MentalHospital: A Virtual Environment for Evaluating Psychiatric Clinical Encounters

Research protocol: Evaluating Chain-of-Thought Reasoning in LLMs for Complex Clinical Psychiatric Cases

Background Psychiatric medicine presents unique diagnostic and therapeutic challenges, often involving multimorbidity, polypharmacy, and atypical presentations requiring complex reasoning. Artificial Intelligence, particularly Large Language Models (LLMs), is emerging as a support tool in these settings. However, the clinical validity, interpretability, and reliability of LLMs in psychiatry remain largely unexplored, particularly their ability to generate transparent, guideline-consistent reasoning. Aims This study evaluates the clinical reasoning capabilities of LLMs in complex psychiatric scenarios. The primary aim is to assess the validity of their long chain-of-thought (CoT) reasoning. A secondary aim is to determine whether LLM assistance improves clinicians’ diagnostic accuracy. Method Three LLMs, Gemini 2.5, Grok, and DeepSeek R1, will be assessed using ten complex psychiatric cases sourced from a non-public clinical manual and rated with the Amsterdam Clinical Challenge Scale. Each model's CoT response will be evaluated by blinded panels of psychiatrists, residents, and general practitioners using standardized metrics for factual accuracy, coherence, and medical plausibility. In a second phase, clinicians will answer thirty diagnostic questions with and without support from the best-performing LLM. The study uses step-by-step reasoning prompts and few-shot examples to elicit detailed responses and includes bias mitigation strategies such as randomization, blinding, and statistical controls. Results Analyses will assess inter-rater reliability, metric redundancy, and reasoning quality. Closed-source models are expected to outperform open-source ones. LLM assistance is anticipated to improve diagnostic accuracy, especially among non-specialists. Conclusions This study provides a framework for evaluating LLMs in psychiatry and supporting their safe, evidence-based integration into mental health practice.

[Virtual Mental Health Assistance: Online Platform for Psychiatric Assessment and Follow-Up].

Large language model applications for real-time clinical mental health assessment: Current potential and future directions.

Large language models are best understood as emerging assessment-support tools rather than replacements for clinical evaluation because the limited pace of academic validation means that, at present, LLMs are best understood as emerging assessment-support tools rather than replacements for clinical evaluation.

A Real-World Evaluation of Large Language Model-Generated Hospital Courses in Pediatrics.

BACKGROUND Large language model (LLM)-generated hospital courses are increasingly integrated into electronic health records (EHRs), yet their accuracy and safety in pediatric populations remain poorly characterized. OBJECTIVE To evaluate the accuracy, text quality, and perceived potential harm of EHR-integrated and LLM-generated hospital courses in pediatric inpatient care during early clinical implementation. METHODS We conducted a descriptive evaluation from June 10 to August 8, 2025, at an academic freestanding children's hospital using an Epic EHR with an integrated LLM tool (GPT-4o and GPT-4.1). Clinicians across multiple roles, including attending physicians, residents, and advanced practice providers, reviewed LLM-generated hospital courses for their own patients. Clinicians identified and categorized errors (hallucinations, inaccuracies, or omissions). They also rated text quality (comprehensiveness, conciseness, coherence) on a 5-point scale and perceived harm on an 8-point scale. RESULTS A total of 129 LLM-generated hospital courses were reviewed (median length of stay, 3 days; IQR, 2-7) by 50 involved clinicians. Hallucinations occurred in 21% (95% CI, 14%-29%) of the hospital courses, inaccuracies in 41% (53/129; 95% CI, 33%-50%), and omissions in 24% (31/129; 95% CI, 17%-32%). Overall, perceived harm ratings were low (median, 0; IQR, 0-1). Text quality ratings were high (median [IQR]: comprehensiveness, 4 [3-5]; conciseness, 4 [4-5]; coherence, 4 [4-5]) and comparable with prior literature. CONCLUSION In this pediatric evaluation of LLM-generated hospital courses reviewed by frontline clinicians, errors were common, but perceived potential harm was low, even assuming use without clinician correction. These findings support the use of LLM-generated hospital courses as starting drafts when paired with clinician review and institutional safeguards.

Large Language Models for Differential Diagnosis: A Survey of Performance, Collaboration, and Technical Strategies

Current evidence supports clinician-supervised use of artificial intelligence (AI) systems rather than autonomous diagnosis, pending prospective and specialty-specific evaluation.

Performance of a domain-specific large language model in answering patient questions in psychiatry

MIND was able to answer many questions about escitalopram in a manner deemed accurate, complete, and safe by psychiatrists the majority of the time, and represents a step towards building safe LLM systems to enhance patient education in psychiatry.