Evaluating the diagnostic performance of OpenBioLLM in neurology: A case-based assessment of a medical large language model

In the evolving field of neurological healthcare, deep learning technologies are gaining recognition for their potential to enhance diagnostic accuracy. Transformer-based models, particularly large language models (LLMs) such as OpenBioLLM, have shown promise in processing large datasets typical of neurological assessments. This study evaluates the diagnostic capabilities of OpenBioLLM in the realm of neurological conditions. The primary aim of this research is to assess the diagnostic accuracy, comprehensiveness, supplementation, and fluency of OpenBioLLM when applied to complex neurological case studies. Twenty-five complex neurology cases were selected from "Clinical Cases in Neurology." OpenBioLLM was used to generate diagnoses and rationales for each case. Two independent medical doctors evaluated the responses based on accuracy, comprehensiveness, supplementation, and fluency, with discrepancies resolved by a third assessor. Statistical analyses included one-way ANOVA, Bartlett's test, and Spearman's rank correlation. OpenBioLLM achieved a mean accuracy score of 38%, a comprehensiveness score of 52%, a supplementation score of 24%, and a fluency score of 100%. The model could localize neurological lesions but often struggled with identifying the correct pathophysiological causes. Accuracy scores did not significantly vary by neurological disorder type. While OpenBioLLM shows potential in diagnosing neurological conditions, its performance metrics suggest it is not yet a reliable standalone tool. Future research should focus on fine-tuning the model and improving its reasoning capabilities to enhance diagnostic accuracy.