Prediction of infection in the emergency department—a machine learning model

Background: The integration of artificial intelligence (AI) into emergency medicine holds promise for enhancing early diagnosis and clinical decision-making. Traditional diagnostic approaches often rely on physician judgment or early warning scores that may delay identification of infections, especially when these tools prioritize outcomes such as mortality or ICU admission over early infection detection. Objective: This study aimed to identify the most informative predictors of infection at emergency department (ED) admission using machine learning (ML), and to develop a predictive model for infection among acutely admitted patients with suspected infection. Methods: Four ML algorithms were evaluated: Logistic Regression (LR), K-Nearest Neighbors (KNN), Decision Tree (DT), and Random Forest (RF). Sequential forward selection was applied to LR and KNN to optimize predictor input. DT and RF incorporated intrinsic feature selection. Model performance was assessed using sensitivity, specificity, positive/negative predictive value (PPV/NPV), accuracy, area under the ROC curve (AUC), average precision score (AP) with five-fold cross-validation. Results: RF model outperformed the other ML models, achieving a specificity of 75%, a PPV of 94%, and an accuracy of 80% while DT showed marginally higher sensitivity but poorer specificity. LR and KNN demonstrated intermediate performance. The RF model had the highest mean AUC of 84% and AP of 96%. Conclusion: RF model using readily available clinical variables-including C-reactive protein, leucocyte count, temperature, diastolic blood pressure and heart rate-can effectively predict infection at ED admission. This supports the potential of ML to enhance early infection detection and guide timely treatment in emergency settings.