📝 SummarXiv

Abstract

Background: Large language models (LLMs) can be useful in diagnosing medical conditions, but few studies have evaluated their consistency in recommending appropriate medication regimens. The purpose of this evaluation was to test GPT-4o on three medication benchmarking tests including mapping a drug name to its correct formulation, identifying drug-drug interactions using both its internal knowledge and using a web search, and preparing a medication order sentence after being given the medication name. Methods: Using GTP-4o three experiments were completed. Accuracy was quantified by computing cosine similarity on TF-IDF vectors, normalized Levenshtein similarity, and ROUGE-1/ROUGE-L F1 between each response and its reference string or by manual evaluation by clinicians. Results: GPT-4o performed poorly on drug-formulation matching, with frequent omissions of available drug formulations (mean 1.23 per medication) and hallucinations of formulations that do not exist (mean 1.14 per medication). Only 49% of tested medications were correctly matched to all available formulations. Accuracy was decreased for medications with more formulations (p<0.0001). GPT-4o was also inconsistent at identifying drug-drug-interactions, although it had better performance with the search-augmented assessment compared to its internal knowledge (54.7% vs. 69.2%, p=0.013). However, allowing a web-search worsened performance when there was no drug-drug interaction (median % correct 100% vs. 40%, p<0.001). Finally, GPT-4o performed moderately with preparing a medication order sentence, with only 65.8% of medication order sentences containing no medication or abbreviation errors. Conclusions: Model performance was overall poor for all tests. This highlights the need for domain-specific training through clinician-annotated datasets and a comprehensive evaluation framework for benchmarking performance.