Document resource
Background and Importance Drug-related problems (DRPs) are a common and preventable source of morbidity in emergency departments (EDs). Machine learning (ML) techniques may enhance early DRP detection and risk stratification, supporting clinical pharmacists in optimising patient safety.Aim and Objectives To develop and validate the performance of three decision models for predicting DRPs in ED patients–comparing a conventional logistic regression model with two ML-based models–and to evaluate the potential of the best-performing model to prioritise valuable pharmacist-led interventions.Material and Methods A retrospective observational study in the ED of a tertiary hospital (March–June 2025) was conducted. Adult patients (≥18 years) receiving ≥1 prescribed medication during standard pharmacy working hours (Monday–Friday) were included. Predictors included demographics, frailty score, ED length of stay, triage level, admission diagnosis, planned hospital admission, high alert medications, and prior isolation of multidrug-resistant bacteria. A random forest (RF) model, a K-means clustering approach, and a multivariate logistic regression model were developed. Model performance was assessed in separate training (80%) and validation (20%) cohorts using area under the receiver operating characteristic (ROC) curve (AUC), sensitivity, specificity, and accuracy. Shapley Additive explanations (SHAP) analysis was conducted to analyse the RF model, and values were used to interpret the most influential predictors.Results Among 5064 patients (mean age 72.1 years; 53.6% female), 823 (16.2%) experienced ≥1 DRP, most frequently medication reconciliation errors (45.5%). In the training cohort, decision models showed AUCs-ROC of 0.685 for logistic regression, 0.720 for RF, and 0.551 for K-means clustering. The RF model achieved the best balance of sensitivity (0.727) and specificity (0.529), modestly outperforming logistic regression (sensitivity 0.864; specificity 0.378). SHAP analysis confirmed frailty score, age, hospital admission prevision, chronic high alert medication and clinical presentation of heart failure decompensation or dyspnoea as key predictors.Conclusion and Relevance Our study supports the integration of ML-based decision RF models into ED workflows to enhance the detection of DRPs, being a promising alternative to traditional scores based on logistic regression models. These tools could enable pharmacists to rapidly identify patients at higher risk of DRPs, prioritise interventions, and ultimately improve patient safety. Future research should evaluate external validation, and the impact of RF implementation on clinical outcomes.Conflict of Interest No conflict of interest