JMIR Publications

ABSTRACT

Background:

College-aged students experience persistent academic and social stress that can negatively impact their mental and physical health. Digital phenotyping of mental illness using wearable devices and machine learning offers a promising approach to monitoring stress in real time by integrating continuous physiological signals. These multimodal passive data streams allow early stress detection and facilitate just-in-time interventions that can recommend therapeutic solutions to support student well-being.

Objective:

This study aims to systematically review the literature to identify best practices and emerging trends in stress measurement using wearable technology and digital tools among college-aged students. Specifically, we sought to evaluate commonalities in sensor types, datasets, and machine learning approaches used for stress detection.

Methods:

A systematic search was conducted across medical and computer science databases, including Embase, PubMed, IEEE Xplore, and ACM Digital Library, for studies published between January 2020 and July 2025. Two reviewers independently screened records and extracted data where studies were included if they focused on stress detection using wearable or digital tools among college-aged students. Data extraction included sensor modalities, dataset characteristics, and machine learning algorithms used for stress prediction.

Results:

126 studies met the inclusion criteria and were included in the review out of our original 776 articles. Electrodermal activity (EDA) was the most frequently used physiological signal, appearing in 60.3% of studies, and wrist-worn wearable devices were the predominant sensing modality. Among studies that compared algorithms, Support Vector Machines (SVMs) were identified as the most commonly applied and best-performing model in 33.3% of cases. 65.2% of included studies relied on pre-existing datasets, and approximately 82% of those used the WESAD dataset, which contains only 15 participants. Demographic reporting was inconsistent, as 28.6% of studies did not report sex distribution, and only one study justified its sample size. The use of temporal modeling algorithms was limited, despite their importance for capturing the dynamic, time-varying nature of stress in real-world settings.

Conclusions:

Wearable sensing and machine learning show strong potential for real-time stress detection and personalized support among college-aged students. However, the evidence base remains limited by small and homogeneous datasets, inconsistent methodologies, and insufficient use of temporal modeling techniques that more accurately capture the dynamic nature of stress. Considering recent advances in sensing modalities, datasets, and analytical approaches, this review highlights these recurring gaps and emphasizes the need for more diverse datasets and advanced modeling approaches tailored to student stress patterns. Strengthening these areas could improve the accuracy and robustness of stress detection systems, enabling earlier intervention and reducing the mental health burden in students.