Large language models in clinical psychiatry: Applications and optimization strategies

Abstract

Psychiatric disorders constitute a complex health issue, primarily manifesting as significant disturbances in cognition, emotional regulation, and behavior. However, due to limited resources within health care systems, only a minority of patients can access effective treatment and care services, highlighting an urgent need for improvement. large language models (LLMs), with their natural language understanding and generation capabilities, are gradually penetrating the entire process of psychiatric diagnosis and treatment, including outpatient reception, diagnosis and therapy, clinical nursing, medication safety, and prognosis follow-up. They hold promise for improving the current severe shortage of health system resources and promoting equal access to mental health care. This article reviews the application scenarios and research progress of LLMs. It explores optimization methods for LLMs in psychiatry. Based on the research findings, we propose a clinical LLM for mental health using the Mixture of Experts framework to improve the accuracy of psychiatric diagnosis and therapeutic interventions.

Keywords: Large language models; Clinical psychiatry; Mixture of experts; Mental health; Research progress

Core Tip: This article comprehensively reviews the application scenarios and research advancements of large language models (LLMs) in psychiatry, ranging from outpatient reception, diagnosis and therapy, clinical nursing, medication safety, to prognosis tracking. It explores optimization methods for LLMs in psychiatry. These methods combine the techniques such as pre-training, supervised fine-tuning, retrieval-augmented generation, agent systems, and prompt engineering. Based on the research findings, we propose a clinical LLM for mental health using the Mixture of Experts framework. This approach addresses the shortcomings of single LLM system and aims to improve the accuracy of psychiatric diagnosis and therapeutic interventions.