ChatGPT in the diagnosis and management of complex polyneuropathies: comparative analysis with neurologists using real-world cases.

Abstract

Polyneuropathies are common and often require specialist expertise for accurate diagnosis. This study evaluated the diagnostic performance of ChatGPT-4o on real-world polyneuropathy cases, comparing it to peripheral neuropathy specialists and non-specialist neurologists. One hundred cases were selected from two tertiary centers in Milan, Italy. Standardized summaries included clinical, laboratory, and electrophysiological data. ChatGPT-4o was prompted to provide a leading diagnosis, two differentials, and a confirmatory test. Neurologists reviewed the same cases and generated comparable outputs, then could revise their responses after viewing ChatGPT-4o's suggestions. ChatGPT-4o achieved 65.5% leading diagnosis accuracy, comparable to non-specialists (63.0%) but lower than specialists (74.0%, p = 0.002). For differential diagnoses, it outperformed non-specialists (82.0% vs. 77.5%, p = 0.043) and recommended more appropriate tests (68.0% vs. 53.0%, p < 0.001). After reviewing ChatGPT-4o outputs, non-specialists revised their assessments in 21.8% of cases, improving accuracy. ChatGPT-4o shows potential as a diagnostic aid, particularly in non-specialist or resource-limited settings.