BackgroundPsychological stress is a common and dynamic mental health concern, but conventional assessment methods often rely on self-report questionnaires, clinical interviews, or contact-based physiological sensors, which may limit continuous and unobtrusive monitoring in everyday settings. Noncontact sensing technologies combined with artificial intelligence (AI) provide a potential approach for low-burden assessment by capturing physiological or physiology-related information without direct body contact.ObjectiveThis systematic review and meta-analysis aimed to evaluate the performance of AI models using noncontact physiological signals for psychological stress recognition, summarize current sensing and modeling approaches, and identify methodological limitations affecting practical translation.MethodsWe searched PubMed, IEEE Xplore, Web of Science, Embase, and APA PsycINFO from inception to May 6, 2026, and performed backward citation searching. Eligible studies used noncontact sensing modalities, including remote or imaging photoplethysmography, thermal imaging, or wireless radar, combined with machine learning or deep learning methods to identify psychological stress. Narrative synthesis was conducted for all included studies, and quantitative meta-analysis was performed for binary stress classification tasks.ResultsTwenty-one studies were included, most of which were conducted in laboratory or controlled settings. For binary stress classification, 8 studies contributing 17 model-level estimates and 13,910 sample instances were included in the primary meta-analysis. The pooled accuracy was 81.3% (95% CI 68.4%–91.5%), with substantial heterogeneity. Supplementary analyses showed pooled F1 score, sensitivity, and specificity values of 0.791, 0.809, and 0.697, respectively. Exploratory subgroup analyses suggested that stress label sources were associated with model performance, whereas sensing modality, signal category, validation approach, and model type showed no statistically significant differences.ConclusionsCurrent evidence supports the technical feasibility of noncontact AI-based approaches for psychological stress recognition. However, available studies remain limited by small samples, heterogeneous stress labels, insufficient participant-independent validation, incomplete performance reporting, and limited evaluation in real-world settings. Future research should prioritize standardized stress labeling, transparent model validation, comprehensive reporting of performance metrics, and naturalistic evaluation to improve the reliability and applicability of noncontact AI-based mental health monitoring approaches.