Fiordelisi Marisa, Masucci Simona, Bianco Alessandra, Bellero Marco, Toma Diana, Campo Nicoletta, Zichi Clizia, Marino Donatella, Sperti Elisa, Valabrega Giorgio, Cena Clara, Fazzina Giovanna, Gasco Annalisa
Azienda Ospedaliera Ordine Mauriziano di Torino.
Università di Torino.
Recenti Prog Med. 2024 Nov;115(11):558-559. doi: 10.1701/4365.43601.
This study explores the potential use of ChatGPT, an AI-based language model, in assessing herbal-drug interactions (HDi) to enhance clinical decision-making. HDi can pose significant health risks by reducing drug efficacy or causing unwanted side effects. Clinical pharmacists play a key role in identifying these HDIs, and currently, there are limited tools available for checking drug interactions. The research focuses on a case study of a rectal adenocarcinoma patient treated with capecitabine and 26 supplements, which contain a total of 80 herbal substances. ChatGPT 3.5 was asked three questions regarding potential HDIs: "Are there possible HDIs?", "What is the pharmacokinetic mechanism?", and "What is the bibliographic source of the interaction?". The results were reviewed by an oncology clinical pharmacist and compared to existing databases and independent bibliographic research. The findings highlight ChatGPT's advantage in processing large amounts of data quickly, with 16% of interactions classified as "unlikely", confirmed by the pharmacist. However, 73% of the suggested mechanisms were false positives, and 4% were categorized as "hallucinations". Additionally, most of the bibliographic sources provided by ChatGPT were outdated or unavailable. While ChatGPT proves useful for initial HDI screening, its limitations include outdated data (last updated in January 2022), lack of access to private databases, and occasional inaccuracies. Further applications of AI in this area are recommended, though expert validation remains essential in the clinical decision-making process.
本研究探讨了基于人工智能的语言模型ChatGPT在评估草药-药物相互作用（HDI）以加强临床决策方面的潜在用途。HDI可能会通过降低药物疗效或引起不良副作用而带来重大健康风险。临床药剂师在识别这些HDI方面起着关键作用，而目前用于检查药物相互作用的工具有限。该研究聚焦于一名接受卡培他滨治疗的直肠腺癌患者以及26种补充剂的案例研究，这些补充剂共含有80种草药成分。针对潜在的HDI向ChatGPT 3.5提出了三个问题：“是否存在可能的HDI？”“药代动力学机制是什么？”以及“相互作用的文献来源是什么？”。结果由一名肿瘤临床药剂师进行审核，并与现有数据库及独立的文献研究进行比较。研究结果凸显了ChatGPT在快速处理大量数据方面的优势，药剂师确认其中16%的相互作用被归类为“不太可能”。然而，73%的建议机制为假阳性，4%被归类为“幻觉”。此外，ChatGPT提供的大多数文献来源过时或无法获取。虽然ChatGPT在初始HDI筛查中被证明是有用的，但其局限性包括数据过时（最后更新于2022年1月）、无法访问私人数据库以及偶尔出现的不准确情况。尽管在临床决策过程中专家验证仍然至关重要，但建议在该领域进一步应用人工智能。
