Abstract
BACKGROUND: Critically ill patients are managed with complex medication regimens that require medication management to optimize safety and efficacy. When performed by a critical care pharmacist (CCP), discrete medication management activities are termed medication interventions. The ability to define CCP workflow and intervention timeliness depends on the ability to predict the medication management needs of individual intensive care unit (ICU) patients. The purpose of this study was to develop prediction models for the number and intensity of medication interventions in critically ill patients. METHODS: This was a retrospective, observational cohort study of adult patients admitted to an ICU between June 1, 2020 and June 7, 2023. Models to predict number of pharmacist interventions using both patient and medication related predictor variables collected at either baseline or in the first 24 hours of ICU stay were created. Both regression and supervised machine learning models (Random Forest, Support Vector Machine, and XGBoost) were developed. Root mean square derivation (RMSE), mean absolute error (MAE), and symmetric mean absolute percentage error (sMAPE) were calculated. RESULTS: In a cohort of 13 373 patients, the average number of interventions was 4.7 (standard deviation [SD] 7.1) and intervention intensity was 24.0 (40.3). Among the ML models, the Random Forest model had the lowest RMSE (9.26) while Support Vector Machine had the lowest MAE (4.71). All machine learning models performed similarly to the stepwise logistic regression model, and these performed better than a base model combining severity of illness with medication regimen complexity scores. CONCLUSIONS: Intervention quantity can be predicted using prediction models that incorporate patient-specific factors in the first 24 hours of admission. In this case, machine learning methods did not provide a substantial advantage in performance, but given that inter-institutional variation in intervention documentation precludes external validation, our results provide a framework for workload modeling at any institution where the proposed models here could be evaluated.