Multimodal deep learning for detecting current depressive symptom status using wearable time-series data and...

In high-risk occupational groups, current mental health status may be underreported because of stigma or concerns about disadvantage. Wearable data can capture short-term behavioral and physiological changes, but may be limited in capturing relatively stable biological characteristics. Structural brain magnetic resonance imaging (MRI) may reflect such individual differences and provide complementary information when combined with dynamic wearable signals. We evaluated whether a multimodal deep learning approach combining wearable-derived time-series data with structural brain MRI radiomics could help identify current depressive symptom status in firefighters and prosecution investigators. A total of 291 participants (184 firefighters and 107 prosecution investigators) were included, with a mean monitoring duration of 4.8 weeks. Elevated current depressive symptoms were defined as a Patient Health Questionnaire-9 (PHQ-9) score of 10 or higher at the final observation. Temporal deep learning models were evaluated under wearable-only and multimodal configurations. Multimodal models consistently outperformed wearable-only models, and the final multimodal LSTM model showed the best discriminative performance (AUROC = 0.867; 95% CI 0.814-0.914). In subgroup robustness analysis, performance was generally maintained across broader monitoring-duration categories and occupational groups. SHAP analysis identified the left paracentral lobule, right lateral ventricle, left amygdala, fragmented activity patterns, and minimum oxygen saturation during sleep as major contributors. These findings suggest that a multimodal approach integrating wearable-derived dynamic behavioral characteristics with relatively stable neuroanatomical information from structural brain MRI may serve as a supportive tool for more objectively identifying current depressive symptom status in high-risk occupational populations.