Stress Detection Using Machine Learning With Multimodal Data: A Systematic Literature Review

Abstract

Stress is a major determinant of mental health and productivity. Consequently, continuous, unobtrusive stress detection using wearable sensors and machine learning (ML) has become a key priority in digital health. This paper presents a Systematic Literature Review (SLR) of 19 peer-reviewed articles, selected from 36 initial papers via structured inclusion/exclusion criteria focusing on studies from 2021-2025 that report quantitative ML performance. We employed a quantitative and qualitative synthesis to analyze and map five key dimensions: sensing modalities, ML/DL algorithms, datasets, validation protocols, and societal feasibility. Findings reveal a clear state-of-the-art: multimodal physiological fusion (notably PPG, EDA, and ACC) paired with hybrid deep models (CNN-LSTM) consistently achieves the highest accuracy (85–96%) on benchmark datasets. Our research reveals a significant lab-to-field gap. Most studies utilize intra-subject or k-fold cross-validation, whereas the more robust Leave-One-Subject-Out (LOSO) validation is hardly employed, constraining model applicability. Furthermore, fewer than 15% of studies explicitly address vital practical constraints such as privacy, computational efficiency (Edge AI), or power consumption. This review methodically quantifies the gap, emphasizing that current models, despite their accuracy, are not yet suitable for real-world implementation. We conclude with actionable directions toward generalizable, lightweight, and privacy-aware stress-aware systems.