We conducted a scoping review to characterize current research on stress detection using wearable and digital tools among college-aged students. This review synthesizes studies published between January 2020 and December 2025 to summarize commonly used wearable devices, physiological signals, datasets, and machine learning or deep learning approaches for identifying high-stress moments. By organizing existing methods and conducting a quality assessment, this review provides an overview of methodological practices and highlights areas for future research in wearable-based stress measurement. This scoping review adhered to the methodological framework proposed by Arksey and O’Malley [34], which includes identifying the research question, identifying relevant studies, study selection, charting the data, and collating, summarizing, and reporting the results. Finally, this scoping review was conducted and reported in accordance with the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines to ensure transparency and reproducibility [35].

No formal review protocol was registered for this scoping review, as the objective was to map the scope and characteristics of existing evidence in stress prediction research using wearable technology.

We defined eligibility criteria to ensure that only relevant and methodologically appropriate studies were included in this review. Studies were included if they measured or classified psychological stress using physiological signals from a tool, wearable, or sensor. Only experimental or observational studies published in English were considered. The target population was college students aged 18‐24 years. Studies that partially included this age range were eligible if they explicitly mentioned students as a distinct group or if the mean age, along with the SD, fell within the target population. Studies were excluded if they focused on nonpsychological stress (eg, mechanical stress), were review papers, extended abstracts, or prototype descriptions without a defined study population. Papers without clear population details or those identifying participants solely by employment (eg, “office workers” or “hospitalized patients”) were also excluded.

We searched IEEE Xplore, ACM Digital Library, PubMed, and Embase for conference and journal papers covering studies published between January 2020 and December 2025, a time frame selected to capture recent developments in wearable sensing technologies and stress detection methodologies.

We used a combination of terms related to the key concepts of psychological stress, wearable devices, and sensors (full search per database is provided in Multimedia Appendix 1). We extracted each database searched and the platform used, including IEEE Xplore, ACM Digital Library, PubMed, and Embase, in accordance with PRISMA-S (PRISMA literature search extension) [36], and all databases were searched independently rather than through a multidatabase platform. No multidatabase searching or study registry searching was conducted. No additional online resources (eg, tables of contents, print conference proceedings, and websites) were browsed. No additional search methods were used, including citation searching, contacting authors or experts, or setting up citation alerts. The full search strategies for each database are provided in Multimedia Appendix 1, including the specification that no filters or limits other than language (English) and publication date (January 2020 to December 2025) were applied. Search strategies were developed with input from 2 academic librarians; however, search strategies from prior reviews were not reused, and no formal peer review of the search strategy was conducted. No additional methods were used to update the search. Therefore, searches were limited to studies published in English within the specified date range. No restrictions were applied based on study design. All retrieved records were initially screened. Following screening, records were imported into Rayyan (Rayyan Systems Inc) [37], where duplicate entries were identified and removed. The deduplicated set of records was then used for abstract and full-text screening.

Two independent reviewers screened all records using a 2-stage selection process. Studies were checked for eligibility by 2 reviewers independently screening titles and abstracts. This first round of filtering focused on relevance. Abstracts were also screened for population. Some papers did not mention population in the abstract and were thus moved to full-text screening. This resulted in 261 papers for full-text screening. During this second round of filtering, studies were also checked for eligibility by 2 researchers independently reviewing the full text. Disagreements at any stage of eligibility and filtering were resolved by the 2 reviewers discussing their reasons for either inclusion, exclusion, or neither. Full agreement was reached for abstract and full-text screening, leading to the final inclusion of 134 papers.

A standardized data-charting form was jointly developed by 2 reviewers to identify and extract relevant information aligned with the review objectives. The form was pilot-tested on a subset of included studies and refined iteratively to ensure completeness and consistency. Two reviewers independently charted data from all eligible studies, compared their entries, and resolved discrepancies through discussion. All data were extracted directly from the published papers, and no additional information was sought from study authors.

To extract consistent information from each paper, we conducted systematic data extraction as outlined in Tables 1-3. Extracted variables included study details (title, authors, publication date, study purpose, and data collection duration), sample characteristics (age, sex, sample size, and demographic information), sensor type, and all available feature categories used in the study (sleep, physiological signals, calorie intake or expenditure, phone use, activity, location, and survey or EMA data). For studies conducting algorithm comparisons, we additionally extracted the types of signals analyzed, devices used, algorithms tested, performance measures, best-performing algorithm, validation strategy, and outcome measures.

Although critical appraisal is not required for scoping reviews, we conducted an assessment of study quality to better contextualize the strengths and limitations of the included evidence. To address the quality of each paper, we scored every paper across 4 categories on a scale from 0 to 2 as described in Multimedia Appendix 2 and shown in Multimedia Appendix 3. Given the diverse study designs among the extracted papers, we adopted a methodology similar to that used by De Angel et al [155]. This approach integrates the AXIS appraisal tool [156] for cross-sectional studies with the Newcastle-Ottawa Scale [157] for longitudinal studies. Papers were assessed using a 3-point scoring system: 2 points for fully meeting the criteria, 1 point for partial fulfillment, and 0 points for nonfulfillment.

Effect measures extracted from the included studies consisted of accuracy, F₁-score, sensitivity, specificity, precision, recall, and other performance metrics reported for stress detection. These measures were used to compare model performance across studies. For population characteristics, the mean age and corresponding SDs were extracted whenever available.

Due to differences in study designs, methodologies, and outcome reporting, results were synthesized descriptively. Key study characteristics, signals measured, algorithms used, and sensor types were organized into structured tables to enable comparison across studies. Frequencies of the most commonly measured signals, best-performing algorithms, and most-used sensors were calculated and visualized using bar plots. Missing summary statistics were extracted as reported, with no additional transformations applied. No meta-analysis, subgroup analysis, or meta-regression was conducted; instead, the synthesis focused on identifying overarching trends across the included studies. Because the focus of this review was to characterize stress detection methods used among college-aged populations, we extracted data elements that were directly relevant to the review objectives, including participant characteristics, sensor types, physiological signals, analytical methods, and model performance outcomes. Broader intervention-related data items (eg, intervention protocols, adverse event reporting, and clinical outcome metrics) did not apply to the observational and experimental studies included in this review. Therefore, the extraction approach was intentionally streamlined to ensure consistency, interpretability, and comparability across heterogeneous study designs. In addition, we developed an evidence gap map to conceptually organize and summarize the literature across study conditions, methodological enablers, analytical approaches, barriers, and outcomes, highlighting recurring patterns as well as persistent gaps, following a prior standardized method [158].

This study is a systematic review of previously published literature and did not involve the collection of primary data from human participants. No new data were generated, and no individuals were directly recruited, observed, or intervened upon as part of this research. Accordingly, a formal review by an Institutional Review Board or Research Ethics Board was not sought. This determination is consistent with standard guidance that systematic reviews relying exclusively on publicly available, previously published data do not constitute human participant research requiring ethics board oversight.

All included studies were previously published in peer-reviewed journals and were assumed to have undergone appropriate ethical review by their respective authors and institutions before publication. No personally identifiable information was accessed, extracted, or reported at any stage of this review. The conduct of this review adhered to the ethical principles outlined in the World Medical Association Declaration of Helsinki and complied with applicable institutional, regional, and international standards for research integrity.

Records were screened from IEEE Xplore, ACM Digital Library, Embase, and PubMed, with most records coming from technical journals. A total of 134 studies were included in the review out of the original 792 records, as illustrated in Figure 1 and Multimedia Appendix 4. Forty-eight records were removed after deduplication. Of the remaining records, 483 were excluded after 744 abstracts were screened for relevance. In total, 127 records were excluded after 261 full texts were screened for relevance and correct population. Summary characteristics of the final 134 included studies are provided in Figure 1 and Table 1.

Our population of interest included college-aged students aged 18‐24 years. In terms of sex demographics, about 72.4% (97/134) of studies specified the number of participants who were male, female, or nonbinary. Among the selected studies, 2 of 134 (1.5%) studies [39,137] failed to mention a sample size. Across the studies that reported sex distribution, most had a higher proportion of male participants than female, indicating a demographic imbalance that may limit the generalizability of findings. In terms of racial demographics, in 42 papers published from 2020 to 2022, about 9.5% (n=13) of papers included race distribution across their sample population, and 21 (50%) studies included other relevant health information, including preexisting conditions, mental health, and underlying illnesses. From papers published from 2020 to 2025, 26 (19.4%) studies were conducted in Europe, 14 (10.4%) studies were conducted in Asia, 2 (1.5%) studies were conducted in the Middle East, 21 (15.7%) studies were conducted in the United States, 3 (2.2%) studies were conducted in South America, and other studies did not explicitly mention where they were conducted. The higher number of studies conducted in Europe and the United States compared to Asia and other regions suggests regional variations in digital health adoption, research funding, and accessibility of wearable technologies. These differences may influence trends in stress detection research, highlighting the need for region-specific digital health strategies to address varying technological infrastructures, health care priorities, and user needs.

More than half (62.8%, n=134) of the studies used preexisting datasets to implement their method of stress measurement. The rest of the studies were experimental in nature and carried out “de novo” data collection. Seventeen studies [32,45,46,54,57,61,65,72,73] were longitudinal in nature, published from 2020 to 2022, and 10 studies [3,76,82,89,91,95,99,105,152,154] were longitudinal in nature, published from 2023 to 2025, meaning data were collected for the same study population over a period rather than collected at one time point cross-sectionally. These longitudinal data consist of repeated observations at the individual level rather than data collected at multiple time points across different populations. Individual-level effects are confounded with cohort effects in cross-sectional studies, so being able to isolate and study the effect of time as a repeated measure is critical. Of these longitudinal studies published from 2020 to 2022, 2 were clear in addressing how they handled missing data. These studies either imputed missing values with each person’s channel-wise mean values of the day, where days with >25% sensor data missing were discarded [45], or removed missing data [57]. It is difficult to collect comprehensive, complete data from sensors longitudinally, where data are not always complete for each participant. About 6.6% (n=9) of studies included a recruitment method for participants. Two studies used volunteers, and 1 study invited participants to participate.

The extracted studies were classified into 3 primary methodological categories: algorithm comparisons (shown in Table 2), the development of custom stress measurement frameworks, and statistical analyses (illustrated in Table 3). Studies focusing on algorithm comparison primarily used 2 approaches: machine learning models, such as support vector machines (SVMs), random forest (RF), k-nearest neighbors, and extreme gradient boosting (XGBoost), which used handcrafted features for stress detection, or deep learning methods, such as convolutional neural networks (CNNs), to automatically extract relevant features [159]. Among the studies reviewed, SVM demonstrated the highest performance, with 33.3% (n=45) of papers identifying it as the best-performing algorithm, as illustrated in Figure 2. In comparison, 11.1% (n=15) of the studies reported CNN as the best-performing model [50,58,67,103]. One study evaluated 3 boosting algorithms—XGBoost, Light Gradient Boosting Machine, and CatBoost—tree-based ensemble methods that iteratively improve weak learners to enhance classification, evaluating the effectiveness of data fusion methods [115].

One paper [31] focused on comparing long short-term memory (LSTM) and a combination of LSTM and CNN. This study found LSTM alone to perform better. Two studies in Table 3 that focused on a single framework supported the use of a single modality or a modest set of signals [66,69]. Studies that focused on the comparison of chest and wearable devices found chest devices to perform better [84,109], but chest devices in combination with wrist devices performed the best [71]. Most of these studies focused on time-agnostic algorithms, as shown in Table 2. We also found studies using wrist wearables (eg, Empatica, Microsoft Smartband 2, Fitbit Charge 2, and Samsung Galaxy Gear Sport Watches) and chest-worn devices, with core physiological signals such as EDA, galvanic skin response, HR, photoplethysmography, HRV, respiration, or temperature, evaluated using k-fold cross-validation, leave-one-out cross-validation, or leave-one-subject-out evaluation, and reported performance metrics such as F₁-score, accuracy, precision, and recall. In the “best” column, classic machine learning models were most often SVM, followed by RF, while deep learning wins were fewer (occasional CNN, deep neural network, and a single LSTM). Few studies in Table 2 incorporated nonphysiological or contextual signals [61,72,73]. Recent studies examining the association between sleep and stress have leveraged data from the Oura Ring [3,99]. Two recent studies using Garmin smartwatch–derived stress scores found significant associations with high- and moderate-intensity positive mood in 1 study [154], while another reported a stronger association of HR with self‐reported stress, and the Garmin stress score demonstrated marginal predictive value [24].

Studies mainly aggregated stress on a binary or 3-tier scale, meaning participants were either identified as stressed or not stressed, as opposed to being measured on a continuous scale. Here, a continuous scale captures stress fluctuation over time rather than binary or categorical labels. Sensors or tools used to measure physiological signals included various wrist, chest, and full-body sensors alongside mobile surveys. Figure 3 details the various devices used and shows that wrist sensors, in general, were the most widely used sensor type. About 72.4% (n=97) of the studies used well-validated stress tests or tasks for their models, such as the TSST [160], mental arithmetic tests, video stimuli, the Stroop color word test, startle response tests, cold-pressor tests, or public speaking, to reliably trigger stress responses while incorporating restful periods as a baseline [22]. About 8.3% (n=11) of the studies used self-reported SMS text messaging surveys in their supervised machine learning models. The various physiological features and signals measured are illustrated in Figure 4. The most common signal was EDA, appearing in 57.5% (n=77) of studies. Figure 4 shows the top signals measured per study, including instances where papers used multiple signals together.

Of the 62.8% of studies that used some preexisting datasets, around 80% (n=67) used the Wearable Stress and Affect Detection (WESAD) dataset, for instance, including papers published from 2020 to 2022 [18,38,42,43,48,55,56,58,59,63,66-71] or a few published from 2023 to 2025 [2,28,75,79,86,94,103,113,122,141]. This dataset was commonly referenced in papers included in this review. This dataset is publicly available and is a widely used dataset for stress and affect detection [161]. The mean age of participants is 27.5 years with a SD of 2.4 years. The sample included 3 females and 12 males for a total of 15 participants. Heavy smokers and pregnant women were excluded, and the participants were composed of graduate students. The signals collected include physiological and motion data from chest-worn and wrist-worn devices. Measurements include blood volume pulse, ECG, EDA, electromyography, respiration, body temperature, and 3-axis acceleration. The protocol used elicits 3 emotional states: baseline, stress, and amusement, followed by a meditation phase. Benchmarks for comparison used the well-studied stress induction method, the TSST, with 0.93 accuracy and 0.91 F₁-score for distinguishing stress, using a linear discriminant analysis classifier, using only chest-based physiological signals.

Although many papers used this same dataset, they experimented with different physiological signals as well as motion data when extracting features for modeling. Modeling and validation methods also varied. The algorithms with the best performance when applied to the WESAD dataset included SVM, RF, XGBoost, k-nearest neighbor, decision tree, deep neural network, self-supervised learning, artificial neural networks, large language models, and CNN. In addition to WESAD, recently published papers used other datasets, including SWELL [29], AffectiveROAD [81], VerBIO [96], S-TEST, or DS-3 [101].

Figure 5 shows a breakdown of quality score assessments for all extracted papers, broken down into 4 categories. Papers were scored 0, 1, or 2 for each category. An explanation of each category’s scoring is provided in Multimedia Appendix 2, and the individual score breakdown by category for each paper is provided in Multimedia Appendix 3. In general, outcomes and sample descriptions were clearly stated, with most papers having a quality score of 2. However, representativeness and justification of sample size were areas in which many papers did not perform as well. Representativeness was cited as a common issue across many papers, as samples were limited due to recruitment processes for participants or the data that were available. The samples were also limited by age due to the demographic of interest in this review. Around 27.6% (n=37) of papers failed to give sex demographic information. Most papers analyzed used experimental data from other sources or open-source, publicly accessible datasets such as the WESAD dataset, which did not justify the chosen sample size. From papers published from 2020 to 2022, only 2% of papers failed to give sample size information; however, sample size justification was rarely given, although the papers that did address this issue cited their voluntary recruitment process as a limitation. Almost none of the studies analyzed did a power analysis to determine sample size before running the stress studies, which is a major shortcoming. Across recent papers published from 2023 to 2025, almost all clearly defined outcomes and described their samples, but very few addressed representativeness, and only 3 papers [24,143,154] justified their sample size published, highlighting a major gap in methodological rigor.

Finally, these findings point to substantial heterogeneity and a meaningful risk of bias across the included studies. The wide variation in sample sizes, inconsistent reporting of demographic characteristics, limited disclosure of health information, and strong geographic skew toward Europe and the United States contribute to structural differences that complicate direct comparison of results. This heterogeneity is further shaped by the heavy reliance on the WESAD dataset, a publicly available dataset with only 15 predominantly male participants, with a mean age of 27.5 years, which results in many studies concluding a small and demographically narrow sample. Such repeated use of a single dataset increases the likelihood that reported model performance reflects the characteristics of WESAD participants rather than capturing variability among college-aged students. Accordingly, the synthesized findings should be interpreted with caution, acknowledging that both heterogeneity in study design and risk of bias in sampling and reporting may influence observed performance patterns and limit the extent to which results can be generalized. Using a relational synthesis approach, Figure 6 presents an evidence gap map that synthesizes methodological enablers, study conditions, stress prediction approaches, barriers, and outcomes observed across the included studies. The map illustrates a research landscape shaped by publicly available datasets, standardized in-laboratory stress protocols, and widespread use of wrist-worn physiological sensors. At the same time, it highlights recurring constraints including a predominance of laboratory-based study designs, heavy reliance on publicly available datasets, and limited demographic representativeness. While many studies report strong classification performance using classical machine learning models under controlled conditions, comparatively fewer examine temporal stress dynamics, personalization, or real-world deployment.

In this scoping review, we examined how stress is measured among college-aged students using wearable technologies and machine learning methods between 2020 and 2025, to identify commonly used wearables, the most informative physiological signals, and the best-performing algorithms. Across the literature, we found that SVMs among traditional machine learning models and CNNs among deep learning models were the strongest performers for stress classification. Wrist-worn devices were the predominant sensor platform, and EDA was the most frequently measured and most informative signal. However, most studies relied on small, homogeneous samples, frequently used controlled laboratory datasets such as WESAD, and commonly used binary (stressed vs not stressed) labeling approaches, raising concerns about representativeness and ecological validity. Our quality assessment further revealed inconsistent demographic reporting, insufficient justification of sample sizes, limited attention to social determinants of stress, and substantial variation in how psychological stress was defined, elicited, and validated across studies.

Regarding stress prediction model performance, the strong performance of SVMs can be attributed to their robustness in handling high-dimensional physiological data [33,144], their ability to generalize well by maximizing the margin between classes, and their effectiveness in small and imbalanced datasets, which are common in stress detection studies [162]. Additionally, the flexibility of SVM in using different kernel functions [163] allows them to model complex, nonlinear relationships in physiological signals without requiring deep feature extraction. These advantages likely contribute to their superior performance compared with other traditional machine learning models in stress classification. However, SVMs are computationally expensive and may not be practical for real-time applications [164]. More efficient and scalable approaches are needed to enhance practicality in the field. Deep learning models, particularly CNNs, outperformed traditional machine learning approaches in comparative analyses [82,85]. Although CNNs capture spatial patterns in temporal data, they do not have memory in their architecture, reducing their effectiveness on longitudinal temporal data [165], indicating a need for algorithms that explicitly model temporal patterns, such as RNNs [74]. One study focusing on the comparison of various machine learning and deep learning methods attempted to use a version of an RNN in the form of an LSTM. This paper reported the greatest performance with LSTM alone, as opposed to a combination of LSTM and CNN, indicating some value in noting and using temporal patterns. In addition, emerging evaluations of large language models for stress prediction [135] did not perform well and suggest that parameter count does not consistently correlate with performance. For example, GPT-3.5-Turbo performed comparably to GPT-4 on WESAD [109]. These findings indicate that identifying key biomarkers is essential for improving model efficiency [115]. From 2023 to 2025, published literature emphasized personalization and multitask learning to enhance stress-prediction performance and generalizability [70,79,98,107,112,127]. In addition, 1 study explored stress detection in a virtual reality environment integrated with an Internet of Things system, demonstrating the potential of immersive technologies for stress monitoring [85].

Wrist wearables were most commonly considered [166] as they seem less encumbering than full body or chest wearables [22] while attaining better measurement of physiological signals than surveys or smartphones. Other wearable sensors used across studies included chest wearables, full body sensors, or some combination of chest and wrist wearable signals. We saw that EDA was the most frequently measured signal across papers and is important in stress detection [167], since it provides valuable information about a person’s sympathetic nervous system activity, which is closely linked to emotional responses, including stress. Most papers used multiple signals in their model building, with EDA most commonly contributing to a more accurate model. For instance, building a stress detection model incorporating both HR and EDA [22,26,81] data might allow for a more comprehensive, accurate, and context-aware assessment of stress and other emotional responses. Ensuring the reliability and reproducibility of physiological measurements is crucial for real-world stress detection [26]. Variability in sensor accuracy, signal quality, and environmental factors can impact consistency [22]. Validating models across diverse settings improves generalizability and practical applicability [168].

We saw that most studies used a binary model of stress in which an individual is identified as either stressed or not stressed. A few studies extended beyond binary classification by using multiclass stress prediction (eg, 3-class [62] or 5-class [59,125] models), which allows a somewhat finer-grained view but still treats stress as discrete states. There is a need for a model more in line with how human stress manifests, such as a continuous scale [26,169]. For example, an individual might feel mildly stressed, which is worth noting and which cannot be captured on a binary scale of stress [150]. On a binary scale, mild stress may be interpreted as either diminished or heightened stress. A continuous scale for stress monitoring is valuable for capturing individual differences and for understanding the dynamic nature of stress [150].

We found a lack of detailed explanations on how psychological stress was identified. Accurately distinguishing psychological stress from other physiological responses is complex, as HR alone is insufficient for stress detection [154]. For example, HR alone cannot reliably indicate stress, as an elevated HR may result from various factors [170], such as jogging or facing an unprepared mathematics test. A stress detection model based solely on HR data could misclassify natural variations in HR, such as those caused by excitement or physical activity during social events, as stress, leading to inaccurate assessments [169]. One critical detail to note in studies of stress is the differentiation between physiological and mental stress. This distinction is complicated for wearable devices [154]. To accommodate this, studies need to look at a participant’s resting data while they are confirmed to be stressed, as well as their accelerometer data, if necessary, to check movement patterns, and consider these factors while detecting significant stress moments [169]. One’s activity must be noted to clearly identify psychological stress. Many studies used some well-validated stress tasks to account for this concern, but could benefit from clearer explanations of how their stress tasks accommodate this issue. These stress tasks mostly used tests such as mental arithmetic, Stroop test, public speaking, or cold-pressor tests, with participants putting their hands in ice water, to benchmark stress [22,26]. By contrast, other datasets (eg, “A Wearable Exam Stress Dataset for Predicting Cognitive Performance in Real-World Settings” [124]) inferred stress levels indirectly from examination grades, raising concerns about the accuracy of stress labeling. Studies that did not incorporate a stress task often used self-report surveys to monitor whether someone is stressed [168,171]. Self-report measures often face challenges with accuracy and completeness [172]. While frequent and timely survey prompts can improve accuracy, they do not fully address issues of completeness. Additionally, repeated survey checks may increase participant burden, potentially leading to survey fatigue and lower response rates [173]. There is also a need for better transparency regarding the wording of questions and the frequency of surveys to ensure consistency and minimize bias [174].

When analyzing the quality of research, we saw a need for larger sample sizes [175]. Larger sample sizes help reduce bias, provide a better representation of the target population, and lower the impact of outlier participants [176]. We observed that many studies relied on the WESAD dataset [177], a widely used dataset for stress and affect detection. However, WESAD includes only 15 participants, making it a limited representation of the college student population. Additionally, since WESAD data were collected in a controlled laboratory setting, they do not reflect real-world (“in the wild”) stress detection, where external factors and daily life variability play a significant role [171,178,179]. In fact, 1 study that used WESAD achieved strong performance under laboratory conditions but failed to generalize effectively in real-world settings [119], further underscoring the limitations of laboratory-based datasets.

Many studies did not report racial or ethnic demographics or have a representative sample regarding sex. This was a commonly identified issue within papers, as many samples relied on volunteers. Many papers also failed to report on other demographics of their samples besides sex or ethnicity, such as populations for exclusion. This includes excluding populations taking certain medications, populations with certain mental health histories, populations engaging in drug use, or pregnant populations. Knowing the populations for exclusion is crucial for replicability and transparency, as well as for bias detection and interpretation of results [180-182]. Although our population of interest was students, there is a need for more varied student demographics in samples regarding sex, race, and ethnicity, capturing different social determinants [183]. Given that stress is influenced by various social determinants [184,185], future studies should incorporate factors such as socioeconomic status, neighborhood context, physical environment, racial minority representation, and health-lifestyle interactions [186]. Including these elements would provide a more comprehensive understanding of stress in college students. One paper mentioned that its sample may not be representative because participants were recruited from an elite, private university [32]. Along these lines, there is a need for better justification of sample selection as well as sample size. Finally, missing data present a significant challenge in stress studies, affecting both comparability across studies and the reliability of findings [187]. The way missing data is handled, whether through imputation, exclusion, or other techniques, can influence study outcomes and lead to biased conclusions [188]. There is a need for more complete data and more detailed descriptions of how missing data were handled, particularly in longitudinal studies [189].

Prior literature reviews have explored various aspects of stress detection using wearable technology and machine learning. A meta-analysis examined the effectiveness of wearable AI in diagnosing and predicting stress among students, while emphasizing the need for real-world validation and improvements [190]. Another review categorized stress detection approaches based on different wearable sensor types and environments such as driving, studying, and working [191]. A separate study systematically assessed biosignal responses to psychological stress, analyzing electroencephalogram, ECG, EDA, HRV, respiration, and temperature to evaluate their reliability and consistency [192]. A prior review also examined machine learning techniques used in stress monitoring research, focusing on model generalization when training on public datasets [20]. Another review focused on wearable technologies and smart devices for detecting depression, anxiety, and stress, discussing physiological markers such as HRV, EDA, and electroencephalogram, along with their market availability [193]. Finally, a review analyzed physiological parameters such as HR, temperature, humidity, blood pressure, and speech, exploring various stress detection sensors and machine learning-based classification techniques [194]. Our scoping review extends this literature by specifically focusing on stress measurement in college-aged students, reviewing recent papers published from January 2020 to December 2025, analyzing common datasets, sensor types, and the best-performing machine learning algorithms used in research. We also evaluate weaknesses in current methodologies through a quality assessment while identifying best practices in study design, feature selection, sensor use, and algorithmic approaches.

Taken together, the findings of this scoping review highlight that progress in wearable-based stress detection for college-aged students [3,32,46,73] is constrained primarily by methodological and conceptual design choices rather than sensor availability for digital phenotyping of stress [195] or algorithmic capacity [18,28,30]. While multimodal physiological sensing, particularly EDA combined with cardiac measures, shows consistent promise [22,26], the field remains highly reliant on small, controlled datasets such as WESAD [177] and binary stress formulations that fail to capture the continuous [26,169], context-dependent nature of stress in students’ daily lives [171]. Advancing this area will require a shift toward larger [175], more diverse cohorts that reflect different social determinants of health [186], and real-world datasets that support generalizable human behavior modeling [168,196]; along with transparent reporting of participant characteristics, exclusion criteria, and missing data handling [189]; and modeling approaches that explicitly account for temporal patterns [95], personalization [1], and contextual information from students’ behavioral patterns [152]. These improvements are not only methodological but also ethical, and without representative samples and robust validation in real-world settings, stress detection systems might risk reinforcing bias [197] and producing misleading inferences when deployed in student populations [183]. By synthesizing recent evidence and identifying persistent gaps, this review provides a foundation for designing more reliable, interpretable, and equitable stress monitoring systems that can support just-in-time interventions and inform institutional strategies to improve student mental health [5].

Our focused and systematic approach targeting stress in college students in recent years allows for a more detailed analysis. Recency allows for analysis of the most up-to-date and commonly used sensors as well as the newest algorithms. By systematically categorizing the approach taken by each study, along with the devices used and signals measured, we can synthesize the information, establish trends, and make conclusions about best-performing methods and practices. Many studies relied on commonly used datasets, such as WESAD. Using the same dataset across different research projects enables benchmarking, allowing for direct comparison of methodologies and an understanding of why results may vary across approaches. A common challenge in the reviewed papers was the inclusion of multiple populations or datasets within a single study. While our primary focus was on college students, some papers analyzed mixed populations or multiple datasets. However, as long as college students were included, these studies were still considered in our review. Many papers also used overlapping datasets such as the WESAD dataset, although different papers used different parts of the dataset along with different models. This may lead to some redundancy in findings. The commonly used dataset, WESAD, with only 15 participants, had limited sample sizes, introducing potential bias and reducing the likelihood of capturing a truly representative population. Additionally, only studies published in English were included, as this was the language accessible to our reviewers, which may have led to the exclusion of relevant research.

This scoping review provides a focused synthesis of wearable- and digital tool–based stress detection research specifically among college-aged students, a population often overlooked or aggregated with broader adult samples in prior reviews. Current research highlights the need for larger and more diverse samples to improve representativeness, as many studies rely on a limited number of existing datasets, potentially leading to overlapping findings. Greater diversity in sex and ethnic demographics, along with clearer justification of sample sizes and improved demographic reporting, is essential for understanding population-level stress patterns. Methodologically, most studies conceptualized stress as a binary state (stressed vs not stressed), failing to capture variations in intensity, such as mild or moderate stress that can be chronic and clinically meaningful. Few studies used algorithms such as RNNs, which can capture temporal patterns, despite the importance of tracking stress progression over time. Greater emphasis on time-dependent modeling could enhance the understanding of how stress evolves. Many studies failed to clearly distinguish between psychological stress and physiological stress responses, despite the critical need for distinct measurement approaches. More precise definitions and methodologies are necessary to differentiate between these 2 aspects of stress effectively. In real-world settings, these limitations constrain the generalizability and clinical usefulness of stress detection systems.

To strengthen the credibility and generalizability of future research, studies should provide clear justifications for their sample sizes and, where possible, aim to recruit larger cohorts that reduce bias and improve statistical reliability. The field would also benefit from the development and use of more varied datasets, which can limit overlap across studies and reduce potential sources of bias. Increasing diversity in participant recruitment is essential; researchers should ensure representation across race, sex, socioeconomic status, and environmental contexts, as well as variation in behavioral and lifestyle factors such as sleep duration and efficiency, physical activity, phone usage, social media engagement, and mobility patterns. Detailed demographic reporting should accompany all studies to enhance transparency and enable meaningful comparisons across research efforts. Future analytical approaches should incorporate algorithms capable of capturing temporal patterns to model fluctuations in stress over time. Rather than relying solely on binary stress categorizations, researchers should develop models that characterize stress as a dynamic and progressive state, allowing for the detection of mild, moderate, and chronic stress levels. Clear explanations of baseline stress measurements are also needed to ensure that resting conditions are consistently defined and comparable across studies. Finally, stress prediction models should increasingly focus on personalization while maintaining robust privacy protections for participants.