This paper proposes an interactive depression detection framework (PDIMC) for early diagnosis of depression. Existing depression detection studies utilize multilayer neural network models to capture the hierarchical structure of clinical interview conversations, but they are limited by their inability to explicitly model inter- and intra-topic correlations and their inability to allow clinician intervention. PDIMC utilizes contextual learning techniques to identify topics in clinical interviews and model inter- and intra-topic correlations. Furthermore, it provides an interactive feature that allows clinicians to adjust topic importance based on their interests through AI-based feedback. On the DAIC-WOZ dataset, it achieves performance improvements of 35% and 12%, respectively, compared to the previous best-performing model, demonstrating the effectiveness of integrating topic correlation modeling and interactive external feedback.
