Can ChatGPT perform as well as pharmacists in identifying potentially inappropriate medications (PIMs) in elderly cancer patients?

e23121 Background: Polypharmacy remains a significant issue among older adults, primarily due to its association with an increased risk of potentially inappropriate medications (PIMs). Using PIMs in this population is associated with adverse drug events, heightened healthcare costs, and poorer clinical outcomes. ChatGPT, a sophisticated language model developed by OpenAI based on the Generative Pre-trained Transformer (GPT) architecture, has demonstrated potential utility in a variety of healthcare applications. Despite these advancements, the comparative effectiveness of ChatGPT in identifying PIMs and its efficiency relative to clinical pharmacists remains under-explored. This study sought to evaluate the capability of ChatGPT to detect PIMs in elderly cancer patients and to compare its time efficiency with that of pharmacists in the drug reconciliation process. Methods: A cross-sectional study was conducted involving elderly cancer patients receiving treatment at the outpatient cancer clinic of Ramathibodi Hospital. PIMs were identified and categorized based on the American Geriatrics Society (AGS) 2023 Updated Beers Criteria for Potentially Inappropriate Medication Use in Older Adults. Comprehensive demographic and clinical data, including diagnoses and prescribed medications, were input into ChatGPT version 4.0 using standardized prompts. The results generated by ChatGPT were subsequently compared with independent evaluations performed by two clinical pharmacists. The time required for medication reconciliation by ChatGPT and the pharmacists was measured and analyzed. All statistical analyses were conducted using Stata version 18. Results: The study included a cohort of 200 patients, with a median of 6.08 medications per patient (range: 0–22). Polypharmacy was identified in 55.0% of participants, with these patients exhibiting a significantly higher prevalence of PIMs compared to those prescribed fewer than 5 medications (P < 0.001). Pharmacists identified PIMs in 29.5% of cases, with the most frequently identified PIMs being benzodiazepines, metoclopramide, and proton pump inhibitors. The agreement between ChatGPT and the pharmacists was limited, with a Pearson correlation coefficient of r = +0.43. In terms of time efficiency, pharmacists required an average of 3.04 minutes to complete the drug reconciliation process, while ChatGPT completed the same task in 2.09 minutes, significant difference (P < 0.001). Conclusions: ChatGPT shows potential as an artificial intelligence tool for enhancing medication reconciliation processes in healthcare settings. While it offers a time-efficient solution for identifying PIMs, the level of agreement with clinical pharmacists remains moderate. Further refinement and validation of this technology are necessary to ensure its reliability and safety before broader clinical implementation.