Diagnostic accuracy of a large language model in rheumatology: comparison of physician and ChatGPT-4

Abstract

AbstractPre-clinical studies suggest that large language models (i.e., ChatGPT) could be used in the diagnostic process to distinguish inflammatory rheumatic (IRD) from other diseases. We therefore aimed to assess the diagnostic accuracy of ChatGPT-4 in comparison to rheumatologists. For the analysis, the data set of Gräf et al. (2022) was used. Previous patient assessments were analyzed using ChatGPT-4 and compared to rheumatologists’ assessments. ChatGPT-4 listed the correct diagnosis comparable often to rheumatologists as the top diagnosis 35% vs 39% (p = 0.30); as well as among the top 3 diagnoses, 60% vs 55%, (p = 0.38). In IRD-positive cases, ChatGPT-4 provided the top diagnosis in 71% vs 62% in the rheumatologists’ analysis. Correct diagnosis was among the top 3 in 86% (ChatGPT-4) vs 74% (rheumatologists). In non-IRD cases, ChatGPT-4 provided the correct top diagnosis in 15% vs 27% in the rheumatologists’ analysis. Correct diagnosis was among the top 3 in non-IRD cases in 46% of the ChatGPT-4 group vs 45% in the rheumatologists group. If only the first suggestion for diagnosis was considered, ChatGPT-4 correctly classified 58% of cases as IRD compared to 56% of the rheumatologists (p = 0.52). ChatGPT-4 showed a slightly higher accuracy for the top 3 overall diagnoses compared to rheumatologist’s assessment. ChatGPT-4 was able to provide the correct differential diagnosis in a relevant number of cases and achieved better sensitivity to detect IRDs than rheumatologist, at the cost of lower specificity. The pilot results highlight the potential of this new technology as a triage tool for the diagnosis of IRD.