Systematic review and meta-analysis wearable device artificial intelligence depression detection and depressive episode prediction performance (Preprint)

BACKGROUND Depression is a globally prevalent psychiatric disorder, projected to become the leading cause of disease burden by 2030. Traditional diagnostic methods rely heavily on subjective assessments, lacking objective biomarkers and real-time monitoring capabilities. Recent advancements in wearable technology and artificial intelligence (AI) offer promising tools for depression detection and monitoring. However, the diagnostic accuracy of wearable-based AI algorithms remains variable and insufficiently evaluated. OBJECTIVE This study aims to systematically review and meta-analyze the diagnostic and predictive performance of wearable device–based AI models in detecting depression and predicting depressive episodes, and to identify key factors influencing diagnostic outcomes. METHODS Following PRISMA-DTA guidelines, a comprehensive search of PubMed, Embase, Web of Science, and PsycINFO was conducted up to May 27, 2025. Eligible studies applied AI algorithms to data from wearable devices for depression detection or episode prediction. Data on diagnostic performance (sensitivity, specificity, diagnostic odds ratio [DOR], area under the curve [AUC]) were extracted and analyzed using a bivariate random-effects model. Risk of bias was assessed using the PROBAST+AI tool, and evidence certainty was evaluated via GRADE. RESULTS Sixteen studies comprising 32 internal validation datasets were included, totaling 1,189 patients and 13,593 samples. The pooled sensitivity and specificity for depression detection were 0.89 (95% CI, 0.83–0.93) and 0.93 (95% CI, 0.87–0.96), respectively. The pooled DOR was 110.47 (95% CI, 33.33–366.17), and the AUC was 0.96 (95% CI, 0.94–0.98). Random forest algorithms demonstrated the highest diagnostic performance (sensitivity: 0.89; specificity: 0.91; AUC: 0.97). Subgroup analyses revealed that study design, AI method, reference standard, and data input type significantly influenced diagnostic outcomes (P < 0.05). For depressive episode prediction, based on 3 datasets, sensitivity was 0.86 (95% CI, 0.80–0.91) and specificity was 0.65 (95% CI, 0.59–0.71). The overall risk of bias was low to moderate, and no publication bias was detected. CONCLUSIONS Wearable device–based AI models show high performance in detecting depression and moderate potential in predicting depressive episodes. However, substantial heterogeneity exists across studies, with predominant retrospective designs and reliance on public datasets limiting generalizability. Further large-scale, prospective studies are needed to validate these findings and standardize methodologies for clinical implementation.