Development and validation of a machine learning-based clinical decision support tool for stratifying intravenous medication risk in hospitalized patients with heart failure.

INTRODUCTION: Hospitalized patients with heart failure (HF) frequently receive multiple high-risk intravenous (IV) medications, placing them at a substantial risk of clinically significant drug-related problems (DRPs). Timely severity-based stratification of IV medication-related risks remains challenging, particularly in the context of complex regimens and organ dysfunction. AIM: To develop and externally validate a clinically applicable machine learning-based stratification tool for classifying IV medication-related risk severity in hospitalized patients with HF and to support pharmacist-led medication safety management through a web-based clinical decision support tool. METHOD: This multicenter retrospective study included 1,884 adult patients hospitalized with HF from seven tertiary hospitals, with an independent external validation cohort of 100 patients. IV medication-related DRPs were identified and classified by senior clinical pharmacists using the Pharmaceutical Care Network Europe (PCNE) DRP classification (version 9.1) and stratified by severity using the National Coordinating Council for Medication Error Reporting and Prevention (NCC MERP) index. Medication risk severity (levels C-F) was used as a multiclass stratification outcome. Candidate clinical and medication-related variables reflecting patient characteristics, organ function, and medication burden were selected using least absolute shrinkage and selection operator (LASSO) regression, and six machine learning algorithms for multiclass risk stratification were developed and compared. RESULTS: Overall, 1405 patients (74.6%) experienced at least one IV medication-related DRP, with treatment safety problems predominating. Eleven clinically interpretable predictors, including neutrophil percentage, fibrinogen, serum albumin, and creatinine clearance were retained in the final models. In internal validation, the random forest (RF) model achieved the highest discriminative performance (AUC = 0.934), whereas in external validation, the artificial neural network (ANN) demonstrated the best performance (AUC = 0.921). Considering its consistent performance across datasets, ANN was selected as the final model, achieving AUC of 0.889 and 0.921 in the internal and external validation, respectively. Based on this model, a web-based clinical decision support tool was developed to provide individualized IV medication risk stratification at the point-of-care. CONCLUSION: A machine learning-based clinical decision support tool that incorporates routinely available clinical and medication-related variables can accurately stratify IV medication-related risk severity in hospitalized patients with HF.