Development of a machine learning-based risk warning model for potentially inappropriate medication use among older patients in a hematology service.

Abstract

IntroductionOlder hematology patients often have multiple comorbidities and receive complex medication therapies; this can increase the risk of potentially inappropriate medication (PIM) use. A machine learning model could potentially serve as an automated screening and warning tool to facilitate the identification of PIM and support timely medication review in these patients. The aim of this study is to develop machine learning-based model that provide warnings regarding PIMs use in elderly patients hospitalized in the hematology service.MethodsThis retrospective pilot study was conducted on patients over 65 years of age who were admitted to the hematology service of a university hospital in Türkiye between 2018 and 2020. PIMs were determined according to the American Geriatrics Association's 2023 Beers Criteria. Python (v3.12.7) was used for data analysis in this study. The stratified 5-fold cross-validation method was used. To estimate PIM, seven different classification algorithms were used. Model performance was compared using accuracy, precision, recall, F1 score, receiver operating characteristics (ROC) curve, and area under the curve (AUC), brier score evaluation parameters.ResultsThis study included 190 patients. 45.8% of patients had PIM. The logistic regression model demonstrated the highest performance with 0.87 accuracy, 0.83 recall, 0.86 F1 score, and 0.91 AUC.ConclusionThe logistic regression model was determined to be the best-performing model. The developed patient-based automated screening model is promising and can be considered a reliable early warning mechanism for clinical pharmacists and physicians.