Construction of multimodal dataset for early depression detection and performance evaluation of depression detection model

Depression is a serious social issue, and early detection is crucial for improving mental health care. This study constructed a Japanese multimodal dataset for early depression detection, collecting text, audio, video, and heart rate data from 30-minute interviews with professional counselors. Psychological questionnaires were also administered before and after interviews. Feature extraction and correlation analysis with questionnaire results confirmed the dataset's reliability, aligning with existing psychological research. The dataset was then used to train DepMamba, a latest depression detection model, comparing three training methods. Fine-tuning the constructed dataset after pre-training on DAIC-WOZ significantly improved recall, enhancing depression detection performance. However, data limitations and label imbalances highlight the need for future dataset expansion and objective annotation. This research advances depression detection technology and supports the practical application of mental health care.