Utilizing LLMs for Stroke Localization: The study explores the efficacy of LLMs like GPT-4 in localizing acute stroke lesions using clinical data.
Methodological Approach: GPT-4 is prompted with questions derived from history and neurologic examination, then compared with imaging-based localization across multiple trials.
High Performance Metrics: GPT-4 exhibits strong performance in specificity, sensitivity, precision, and F1-score, particularly in identifying stroke side and brain region.
Accurate Clinical Reasoning: The model demonstrates the ability to generate accurate neuroanatomical localization and detailed clinical reasoning from raw text data.
Limitations and Implications: Errors stem from incomplete data and logic failures, suggesting the need for further refinement despite promising potential in assisting with stroke diagnosis.