Performance of large language models on advocating the management of meningitis: a comparative qualitative stud

Abstract

Objectives We aimed to examine the adherence of large language models (LLMs) to bacterial meningitis guidelines using a hypothetical medical case, highlighting their utility and limitations in healthcare. Methods A simulated clinical scenario of a patient with bacterial meningitis secondary to mastoiditis was presented in three independent sessions to seven publicly accessible LLMs (Bard, Bing, Claude-2, GTP-3.5, GTP-4, Llama, PaLM). Responses were evaluated for adherence to good clinical practice and two international meningitis guidelines. Results A central nervous system infection was identified in 90% of LLM sessions. All recommended imaging, while 81% suggested lumbar puncture. Blood cultures and specific mastoiditis work-up were proposed in only 62% and 38% sessions, respectively. Only 38% of sessions provided the correct empirical antibiotic treatment, while antiviral treatment and dexamethasone were advised in 33% and 24%, respectively. Misleading statements were generated in 52%. No significant correlation was found between LLMs’ text length and performance (r=0.29, p=0.20). Among all LLMs, GTP-4 demonstrated the best performance. Discussion Latest LLMs provide valuable advice on differential diagnosis and diagnostic procedures but significantly vary in treatment-specific information for bacterial meningitis when introduced to a realistic clinical scenario. Misleading statements were common, with performance differences attributed to each LLM’s unique algorithm rather than output length. Conclusions Users must be aware of such limitations and performance variability when considering LLMs as a support tool for medical decision-making. Further research is needed to refine these models' comprehension of complex medical scenarios and their ability to provide reliable information. Data are available upon reasonable request.