Integrating AI-driven technologies and facial-semantic features for depression detection: A cross-sectional study

Depression is a major global health concern, still individuals with depressive tendencies remain undetected in outpatient settings due to the limitations of self-report, stigma, and reporting bias. Advances in artificial intelligence (AI) offer opportunities for objective, scalable, and unobtrusive methods of early detection. This study examined predictive accuracy of two AI-driven systems, the iSeeME facial-expression model and the Event-Driven Depression Tendency Warning System (EDDTW-V2) in identifying depressive symptoms among high-risk outpatients in Taiwan.

A cross-sectional study was conducted with 62 outpatients recruited from psychiatric and surgical-oncology clinics at a medical center in southern Taiwan. Participants completed standardized depression assessments, including the Hamilton Depression Rating Scale (HDRS), Beck Depression Inventory-II (BDI-II), and Patient Health Questionnaire-9 (PHQ-9). Facial-expression data were analyzed using iSeeME, while narrative transcripts were processed with EDDTW-V2. Predictive validity was assessed through accurate metrics, correlations with clinical scales, and k-means cluster analysis.

Of the 62 participants, 39 were clinically depressed (HDRS ≥7). The iSeeME system achieved precision of 0.761, recall of 0.854, F1-score of 0.805, and accuracy of 0.770, showing stronger correlations with BDI-II (r = 0.442, p < 0.01) and PHQ-9 (r = 0.335, p < 0.01) than EDDTW-V2. Cluster analysis revealed a high-distress subgroup characterized by younger age, single status, fewer children, and predominance of psychiatric patients.

Both AI systems showed potential as complementary tools to traditional assessments. iSeeME was more sensitive to affective-behavioral symptoms, while EDDTW-V2 captured cognitive-linguistic features. They support earlier detection, monitoring, and targeted interventions in outpatient settings.