SemioLLM: Evaluating Large Language Models for Diagnostic Reasoning from Unstructured Clinical Narratives in Epilepsy

Large Language Models (LLMs) have been shown to encode clinical knowledge. Many evaluations, however, rely on structured question-answer benchmarks, overlooking critical challenges of interpreting and reasoning about unstructured clinical narratives in real-world settings. In this study we task eight Large Language models including two medical models (GPT-3.5, GPT-4, Mixtral-8 × 7B, Qwen-72B, LlaMa2, LlaMa3, OpenBioLLM, Med42) with a core diagnostic task in epilepsy: mapping seizure description phrases-after targeted filtering and standardization-to one of seven possible seizure onset zones using likelihood estimates. We conduct quantitative and qualitative analyses, measuring correctness, confidence, calibration, and expert-evaluated reasoning quality and source citation accuracy. Through systematic prompt-engineering and ablation studies, we assess how model performance depends on variations in prompt strategy, clinical role impersonation, narrative length, and language context. Most models yield well-above chance accuracy after prompt engineering that even approaches clinician-level performance. Specifically, clinician-guided chain-of-thought reasoning leads to the most consistent improvements. Performance is further strongly modulated by clinical in-context impersonation, narrative length and language context (13.7%, 32.7% and 14.2% performance variation, respectively). However, reasoning analysis by clinical experts reveal that correct prediction can be based on hallucinated knowledge and inaccurate source citation, underscoring the need to improve interpretability of LLMs in clinical use. Overall, SemioLLM provides a scalable, domain-adaptable framework for evaluating LLMs in clinical disciplines where unstructured verbal descriptions encode diagnostic information. By identifying both the strengths and limitations of LLMs, our work contributes to testing the applicability of foundational AI systems for healthcare. Large language models (LLMs) are increasingly used in healthcare, but their reliability with real patient descriptions remains unclear. This study developed a framework to evaluate eight LLMs on a core epilepsy diagnostic task: determining which brain region causes seizures based on patient descriptions, crucial for treatment management. We assess correctness, confidence, and reasoning quality across 1200 seizure descriptions. Most models performed well above chance, sometimes approaching clinician-level accuracy. Performance was strongly influenced by clinician-written reasoning examples, clinical role context, description length, and language. However, some correct answers relied on fabricated facts and flawed reasoning, raising safety concerns. These findings reveal both potential and limitations of LLMs in healthcare, highlighting the need for careful validation before clinical deployment. Dani et al. investigate the potential and limitation of large language models in a core clinical task in epilepsy care: mapping seizure descriptions to seizure onset zones. Their findings show that prompt engineering, including clinician-guided instruction, impersonation, symptom length, and language, all affect model performance and reasoning.