Inpatients newly diagnosed with rheumatoid arthritis in our rheumatology department between October 1, 2025, and June 2026 will be recruited for the study. Physicians will use GenAI predictions of potential RA complications and co-existing diseases, together with confirmatory diagnostic tests, as additional inputs in the differential diagnosis process.

Generative AI based on multiple large language models (LLMs) is used to predict potential complications and co-existing diseases in patients with rheumatoid arthritis using EHR data available at admission. Physicians use these AI predictions as additional information to adjust their diagnostic plans during differential diagnosis. The impact of this intervention on the final diagnoses at discharge will be measured.

Before the prospective study, the adoptability of the generative AI prediction reports will be validated using EHR records from retrospective RA patients.

In the routine care workflow, large language models (LLMs) are used to predict potential RA complications for each de-identified patient case and generate an AI report listing possible complications and co-existing diseases. Additional diagnostic tests are suggested to verify the predicted conditions. After reviewing the AI report, physicians immediately evaluate each disease prediction using a 5-point Likert scale (1 = complete disagreement; 2 = disagreement; 3 = neutral; 4 = agreement; 5 = complete agreement). The mean score is calculated as a measure of perceived prediction accuracy. Physicians also indicate whether each specific disease prediction could potentially be adopted or used to assist differential diagnosis (binary: 0 or 1). The percentage of positive adoption responses is calculated as a measure of potential adoption rate, or adoptability.

Before patient discharge, physicians make final diagnoses and record which diagnosed complications or co-existing diseases were influenced by GenAI prediction information for each patient. The percentage of cases in which GenAI predictions affected the final diagnosis is calculated as a measure of AI's actual impact on routine diagnostic practice.