Automated Symptom Identification from Clinical Interview Transcripts in Diagnosed Schizophrenia: A Multi-Agent LLM...

We present a multi-agent LLM framework for automated evaluation of symptom-criteria consistency in diagnosed schizophrenia cases. Our system decomposes evaluation into seven specialized agents with RAG-augmented DSM-5/ICD-11 knowledge: symptom extraction, standard matching, differential diagnosis, temporal analysis, counterfactual reasoning, voting, and arbitration. Evaluated on 71 diagnosed cases, our framework achieves 95.8% PPA (Positive Percent Agreement), outperforming single-agent baseline (70.4%) by 25.4% and single-agent+RAG (87.3%) by 8.5%. Because the dataset is case-only, results reflect case detection rather than screening accuracy; specificity is not estimable. Ablation reveals RAG contributes +16.9% and multi-agent specialization adds +8.5%, demonstrating both are critical for psychiatric symptom-criteria evaluation. Overall, our seven-agent RAG framework structures and traces interview-based symptom-to-criterion assessment and offers a scalable automated prototype; clinical use still requires multi-center validation with controls and first-episode cases to confirm specificity and external generalizability.