Benchmarking large language models against practicing clinicians on psychopathological assessment.

Abstract

Psychiatry's reliance on language makes LLMs a natural tool for psychopathological assessment, yet structured, item-level assessments from psychiatric clinical interviews remain under-researched. In this proof-of-concept study, 10 LLMs assessed transcripts of three simulated psychiatric interviews across all 100 items of the Association for Methodology and Documentation in Psychiatry (AMDP) system, benchmarked against 108 early-career clinicians rating full audiovisual recordings, using an expert consensus panel as reference. GPT-5.1 and Gemini-3-Pro-Preview achieved the highest accuracy (0.72; 64th percentile of the clinician distribution) using majority voting across three runs with AMDP definitions as context. GPT-5.1, selected for a marginal advantage, showed per-scenario accuracies of 0.81 (depression), 0.76 (mania), and 0.60 (schizophrenia) versus clinician means of 0.79, 0.68, and 0.58. Clinicians and LLMs showed distinct error profiles: clinicians tended to over-infer symptom presence, whereas LLMs more conservatively flagged items as "not assessable" - most pronounced for observation-dependent items but present even for text-assessable items (19.4% vs. 11.4%, p < 0.001). In post hoc simulated disagreement resolutions (2091 clinician pairs; 35.5% disagreements), LLM and board-certified supervision were associated with more accurate resolutions than unsupervised random clinician selection (p < 0.0002). These proof-of-concept findings require validation in real patient interviews, larger samples, and prospective studies integrating multimodal input.