Machine Learning Frameworks for Wearable-Based Stress Modeling in Naturalistic Settings: Scoping Review

Stress, as commonly recognized, is an integral part of modern life and can significantly affect both mental and physical health. While substantial advancements have been made in measuring physical fitness through wearable devices, the detection and assessment of mental stress remain in their early stages.

The objective of this paper is to review recent studies of wearable-based stress detection in naturalistic settings, with a specific focus on characterizing machine learning frameworks inspired by the model card approach.

This review was conducted using the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) checklist. A total of 353 articles were identified through searches in databases such as PubMed, MEDLINE, ScienceDirect, IEEE, ACM Digital Library, Web of Science, and Embase. Studies were considered eligible if they collected data from healthy adults in naturalistic settings using wearable devices and used machine learning models for stress detection.

A total of 34 articles met the eligibility criteria, including 11 conference papers, 22 journal articles, and 1 preprint published between 2017 and 2024. From these studies, we analyzed key machine learning modeling decisions such as problem formulation, ground truth determination, and machine learning algorithms. Additionally, we examined the major contributions of each study, focusing on the challenges they addressed and the solutions they proposed. Based on these findings, we proposed a model card framework for reporting machine learning-based, wearable-based stress detection.

This scoping review highlights recent trends in machine learning models for stress detection and measurement using wearable signals. It underscores the need for improved standardization in reporting practices for datasets and key machine learning decisions, as well as the importance of addressing critical challenges associated with data collection in real-world settings. We hope this review will support and strengthen ongoing research efforts, promote knowledge sharing, and promote collaboration among researchers-ultimately advancing the field as a community.