Early Sepsis Prediction Using Interpretable Models

Sepsis remains a major cause of morbidity and mortality in Intensive Care Units (ICUs). Timely identification of sepsis can prevent severe complications by enabling early treatment, such as administering antibiotics. Despite advances in diagnostic biomarkers and scoring systems, these approaches often lack the ability to predict sepsis onset with sufficient lead time or fail to generalize across diverse patient populations. In this study, we propose an explainable machine learning (ML) pipeline that leverages data from 45,285 patients in the Medical Information Mart for Intensive Care (MIMIC)-IV dataset. We develop an Ensemble of a Gradient Boosting Model and a hybrid Long Short-Term Memory Network, utilizing 21 clinically relevant features to predict sepsis 12 hours in advance. The method achieves an Area Under the Curve (AUC) of 0.87, an Area Under the Precision-Recall curve of 0.88, and a sensitivity of 0.79 at a specificity of 0.81. Decision Curve Analysis suggests strong clinical utility with maximum net benefit 0.45. We enhance interpretability by using the TE2Rules explainability library, achieving an overall fidelity score of 95% with just 42 rules for a positive prediction and 47 rules for a negative prediction. Argumentation theory raises Ensemble sensitivity to 81%. Metabolic markers, hemodynamics and neurological status can be predictive of sepsis in the next 12 hours. Variable hourly differences suggested by ICU experts and window statistics are proven important for sepsis prediction. This is the first study to investigate sepsis prediction in ICU patients using MIMIC IV, achieving a significantly higher AUC and a longer prediction window, and the first to use rule-based explainable AI and argumentation theory to interpret sepsis predictions. The pipeline is novelly assessed on the eICU Database yielding a comparable AUC of 0.88. This interpretable framework for sepsis prediction can provide healthcare professionals with reliable insights for early intervention, ultimately reducing mortality rates and healthcare costs.