College of Pharmacy & Research Institute of Pharmaceutical Sciences, Seoul National University, Seoul, Republic of Korea; Department of Pharmacy, Seoul National University Bundang Hospital, Gyeonggi-do, Republic of Korea.
Office of eHealth Research and Businesses, Seoul National University Bundang Hospital, Gyeonggi-do, Republic of Korea.
Int J Med Inform. 2024 May;185:105398. doi: 10.1016/j.ijmedinf.2024.105398. Epub 2024 Feb 29.
Drug-related problems (DRPs) are a significant concern in healthcare. Pharmacists play a vital role in detecting and resolving DRPs to improve patient safety. A pharmacy inquiry program was established in a tertiary teaching hospital to document inquiries about physicians' orders, aimed at preventing potential DRPs or providing medication information during order reviews.
We aimed to develop machine-learning models using a pharmacy inquiry database to predict dose-related inquiries based on prescriptions and patient information.
This retrospective study analyzed 20,393 pharmacy inquiries collected between January 2018 and February 2023. Data included prescription information (drug ingredient, dose, unit, and frequency), patient characteristics (age, sex, weight, and department), and renal function. The inquiries were categorized into two classes: dose-related inquiries (e.g., wrong dose and inappropriate regimen) and non-dose-related inquiries (e.g., inappropriate drug form and administration route). Six machine-learning models were developed: logistic regression, support vector classifier, decision tree, random forest, extreme gradient boosting, and categorical boosting. To evaluate the performance of the models, the area under the receiver operating characteristic curve and the accuracy were compared.
The CatBoost model achieved the highest performance (sensitivity: 0.92; accuracy: 0.79). The SHapley Additive exPlanations values highlighted the importance of features in the model predictions, drug ingredients, units, and renal function, in that order. Notably, lower renal function positively contributed to the prediction of dose-related inquiries. Additionally, the subsequent feature importance among drug ingredients showed that drugs such as acetylsalicylic acid, famotidine, metformin, and spironolactone strongly influenced the prediction of dose-related inquiries.
Machine-learning models that use pharmacy inquiry data can effectively predict dose-related inquiries. Further external validation and refinement of the models are required for broader applications in healthcare settings. These findings provide valuable guidance for healthcare professionals and highlight the potential of machine learning in pharmacists' decision-making.
药物相关问题（DRPs）是医疗保健中的一个重要关注点。药剂师在发现和解决 DRPs 以提高患者安全性方面发挥着至关重要的作用。一家三级教学医院建立了一个药房查询计划，记录对医生医嘱的查询，旨在防止潜在的 DRPs 或在医嘱审查期间提供药物信息。
我们旨在使用药房查询数据库开发机器学习模型，根据处方和患者信息预测与剂量相关的查询。
这项回顾性研究分析了 20393 次药房查询，这些查询是在 2018 年 1 月至 2023 年 2 月期间收集的。数据包括处方信息（药物成分、剂量、单位和频率）、患者特征（年龄、性别、体重和科室）和肾功能。查询分为两类：与剂量相关的查询（例如剂量错误和方案不当）和非剂量相关的查询（例如药物剂型和给药途径不当）。开发了六种机器学习模型：逻辑回归、支持向量分类器、决策树、随机森林、极端梯度提升和分类梯度提升。为了评估模型的性能，比较了接收者操作特征曲线下的面积和准确性。
CatBoost 模型表现最佳（敏感性：0.92；准确性：0.79）。Shapley 加法解释值突出了特征在模型预测中的重要性，按重要性顺序依次为药物成分、单位和肾功能。值得注意的是，较低的肾功能对与剂量相关的查询的预测有积极贡献。此外，药物成分之间的后续特征重要性表明，乙酰水杨酸、法莫替丁、二甲双胍和螺内酯等药物强烈影响与剂量相关的查询的预测。
使用药房查询数据的机器学习模型可以有效地预测与剂量相关的查询。需要进一步进行外部验证和模型改进，以便在医疗保健环境中更广泛地应用。这些发现为医疗保健专业人员提供了有价值的指导，并强调了机器学习在药剂师决策中的潜力。
