Towards an Explainable Model for Sepsis Detection Based on Sensitivity Analysis

Sepsis is a life-threatening condition which responsible for high proportion of intra-hospital deaths and related healthcare costs each year. Early detection treatment sepsis episodes critical, since an early may highly improve prognosis. This study proposed original method to increase the interpretability set machine learning models onset. Open data from electronic medical records 40,336 patients monitored in intensive care units (ICU), provided by PhysioNet/Computing Cardiology Challenge 2019 used this paper. We including preprocessing, feature engineering, model construction tuning, as well analysis final stage. A total 24 were developed analyzed. The best model, based on 142 features achieved 0.4274 utility score. compact integrates only 20 selected features, score 0.3862. Meanwhile, sensitivity allows identification most relevant markers detect onset sepsis, their interdependence relative importance decision. optimized machine-learning predicting real-time way with performance, interpretable information significant biomarkers analyzed through novel method.