Balancing Accuracy and Simplicity in an Interpretable System for Sepsis Prediction Using Limited Clinical Data

Sepsis is a life-threatening condition caused by an excessive immune response to infection, and even a one-hour delay in treatment can result in irreversible organ damage and increased mortality. This study aimed to develop an interpretable and efficient machine learning-based system for early sepsis prediction using routinely collected electronic health record (EHR) data. The research question focused on whether high predictive performance could be achieved using only a minimal set of clinical features. Data were obtained from intensive care units and general wards in the PhysioNet Computing in Cardiology Challenge 2019 dataset. Thirty-seven predefined clinical features were extracted and systematically analyzed to assess their predictive contributions. Several machine learning models were trained and evaluated using area under the receiver operating feature curve (ROC-AUC) and accuracy metrics. The proposed model achieved an ROC-AUC of 0.929 and an accuracy of 0.926 when using all features. Remarkably, comparable performance was maintained (ROC-AUC = 0.912, accuracy = 0.907) when only 10 carefully selected features were used. The system outperformed existing state-of-the-art approaches while relying solely on commonly available clinical parameters. An interpretable, feature-efficient sepsis prediction system was successfully developed, demonstrating strong performance with minimal data requirements. The approach is well-suited for resource-limited healthcare settings, such as rural hospitals, and has the potential to reduce diagnostic burden while enabling timely intervention to improve patient outcomes.