delayed help seeking. Future studies should look to build on this knowledge to inform the creation of interventions that encourage timely and accurate diagnosis. In practice, this could lead to the development of targeted screening tools that assist clinicians in recognizing at-risk individuals more effectively. The integration of clinical expertise with ML innovation is crucial for the development of tools that are both advanced and clinically relevant. Collaboration between health care professionals, patients, and artificial intelligence developers will be essential in creating user-centered tools that address real-world needs. This collaborative approach will likely result in the development of artificial intelligence applications that are more intuitive and effective in clinical settings.