We adopt a scoping review approach to systematically identify gaps in the existing literature. Although stress detection has long been studied in laboratory settings, it remains challenging to translate these findings into real-life contexts. As such, our primary goal is to map current research on stress detection, identify existing gaps, and provide an overview of ML frameworks inspired by the concept of model cards. Specifically, this review examines how ML frameworks are used to detect stress in naturalistic settings using wearable sensors. The review is guided by the methodological framework proposed by Arksey and O’Malley [24] and adheres to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines (Checklist 1) [25].

The primary search for this scoping review was conducted in April 2024 across 7 databases: PubMed, MEDLINE, ScienceDirect, IEEE Xplore, ACM Digital Library, Web of Science, and Embase. These databases were chosen for their extensive coverage of recent physiology research on stress detection. Our aim was to identify articles examining stress detection in naturalistic settings using biosignals and wearables. We restricted our search to English-language, peer-reviewed studies published between January 2017 and April 2024 that were available in full text. Initially, we tested various search strings—for instance, using “stress detection” alone yielded too many results, whereas “stress detection” AND (“naturalistic setting” OR “field study” OR “real life”) yielded too few. Ultimately, using more generic terms such as “stress detection,” “wearables,” and “physiology,” combined with AND operators, produced a sufficiently broad yet focused list of results. Research articles involving either primary or secondary data analyses (or both) were included. Primary data analysis refers to studies where data were collected by the authors, whereas secondary data analysis refers to studies using data collected by other researchers, which were either openly available or accessible upon request.

Studies were deemed eligible for inclusion if they used data collected in naturalistic (real-world or in-the-wild) settings. Here, naturalistic refers to situations in which stress is not experimentally induced or linked to known stress-inducing events (eg, taking a test), but instead reflects stress as it occurs in daily life. This criterion applied even if some studies use multiple datasets, including those from controlled or laboratory environments, provided that at least one dataset was collected in the wild. This decision was motivated by the fact that stress labeling mechanisms in laboratory studies differ fundamentally from those used in real-life or naturalistic contexts, where stress must be explicitly self-reported or otherwise captured from participants during their daily lives, which pose unique challenges. Given the study’s focus on acute stress, only research involving healthy populations in everyday contexts was considered. Consequently, the use of wearable devices was a prerequisite, as they facilitate continuous monitoring in daily life. Recognizing the surge in wearable technology adoption in recent years, the review encompassed studies published between 2017 and 2024. The specific eligibility criteria pertinent to this scoping review are detailed in Textbox 1.

We began by applying basic eligibility filters (language and publication year) within the scientific databases themselves. After eliminating duplicate entries, the first author compiled all the retrieved records into a spreadsheet and conducted the initial title and abstract screening using preagreed eligibility criteria. The third author reviewed and verified the screened records and decisions regarding the inclusion and exclusion in the spreadsheet against the same criteria. During abstract screening, we identified publicly available, on-request datasets containing naturalistic wearable data for stress detection and reviewed their citations, adding any studies that used these datasets to our review. Any ambiguities or discrepancies were resolved through discussion until consensus was reached. Since formal full parallel double-screening was not performed, Cohen κ or percentage agreement was not calculated. Finally, the remaining full-text articles were reassessed for eligibility using the specified inclusion and exclusion criteria.

Data extraction involved classifying studies into predefined categories to facilitate evidence synthesis, with a particular focus on the design of ML frameworks inspired by model cards. Extracted data elements included availability of datasets such as indicating whether datasets were open source, available upon request, or collected within the study design, demographic information such as participants' professions, total number of participants, and geographic location, details of wearable devices and biosignals used, duration of data collection, primary research focus such as problem addressed which are either originating from domain-centric or data-centric approaches, and ML models used. Additionally, we examined the formulation of problem statements, including the selection of input and output windows for nowcasting or forecasting tasks. We also assessed validation practices, with particular attention to measures implemented to prevent data leakage, such as appropriate data partitioning and cross-validation techniques.

As shown in the flowchart in Figure 1, a total of 353 articles were identified through database searches: 100 from PubMed, 16 from IEEE, 167 from ScienceDirect, 14 from MEDLINE, 4 from Web of Science, 28 from ACM Digital Library, and 24 from Embase. During the title and abstract screening, 258 articles were excluded for not meeting the eligibility criteria.

During the abstract screening phase, we identified additional datasets that were collected using wearables to model wellness-related constructs in naturalistic settings, for example, TILES-2018 [26], Lifesnaps [27], Tesserae [28], and SWEET [29]. In total, we find 227 secondary analysis papers citing those datasets, with 45 of them using TILES-2018, 106 using SWEET, 70 using Tesserae, and 6 using Lifesnaps. After further title and abstract screening, 163 studies were excluded. This left a total of 146 studies considered for full-text review, with a mixture of primary and secondary analyses. Of these, 71 studies were excluded from the primary data analysis category, and 41 studies were excluded from the secondary data analysis category based on the eligibility criteria. In the end, 34 articles were selected for inclusion in this scoping review. For a detailed description of the search and selection strategy using PubMed, please refer to Multimedia Appendix 1.

EMA was most frequently administered daily [48,51,54], capturing a range of psychological constructs. In fewer instances, more frequent assessments were conducted (eg, every 30 minutes or every 2 hours throughout the day) [30-32]. In all studies, EMA prompts were triggered by predefined schedules, with most being delivered via mobile apps, often through custom-developed apps [21,48,51]. A smaller number of studies used SMS messages to deliver EMA surveys [38,39,50].

Since the data were collected in real-world contexts, participants were not expected to respond promptly to EMA probes as they may have been busy with other tasks. To control the delay in response and ensure the accurate timing of the label, studies may put an upper bound on the delay, that is, the probe will expire after a certain period of time [32] to ensure the relevance of labels.

The most commonly collected psychology constructs in the studies were stress, which occurred in 26 [19-21,30-39,41-48,50,51,55,57-59] out of 34 (76%) studies, anxiety in 9 [22,39,43,50,51,53,55,56,59] out of 34 (26%) studies, positive or negative affect in 9 [22,39,41,43,49-51,53,54] out of 34 (26%) studies. Various instruments were used, including the Short State-Trait Anxiety Inventory (S-STAI) [47,59], which assessed both stress and anxiety, and the Perceived Stress Scale (PSS) [21,48]. Anxiety was also measured using the State-Trait Anxiety Inventory (STAI) [42,56]. The Positive and Negative Affect Schedule (PANAS) short scale [49,53] was used to assess both positive and negative effects.

Additionally, different scales were used, such as the Likert scale [31,36,37], Numeric Rating Scale [33], and Visual Analog Scale [35], which is a gliding scale ranging from “not at all” to “extremely.” In a few studies, discrete ordered scales were used to categorize stress levels as low, medium, or high [30,58].

There were various preprocessing decisions researchers have made with regard to the stress labels in the reviewed studies. In most studies, the output targets in the ML model maintained the same granularity as the ground truth stress label. However, some studies applied downsampling techniques. For instance, when ground truth labels were collected every 30 minutes, they were aggregated into daily values [20,31] in modeling.

In classification studies, these outputs collected on a numeric scale were often categorized into bins or classes, such as low, medium, or high levels of stress or anxiety [19]. Most studies used binarized output targets for classification, while only a few applied multiclass classification. Due to the ordinal nature of the data, multiple categories were combined into fewer groups. For example, on a 5-point scale, ratings of 0‐1 were classified as stress, while ratings of 2‐4 were classified as no stress [19,32]. Another example involves merging the last 3 categories of a 5-point scale into a single “high stress” group, with the first 2 categories remaining as “low” and “medium stress” [44]. For nondiscrete values, such as on a 100-point scale, values below 50 were marked as low stress, and values above 50 were marked as high stress [35].

Two primary methods were used to define classification label thresholds: a personalized approach and a global approach. The personalized approach involved setting thresholds specific to individual participants, whereas the global approach applied the same thresholds across all participants. For personalized thresholds, various statistical methods were used, including majority voting, median calculation, and z-score analysis. The z-score method was the most commonly used [21,46,48,56], where z-scores were calculated for each individual to account for subjective variability. Data were then categorized into 2 classes; for example, negative classes included those with z-scores below zero, and vice versa. In the global approach, values were discretized using predefined scales or thresholds derived from the entire participant sample. For instance, some studies applied fixed global thresholds to classify stress and anxiety levels across all participants [42].

As part of the modeling framework, we extracted the input and output windows reported in the reviewed studies. The input window refers to the time frame from which input features are extracted, typically derived from physiological and behavioral signals. On the other hand, the output window defines the period during which the model predicts target outcomes corresponding to the ground truth labels, as discussed in the “Ground Truth Stress Labels” section.

We use the terms nowcasting and forecasting to describe the relationship between input and output windows. Nowcasting (Figure 2A-C) refers to models that predict outcomes, such as stress, for the present or immediate future, resulting in a very short output window with no gap between input and output windows. In contrast, forecasting (Figure 2D) predicts stress over future time periods, with short-term forecasting focusing on the immediate next period and long-term forecasting predicting stress at a time point extended further into the future.

Several studies have used an exact overlap between input and output windows as shown in Figure 2A, often with a window size of 24 hours [31,51]. A smaller number of studies have used shorter input-output windows, such as every 20 minutes [32], hourly [44], or every 3 hours [35]. This exact overlap is a defining characteristic of nowcasting studies, as illustrated in Figure 1.

Additionally, many studies included partial overlap, where the input window is either longer or shorter than the output window, as shown in Figure 2B and C. For example, an input window might include data from the previous day, the current day, and the next day, while the output window focuses only on the current day [51,52] or both the previous and current days [54]. In other examples, input window is shorter than the output window; for example, the input window covers 1 minute and the output window spans 15 minutes [30], or where the input window is 2 hours paired with a 1-day output window [56], or an input window that selects an optimal range between 10 and 27.5 minutes while the output window remains fixed at 20 minutes [19].

In the forecasting studies identified, the input window is chosen to be immediately adjacent and trail the output window (Figure 2D). For example, the input window may consist of data from the previous day, with the output window set for the current day [40], in which case the gap is zero. In other instances, the input window might consist of data from the last 35, 30, or 25 days, while the output window could be set for 19, 17, or 12 days [57] into the future, starting from the prediction time, in which case the gap is larger than zero.

It has been observed from the selected studies that the majority of researchers have used a nowcasting relationship between the input and output windows, as illustrated in Figure 3.

The reviewed studies differed in how their training, test, and validation sets were split. Since the wearable dataset is longitudinally collected, it is typical to see multiple input-output pairs extracted from the same participants. Data leakage can occur where data points in the training set and test set may originate from the same participant. This made participant-level splitting relevant to the assessment of potential data leakage and possible inflation of reported model performance.

In our review, we observed that 13 [30,32,35,36,41,42,45,47,54-56,58,59] out of 28 applicable (46%) studies [19-21,30-32,35-38,40-48,50-52,54-59] did not explicitly address potential data leakage issues as inferred from the description of the split of training, test, or validation set, while the rest addressed data leakage issues by using strategies such as leave-one-subject-out cross-validation.

This section provides insights into the types of ML models used. The studies included in this review primarily focused on supervised learning paradigms, addressing both classification and regression tasks. A diverse range of models was used, spanning traditional ML frameworks to end-to-end deep learning architectures. For feature extraction, most studies relied on statistical summaries [30,31,36], while a smaller subset applied Fourier transformations, particularly for analyzing heart rate variability signals [19,20,42].

Traditional ML algorithms, valued for their simplicity and interpretability, such as support vector machines (SVMs), Naive Bayes, and k-nearest neighbors (KNN), were frequently used. Additionally, various ensemble methods, including bagging and boosting techniques, were widely applied [19,44,45,47,51]. Among these, random forest (RF) was the most commonly used in studies using traditional ML algorithms [19,36,38,44,45,49,50,52,54,58,59].

Table 4 summarizes the most frequently used traditional ML models across the studies.

In addition to traditional ML, the use of deep learning models has seen a significant rise. Models explored include recurrent neural networks (RNNs) and long short-term memory (LSTM) networks, convolutional neural networks (CNNs), and LLMs. The choice of models is closely tied to the specific objectives and implications of each study, as discussed below.

The reviewed studies aimed to address specific challenges, each accompanied by proposed solutions. These challenges were either intrinsic to the data or inherent to the domain. Some common challenges and their corresponding solutions are outlined below

Personalization was one of the challenges addressed in the reviewed studies, because of intrinsic between-person variability, as individuals may respond differently to the same stressors. In many of the selected studies, personalization was not considered, thus following a one-size-fits-all approach in modeling. However, a notable number of studies addressed personalization as part of the preprocessing phase. Two key personalization approaches were identified: user-based and group-based. While the user-based approach involves creating a unique model for each participant, the group-based approach groups participants based on selected attributes, such as demographics or gender.

Various clustering algorithms were used in the studies to create cohorts or groups, including k-means, spectral clustering, agglomerative clustering, mean shift, affinity propagation, and Gaussian mixture models (GMM) [51]. Some other studies also used additional techniques to enhance cluster quality, such as iterative clustering [41], which progressively refined participant groupings by repeating the clustering process multiple times. Another method listed in the reviewed studies was dimensionality reduction prior to clustering, which was achieved through principal component analysis (PCA) [43]. PCA simplified the feature space while preserving variance by projecting the data into a lower-dimensional space, whereas self-organizing maps (SOMs) [31] facilitated the organization of complex patterns within the data.

The clusters created during the preprocessing step were further trained using traditional ML techniques (eg, RF, decision trees, and linear regression) [45] as well as deep learning approaches, including metric learning with Siamese neural networks (SNNs) [43]. SNNs transform input data to ensure that similar data points are projected close together, while dissimilar points are spaced apart in the learned feature space. Additionally, feedforward neural networks (FFNs) [41] were used to fine-tune the generalized model by adjusting its weights through iterative training on group-specific data.

Clustering offers several benefits. For example, it may induce more cohesive groups as grouping users with similar characteristics helps the model detect common stress patterns within each group [45]. In another study, the user-based collaborative filtering method is used by leveraging health information from other users. This type of model can fill gaps within groups and reduce the reliance on historical data from individual users [51].

Data collected from wearable sensors inherently possess a time-dependent nature due to its intrinsic properties. Consequently, studies have explored methods to explicitly address these temporal dependencies to improve model accuracy, with the hope that those methods may reveal complex relationships that conventional ML models may overlook. Various models specifically designed to capture these temporal dependencies are listed below, along with a brief summary of their usage in the selected studies.

The feasibility of using all training modalities is often limited in real-world scenarios due to the restricted availability of sensors in wearables. To address this, one study focused on reducing the number of modalities during testing while incorporating more during training [46]. This was achieved using a technique called knowledge distillation, implemented as the “More to Less” framework. Knowledge distillation involves transferring knowledge from a stronger network to a weaker one by using classifier networks from different modalities. During training, the model used all available modalities, with each having its own classifier network. These classifiers exchanged knowledge through adaptive mechanisms defined in the “More to Less” framework, allowing each classifier to learn not only from its own modality but also from those performing better. As a result, testing was conducted with fewer modalities, as the network had already developed robust representations during training. This approach enhanced generalization by leveraging information from multiple modalities, even when fewer input modalities were available during testing.

Bias in relation to demographic groupings was addressed in one reviewed study [56]. In that study, an LSTM model was used by incorporating multitask learning, which included both the protected attribute and the target variable. This approach aimed to mitigate bias related to the protected attribute. To achieve this, different loss weights were assigned. For example, if the target variable (eg, anxiety) is given a weight of 4.5 and the protected attribute (eg, gender) a weight of 0.5, the problem is treated as a multitask learning scenario. Each task has a separate loss function, ensuring that the nonprotected label, such as anxiety in this case, is prioritized.

Contextual information is widely recognized as crucial in studies aimed at identifying the causes or triggers of stress, as learning context helps improve model accuracy. Various types of contextual data considered in these studies are outlined below:

Given the complexity of data collection in the naturalistic setting, there is a limited supply of high-quality datasets, along with challenges in capturing ground truth. To address those challenges, a study [54] explored zero-shot learning within a meta-learning framework. This framework incorporated various base learners, including RF, AdaBoost (AB), gradient boosting (GB), KNN, and LR. The approach enabled the model to learn from diverse tasks, allowing it to quickly adapt to new, unseen tasks with minimal additional training. By training on different tasks and populations, the model developed effective initialization weights, improving its ability to make accurate predictions on unseen data.

Recently, with advancements in AI, emerging technologies such as LLMs have been developed. These models are pretrained on vast amounts of text and can be applied to a wide range of tasks. One study in the review used LLMs along with wearable sensor data through structured prompts that incorporated user context, health knowledge, and temporal information [59]. The study focused on three approaches:

Overall, few-shot prompting outperformed zero-shot prompting, while fine-tuning proved to be the most effective approach.

In Table 3, we presented a detailed analysis of model performance and extracted key evaluation metrics, such as F₁-score, accuracy, precision, recall, and AUC for classification tasks, and MSE for regression tasks. We reported the best-performing model in each study. Where applicable, we also documented the corresponding model configurations and comparison baselines. It should be noted that data leakage considerations are only relevant for models designed to predict momentary stress labels. Six studies [22,33,34,39,49,53] fall outside this scope, including (1) statistical models (often linear or linear mixed-effects models) that fit the entire dataset to examine coefficient significance or overall goodness-of-fit; and (2) models predicting stable psychological constructs derived from one-time baseline surveys, where the unit of analysis is the individual rather than the event or time point. We identified a total of 13 studies [30,32,35,36,41,42,45,47,54-56,58,59] that had potential data leakage issues out of 28 applicable studies [19-21,30-32,35-38,40-48,50-52,54-59].

There are a few caveats in interpreting performance measures:

There are several major findings:

Ethical Considerations

Stress detection applications and associated wearable devices often involve the collection of fine-grained longitudinal data, as well as psychological profile information, either in raw form or inferred through computational models. Such data fall within the realm of sensitive private information and, if inadvertently disclosed, may pose significant risks to individuals. For example, employers could potentially misuse stress or negative affect profiles in human resource–related decision-making processes. In addition, weak data security policies may increase the likelihood of security breaches, which could enable the reidentification of individuals and lead to the leakage of personally identifiable information. Accordingly, research studies on stress detection, as well as the downstream applications built upon these systems, must adopt deliberate and robust measures to safeguard the entire data collection, storage, and processing pipeline. These measures are essential for providing a high level of assurance and protection to individuals, whether they participate as research subjects or engage with such systems as end users of downstream applications. Some emerging approaches may also be explored, such as federated learning, as suggested in one of the reviewed studies [59]. In this approach, raw data remain on the local device where they are recorded, thereby reducing the privacy risks associated with centralized cloud-based data storage and processing.

Equally important is the need for transparent and meaningful informed consent procedures. Data owners should be clearly informed about what types of data are being collected, how the data will be processed and analyzed, who will have access to the data, how long the data will be retained, and whether the data may be reused for secondary research purposes or commercial applications. Such transparency is critical for establishing trust and supporting individuals’ autonomy in making informed decisions regarding participation. In addition, participants should be made aware of potential risks, including the possibility of reidentification or unintended disclosure of sensitive psychological or behavioral information. As noted, among the 14 datasets used across the reviewed studies, some based on primary data collection and others relying on secondary datasets, only 11 explicitly acknowledged obtaining proper institutional review board (IRB) approval, whereas the remaining 3 studies did not report such approval [19,20,32]. This observation highlights the continuing need for stronger ethical oversight and clearer reporting practices in stress detection research involving wearable and behavioral data.

In this section, we propose a model card framework tailored for wearable-based stress modeling, adapted from the general model card template introduced by prior work [23] in the ML community. The specific focus is on models that enable the momentary estimation of current stress or to forecast stress in the future. The original template includes sections such as model details, intended use, factors, metrics, evaluation data, training data, quantitative analysis, ethical considerations, and caveats and recommendations. Given the unique characteristics of wearable-based stress modeling, particularly the reliance on longitudinal, in-the-wild biosignal data collection and self-reported ground truth, we reorganize and customize the model card into two core components: (1) dataset and (2) modeling decisions. This structure foregrounds the data provenance and methodological choices that most strongly influence model validity, transparency, and reproducibility. A detailed model card for wearables, including its components, dimensions, and reporting requirements, is presented in Table 5 below.

Stress detection has become increasingly important due to the rising prevalence of stress in individuals’ daily lives. Advances in wearable technologies have opened new opportunities for detecting stress in real-world settings. This scoping review aims to examine the existing literature on the use of ML applied to data collected from wearables in naturalistic environments for stress detection. In this section, we summarize the principal findings, the identified research gaps, and potential avenues for further investigation, while acknowledging the limitations of this review.

This review identified a lack of standardization in model reporting, which is one of the biggest barriers to advancing the field beyond existing research, and the generalizability of stress detection in naturalistic settings. This is due to substantial heterogeneity in how problem specifications are defined, with no standard format for model reporting, which limits benchmarking across current studies. The reviewed studies differed considerably in the wearables used, ground-truth labeling strategies, input-output window definitions, and ML approaches. We also identified methodological gaps, including potential data leakage, inconsistent reporting of performance metrics, limited handling of class imbalance, and relatively little explicit attention to motion artifacts. In addition, we noted that the reviewed studies addressed a range of challenges, though not uniformly, including personalization, temporal modeling, bias mitigation, modality-related concerns, challenges specific to in-the-wild settings, and the use of LLMs. These findings also motivated the proposed model card specification to improve comparability, transparency, and reproducibility across studies in the future.

In our analysis of model frameworks, characteristics, and performance summarized in Tables 2 and 3, we observed that 13 [30,32,35,36,41,42,45,47,54-56,58,59] out of 28 (46%) applicable studies [19-21,30-32,35-38,40-48,50-52,54-59], specifically, those that developed ML models to predict momentary stress had potential data leakage issues. These issues primarily arise from ignoring the grouped structure of the data, where multiple observations or episodes are contributed by the same participant. When training, validation, and test splits are performed at the observation level rather than the participant level, data points from the same individual may appear in both the training and evaluation sets. This overlap can result in unintended information leakage from training to test data, thereby artificially inflating model performance. A commonly recommended practice to mitigate this issue is to perform data splitting at the subject level, ensuring that all observations from a given participant are assigned to a single split. Given that a substantial proportion of the reviewed studies exhibit this potential methodological flaw, the performance metrics reported in Table 3 should be interpreted with caution, as they may overestimate true generalization performance.

Another issue identified in our analysis of model performance is that a large proportion of studies do not explicitly report baseline model performance. This omission is particularly problematic in the presence of class imbalance, where commonly reported metrics such as accuracy may be misleading. For example, a naïve classifier that always predicts the majority class may achieve deceptively high accuracy, despite providing little meaningful predictive value. Moreover, performance metrics that are more robust to class imbalance, such as the AUC, are rarely reported. While some studies attempt to address this issue by using alternative measures, including balanced accuracy (BACC) or Matthews correlation coefficient (MCC), such practices remain the exception rather than the norm. In addition, we observe that a substantial proportion of studies do not report uncertainty quantification for performance metrics, such as CIs or SDs. The absence of these measures complicates performance comparison across models or approaches, as observed differences may simply reflect random variation rather than meaningful performance gains.

To address these methodological gaps, we explicitly incorporate these considerations into our proposed model card specification, with the goal of encouraging future research to adopt more rigorous experimental designs, performance measures, and reporting practices.

The complexity and richness of data collected from wearables and other devices in real-world settings create ample opportunities to explore a wide range of ML frameworks. While this flexibility allows researchers to pose a variety of research questions and explore various kinds of models, it also poses significant challenges when comparing the utility and performance of different models in tasks such as stress detection and estimation. As demonstrated in our analysis, the lack of standardization in reporting often makes it difficult to extract key specific details about ML modeling decisions, thus limiting the ability to compare model effectiveness across studies. This lack of standard and benchmarking is likely to slow progress in developing robust ML models for stress detection.

Future research could focus on the establishment of benchmark datasets and standardized tasks to enable meaningful comparisons. Additionally, the research community would benefit from adopting consistent reporting standards that ensure critical components of ML frameworks are clearly documented. For example, approaches such as model cards [23] promote transparency and reproducibility. There should also be a stronger emphasis on model quality control, including the adoption of best practices in experimental design to mitigate issues such as data leakage, which may contribute to the inflation of reported model performance. In this paper, we proposed a model card solution specifically tailored to modeling stress in the wild using wearable data, based on our analysis and understanding of the critical issues and related best practices related to modeling decisions, performance analysis, and reporting.

A distinguishing feature of real-life stress studies is the ability to examine stress over extended periods and explore its interactions with contextual factors. Despite these opportunities, we observed that most studies adopt a short-term approach to stress detection, where the input and output windows are closely aligned in time. Only a few recent studies have begun to explore forecasting methods that account for long-term temporal trends in psychological and physiological states.

Among the 4 forecasting-oriented papers, 2 adopt a conventional ML pipeline focused on next-day stress prediction. Paraschou et al [47] report an accuracy of 92.3% from a GB ML model, although the evaluation may be subject to data leakage concerns. Another study [50] systematically explores a wide range of input and output window configurations and finds that LSTM-based models generally achieve the strongest performance. In particular, the highest predictive accuracy is obtained when using data from the preceding 15 days (ie, input window=15 days) to predict next-day stress (ie, output window=1 day), although XGBoost (eXtreme Gradient Boosting) models show competitive performance under several configurations. In contrast, the remaining 2 papers adopt dynamic systems modeling approaches [40,57]. Hadjiantonis et al [40] fit a linear dynamical system model that characterizes next-day changes in emotional arousal as a linear function of the current day’s perceived stress and emotional arousal, with model parameters estimated via linear regression. Li et al [57], however, model day-to-day perceived stress as a nonlinear dynamical system. The study follows a 2-step approach: first, chaos theory is used to identify the most reliable prediction horizon, which shows that an input window of 19 days maximizes predictability for forecasting stress over the subsequent 35 days. Using phase-space embedding and deep neural networks, the proposed method achieves accuracies of 74.4% for binary stress classification and 69.23% for 3-level stress classification under a person-wise 80/20 split.

We encourage future research to broaden the scope of stress modeling beyond momentary detection, incorporating longitudinal perspectives that capture the ebb and flow of stress over time. Such studies could also investigate the interplay between stress and mediating factors such as sleep, as well as coping strategies such as meditation, to provide a more comprehensive understanding of stress dynamics in daily life. Our review suggests that methodological innovations adopting multivariate and longitudinal perspectives are still emerging. Traditional ML pipelines provide a reasonable starting point by framing stress forecasting as a mapping between multimodal signal streams within predefined input windows and corresponding output windows. However, much of the existing work relies on black-box models that offer limited insight into the underlying mechanisms of stress-related dynamics. A complementary line of research draws on dynamical systems–inspired approaches, including both linear and nonlinear models. Within these frameworks, prediction targets often extend beyond a single future time point to a prediction horizon spanning multiple days or weeks. This shift enables more meaningful and theoretically grounded questions to be addressed, such as the evolution trajectory and long-term trends of stress dynamics over time. Chaos theory–based approaches, in particular, provide a useful lens for understanding the limits of predictability in forecasting future stress trajectories. Li et al [57] represent a promising first step in this direction, and future work could further advance this line of inquiry by incorporating richer, multivariate biosignals from wearable devices in both the input and output spaces.

For both ML-based and dynamical systems–based models, a key challenge in fitting highly complex models lies in data sparsity and limited transparency, as these approaches often require large amounts of high-quality longitudinal data while offering limited interpretability. One promising direction for addressing these challenges is the systematic integration of domain knowledge, such as insights from psychological theory, through principled frameworks such as neurosymbolic systems [60,61], which combine data-driven learning with theory-informed constraints and representations. Advancing such approaches will likely require close multidisciplinary collaboration between computing researchers and domain experts in fields such as psychology and health care.

Data collection in real-life settings, where participants are moving, poses significant challenges to the validity of physiological signals due to motion artifacts. Additionally, it remains unclear to what extent models trained on laboratory-controlled datasets generalize to real-world scenarios. Our analysis found that only a few studies explicitly acknowledge and address these challenges. To address motion artifacts, a study [42] proposed subband HRV features in addition to traditional HRV benchmark features (time-domain and frequency-domain features). The proposed features were obtained by splitting the HRV tachogram into low-frequency (LF) and high-frequency (HF) bands and analyzing each band separately. From these bands, the authors computed nonlinear features, such as transfer entropy between LF and HF components, as well as spectral descriptor features, such as centroid, spread, skewness, kurtosis, crest, and spectral entropy. The study found that its Fuse-All model, which combined the traditional benchmark HRV features with the proposed subband features, gave better performance than the benchmark features alone. Another study [36] took a deliberate approach to compare the models developed in the lab versus those in the wild. They used a commercial off-the-shelf heart rate monitor in both laboratory and real-world settings on the same cohort of participants to ensure comparability. Stress labels collected in the wild were obtained through EMA-based self-reports. The results showed that the best-performing pipeline involved removing extreme values through trimming (ie, outlier removal), followed by data standardization using z-score normalization (trim_zscore). Furthermore, excluding periods of high physical activity further improved performance in real-world settings.

In addition to evaluating model performance using common metrics such as accuracy, an equally important aspect of model quality is transparency. Transparent models not only allow for better diagnosis and refinement but also create opportunities for interdisciplinary collaboration with experts from fields such as psychology, who bring deep knowledge of stress and its mechanisms. However, many current modeling approaches rely on black-box methods—particularly those based on deep learning and, more recently, generative AI. While these models may improve performance in certain cases, they also make interpreting results more challenging [59]. Future research should explore modeling approaches that achieve a better balance between transparency and accuracy. By integrating domain knowledge from psychology and physiology, such models could not only predict stress but also explain why certain predictions are made—uncovering not just correlational relationships, but potentially causal ones that offer insights for designing effective stress interventions. Additionally, developing models and systems that facilitate clear communication of model decisions to end users could open a new line of inquiry focused on ensuring human-in-the-loop approaches in stress modeling as explored in the study by Paraschou [47].

With the rise of LLMs, there is increasing interest in using them for mental health prediction tasks such as stress, anxiety, depression, and sleep disorders. One study included in this review demonstrated promising predictive performance using both larger LLMs, such as GPT-3.5 and GPT-4, and a smaller health-focused fine-tuned model, HealthAlpaca. The study suggested that performance depends not only on model size but also on prompting strategy, contextual information, and domain-specific fine-tuning, with smaller fine-tuned models performing comparably to or better than larger general-purpose LLMs. It also showed that the performance of larger models can be improved by providing examples through few-shot prompting or by adding contextual information to zero-shot prompts. However, several limitations remain. Wearable time-series data are high-dimensional, nonlinear, and continuous, making them more challenging for LLMs than conventional language inputs. In addition, the validity and interpretability of LLM-based predictions remain important concerns due to the lack of standardized evaluation benchmarks for nonlinguistic wearable data. These models also have substantial data and computational demands, as well as the risk of false-positive or hallucinated outputs in health-related settings. Future research should therefore examine explainability, validity, reliability, and whether LLM-based approaches provide meaningful advantages over more conventional models in wearable-based stress prediction.

Our review highlights several recurring data quality and validity challenges in wearable-based stress studies conducted in naturalistic settings. In particular, field-collected datasets are often shaped by study design constraints, resulting in nonuniform data distributions, limited sample sizes, short observation periods, and skewed participant characteristics (eg, gender, age, and other demographic attributes). These factors collectively constrain population representativeness and limit the external validity and generalizability of reported findings. In contrast, studies that rely on secondary or publicly available large datasets typically benefit from longer monitoring durations and more heterogeneous participant pools. While such datasets may offer improved statistical power and demographic coverage, they also introduce trade-offs, including reduced control over data collection protocols, labeling procedures, and contextual fidelity. Taken together, these differences underscore the importance of explicitly reporting dataset characteristics and carefully considering how data provenance and study design choices shape model validity and interpretability in real-world stress modeling. In our proposed model solution, we explicitly include sections on dataset characteristics to encourage robust study design, as well as transparency and standardization in reporting.

Limitations of this scoping review include the challenge of ensuring that all relevant studies have been included and none have been overlooked due to language constraints, as well as the exclusion of unpublished or ongoing studies. Additionally, we focused on a limited range of databases, which restricts the scope of the studies retrieved. The search criteria and query string used may also limit the outcomes of the scoping review. Rather than implementing full parallel double-screening of all records, we adopted a variant of the screening approach in which one researcher conducted the initial screening, followed by a review by a second reviewer. As a result, interrater reliability metrics, such as Cohen κ or percentage agreement, were not calculated. The absence of formal double-screening and corresponding reliability statistics may limit the reproducibility of the screening process.

Wearable devices offer valuable opportunities to measure and understand how stress manifests in real-world settings. ML and other advanced computational and statistical methods have increasingly been adopted to develop models that detect, characterize, and understand stress and related psychological constructs from physiological signals captured by wearables, as well as from self-reported data collected at increasingly fine temporal granularity using methods such as EMA. In this focused review, we examined recent studies that apply ML and advanced computational methods to model stress using data from wearable devices in the wild. This review provides an in-depth assessment of current modeling practices and reported performance in this domain and adds to existing reviews by focusing specifically on problem formulation, detailed analysis of model performance, and methodological rigor. In particular, we highlight key problem formulation decisions, such as input and output window configuration, and make the distinction between nowcasting and forecasting tasks, and conduct a detailed analysis of model performance across varied modeling approaches. Through this analysis, we observe a few general patterns of model performance as a function of modeling approaches. We also identify important methodological issues, such as potential data leakage, which may threaten the validity of reported results and inflate reported model performance. We also noted a lack of standardization in datasets, task definitions, and reporting practices, making cross-study comparisons challenging and potentially slowing progress in wearable-based stress detection research. In response to these findings, we propose a model card framework to guide modeling decisions, experimental design, and reporting practices, with the goal of promoting methodological rigor, enhancing transparency, and reproducibility. We hope this review serves as a roadmap for new researchers entering the field, while also offering a framework for researchers working on similar problems to share findings and collaboratively advance the field as a community.