Drug interactions in rheumatoid arthritis: a disparity between electronic database prediction and real-world co-prescription with different large language models
Tonson Lalitkanjanakul, Dhanesh Pitidhammabhorn, Natnicha Jakramonpreeya, Werapat Suechanyapong, Supawit Tangpanithandee, Phisit KhemawootBackground:
Rheumatoid arthritis (RA) is a chronic inflammatory disease that often requires multi-drug regimens, which can increase the risk of drug–drug interactions (DDIs).
Objective:
To determine whether polypharmacy contributes to potential DDIs in patients with RA.
Design:
A retrospective observational study that integrates data from drug interaction databases and medical records, supplemented by artificial intelligence analysis.
Methods:
Twenty-one RA drugs and 54 commonly prescribed medicines were screened for potential DDIs using the Micromedex and WebMD databases and compared with real-world co-prescription at our hospital. The ChatGPT-5 and Gemini 2.5 large language models (LLMs) were also evaluated as “AI judges” versus an expert rheumatologist to assess real-world co-prescription. This study was conducted from January to September 2025.
Results:
Among 75 frequently prescribed medicines in RA patients, 753 database-predicted potential DDIs were reported, with Micromedex showing 464 DDI pairs and WebMD showing 600 pairs. Both databases classified 47 serious potential DDIs as contraindicated or major, with high concordance. The main predicted DDI mechanisms were immunosuppression (27%), methotrexate toxicity (18%), and nephrotoxicity (12%), with tacrolimus (12%), methotrexate (10%), and infliximab (6%) as major culprit drugs. In real-world co-prescriptions, 259 RA patients showed methotrexate/naproxen (46.4%) and methotrexate/sulfasalazine (29.5%) as the most common real-world co-prescribing patterns with potential DDIs. Brennan–Prediger kappa values indicated fair agreement (0.3161) in the electronic database predictions and conditional agreement (0.9807) among preselected serious DDIs. The LLM evaluation revealed that advanced models (e.g., Gemini 2.5 Pro) provided finesse recommendations highly concordant with expert opinion, whereas simpler models were more conservatively aligned with database alerts.
Conclusion:
In both real-world co-prescription and database evaluation, serious potential DDIs were associated mainly with methotrexate toxicity. Based on these exploratory findings, advanced LLMs show potential to assist with the contextualizing rigid database alerts and supporting pragmatic clinical decision-making. However, expert clinical judgment and direct patient consultation remain essential to improve RA patient safety and medication adherence.