Unveiling the Potential and Pitfalls: A Critical Look at Chat GPT’s Role in Drug-Drug Interaction Analysis
The intricate process of analyzing potential drug-drug interactions (pDDIs) is a critical, yet time-consuming, aspect of ensuring patient safety in pharmacotherapy. Healthcare professionals currently rely on multiple clinical decision support systems (CDSSs), which, despite their utility, come with known limitations such as alert fatigue and variability in results. The emergence of large language models (LLMs) developed with artificial intelligence (AI), such as Chat Generative Pre-Trained Transformer (Chat GPT), has sparked considerable interest as a potential alternative to streamline this complex process.
A recent retrospective study aimed to assess the practical utility of Chat GPT in identifying pDDIs compared to established CDSSs (MediQ®, Lexicomp®, Micromedex®). The study involved 30 patients with polypharmacy, and the results were rigorously interpreted by a multidisciplinary team of healthcare professionals for clinical relevance and severity grading.
Key Insights from the Study:
- Significant Disparity in Detection: The expert review, based on established CDSSs, identified a total of 280 clinically relevant pDDIs, including 3 contraindications, 13 severe, and 264 moderate interactions. In stark contrast, the Chat GPT approach resulted in the identification of only 80 pDDIs (2 contraindications, 5 severe, 73 moderate). This difference was statistically significant.
- Critical Safety Omissions: A major concern highlighted by the study was Chat GPT’s almost complete neglect of pDDIs with the risk of QTc prolongation. While the standard method identified 85 such interactions, Chat GPT found only 8 initially, which only slightly improved to 19 even with specific prompts. This is highly problematic given the severe clinical consequences associated with QTc prolongation, including increased risks of ventricular arrhythmias and sudden cardiac death.
- Inconsistency and Errors:
- Upon repeating queries, Chat GPT demonstrated inconsistent results in 90% of cases. This variability is a significant quality control issue, attributed to the probabilistic design of LLMs compared to the fixed, deterministic rules of specialized CDSSs.
- The model also produced obvious errors, such as incorrectly associating pharmacokinetic pDDIs (e.g., CYP3A4 metabolism) with entire drug classes or suggesting non-existent absorption interactions.
- Promising Capabilities: Despite these critical shortcomings in pDDI identification, Chat GPT showed remarkable strengths in other areas:
- It provided acceptable and comprehensible recommendations for specific questions on side effects.
- Its responses were consistently well-structured and easy to understand, often more comprehensive than those provided by physicians in other comparative studies.
- Chat GPT’s user-friendliness and ability to rapidly generate concise overviews for pharmacotherapy are notable advantages over the cumbersome and time-consuming nature of traditional CDSS analysis.
Conclusion and Future Outlook:
The study unequivocally concludes that the use of Chat GPT for the identification of pDDIs cannot be recommended currently. This is primarily due to its failure to detect clinically relevant pDDIs, the presence of obvious errors, and inconsistencies in its results. However, the authors emphasize that Chat GPT remains a promising platform for the future if these significant limitations are addressed. Its potential to overcome some existing weaknesses of CDSSs and its user-friendly interface suggest that, with further development and rigorous validation, AI models like Chat GPT could still play a transformative role in clinical pharmacology and patient care.
Reference for this article: Bischof, T., al Jalali, V., Zeitlinger, M., Jorda, A., Hana, M., Singeorzan, K.-N., Riesenhuber, N., Stemer, G., & Schoergenhofer, C. (2025). Chat GPT vs. Clinical Decision Support Systems in the Analysis of Drug–Drug Interactions. Clinical Pharmacology & Therapeutics. doi:10.1002/cpt.3585
