Unlocking the potential of advanced large language models in medication review and reconciliation: A proof-of-concept investigation

Abstract

Background: Medication review and reconciliation is essential for optimizing drug therapy and minimizing medication errors. Large language models (LLMs) have been recently shown to possess a lot of potential applications in healthcare field due to their abilities of deductive, abductive, and logical reasoning. The present study assessed the abilities of LLMs in medication review and medication reconciliation processes. Methods: Four LLMs were prompted with appropriate queries related to dosing regimen errors, drug-drug interactions, therapeutic drug monitoring, and genomics-based decision-making process. The veracity of the LLM outputs were verified from validated sources using pre-validated criteria (accuracy, relevancy, risk management, hallucination mitigation, and citations and guidelines). The impacts of the erroneous responses on the patients' safety were categorized either as major or minor. Results: In the assessment of four LLMs regarding dosing regimen errors, drug-drug interactions, and suggestions for dosing regimen adjustments based on therapeutic drug monitoring and genomics-based individualization of drug therapy, responses were generally consistent across prompts with no clear pattern in response quality among the LLMs. For identification of dosage regimen errors, ChatGPT performed well overall, except for the query related to simvastatin. In terms of potential drug-drug interactions, all LLMs recognized interactions with warfarin but missed the interaction between metoprolol and verapamil. Regarding dosage modifications based on therapeutic drug monitoring, Claude-Instant provided appropriate suggestions for two scenarios and nearly appropriate suggestions for the other two. Similarly, for genomics-based decision-making, Claude-Instant offered satisfactory responses for four scenarios, followed by Gemini for three. Notably, Gemini stood out by providing references to guidelines or citations even without prompting, demonstrating a commitment to accuracy and reliability in its responses. Minor impacts were noted in identifying appropriate dosing regimens and therapeutic drug monitoring, while major impacts were found in identifying drug interactions and making pharmacogenomic-based therapeutic decisions. Conclusion: Advanced LLMs hold significant promise in revolutionizing the medication review and reconciliation process in healthcare. Diverse impacts on patient safety were observed. Integrating and validating LLMs within electronic health records and prescription systems is essential to harness their full potential and enhance patient safety and care quality.