Skip to content
Open access

Artificial intelligence for patient safety: Pilot evaluation of LLMs for drug-drug interaction analysis

2026 · Zbornik radova · 0 citations

Abstract

Drug–drug interactions (DDIs) represent a significant challenge in the pharmacotherapy of patients with chronic non-communi cable diseases, including cardiovascular diseases and diabetes, due to polypharmacy and increased risk of clinically relevant inter actions. Effective approaches for early DDI identification are essential for improving patient safety. Artificial intelligence (AI)-based tools represent a potential approach for enhancing DDI identification and interpretation. This pilot study evaluated the potential of large language models (LLMs) for DDIs identification and clinical risk assessment. Four LLMs (ChatGPT, Gemini, Claude, and the locally deployed Qwen model), selected based on widespread use and accessibility, were evaluated using a modular framework integrating drug information from the ALIMS database and the RxList platform. A dual-prompting approach was applied, including patient-oriented (layperson) queries and structured pharmacist-oriented (expert) queries based on JSON-formatted drug informa tion. Six DDI pairs were analysed: glimepiride–ciprofloxacin, glimepiride–warfarin, metformin–ibuprofen, insulin aspart–predni sone, atorvastatin–clarithromycin, and semaglutide–insulin aspart. Model performance was assessed by comparing interaction identification and risk severity categorization with reference sources. All models identified clinically relevant interactions, although differences were observed in risk severity assessment. Gemini (r=0.91 for expert queries) and ChatGPT (r=0.86 for both query types) demonstrated the highest agreement with RxList assessments. The atorvastatin–clarithromycin combination represented the most challenging case, with only Qwen and expert Gemini correctly identifying the contraindication. Patient-oriented prompts generated more conservative risk assessments, whereas expert prompts enabled more precise differentiation of clinical relevance. These preliminary findings suggest that LLMs may serve as supportive tools for DDI assessment, although further validation using larger and more representative datasets is required.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.