Electronic-Medical-Record-Based Identification of Sepsis Patients in Emergency Department: A Machine Learning Perspective

Early identification of patients with sepsis is critical to reduce fatality and provide timely care. In this study, we compare the performance of several major machine learning (ML) techniques to identify emergency department (ED) sepsis patients during their first 6 hours of care. The ML techniques used are Decision Tree, Discriminant Analysis, Logistic Regression, k-Nearest Neighbors (KNN), Ensemble Classification, Support Vector Machine (SVM) and Neural Network (NN). We also included a novel genetic-algorithm-optimized rule-based system that we developed previously in a separate study. This comparative study uses 912 sepsis and 975 non-sepsis patients admitted to the ED of the Detroit Medical Center in Michigan, USA. The cases were individually adjudicated by expert physicians. The Fine Gaussian SVM showed the highest sensitivity (95.8%) but with only moderate 77.7% specificity and 80.1% positive predictive value (PPV). The Coarse KNN had the highest specificity (94.5%), but sensitivity was only at 84.9%. The rule-based system attained 90.9% sensitivity, 90.9% specificity and 90.3% PPV. The NN model achieved 92.08% sensitivity, 92.33% specificity, and 91.78% PPV, slightly better than the rest of the ML models and the rule-based system.