PKFAR: psychiatry knowledge-fused augmented reasoning with large language models.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Computational Psychiatry

Background:

• Psychiatric diagnosis is challenged by subjective reporting and complex criteria.
• Current Large Language Models (LLMs) face privacy or computational hurdles for clinical use.
• Existing LLM approaches inadequately address psychiatric complexities.

Purpose of the Study:

• To develop a specialized, knowledge-enhanced LLM approach for accurate psychiatric diagnostics.
• To overcome limitations of commercial and large-scale open-source LLMs in psychiatric applications.
• To improve clinical decision support in psychiatry through advanced AI.

Main Methods:

• Proposed PKFAR (psychiatry knowledge-fused augmented reasoning) system.
• Developed PsychKG: a semantically-augmented psychiatric knowledge graph.
• Implemented a three-stage hierarchical reasoning framework: symptom comprehension, disorder retrieval, and diagnosis reasoning.
• Evaluated on Mentat, MedQA_psychiatry, and MIMIC benchmarks using Qwen3-8B.

Main Results:

• PKFAR achieved 12.4%, 7.5%, and 10.0% accuracy improvements over standard baselines on Mentat, MedQA_psychiatry, and MIMIC, respectively.
• Demonstrated superior performance compared to one-shot CoT of GPT-o3 and DeepSeek-V3.
• Approached the accuracy of the large-scale DeepSeek-R1 model.

Conclusions:

• PKFAR offers an effective balance between computational efficiency and diagnostic precision.
• The knowledge-fused approach and structured reasoning address critical LLM limitations in psychiatric diagnostics.
• PKFAR presents a practical solution for enhancing psychiatric diagnostic accuracy using AI.