This paper proposes a method for integrating Psychological State Tracking (POST) into a conversational system for diagnosing depression. Existing conversational systems for diagnosing depression have limitations, such as failing to adequately capture patients' evolving information, emotions, and symptoms, and lacking a clear framework for dialogue, leading to unnecessary conversations. In this paper, we design POST based on a psychological theoretical model consisting of four components: Stage, Information, Summary, and Next. We integrate this into a large-scale language model (LLM) to generate dynamic psychological states, and propose a system that utilizes these to guide response generation at each turn. Experimental results demonstrate that the proposed method improves the performance of all subtasks of the depression diagnosis conversation over existing benchmarks.
