Interpretable Early Prediction of Sepsis Based on Counterfactual Inference

Abstract

Sepsis represents a life-threatening syndrome characterized by high mortality. Early sepsis prediction is imperative to enable timely clinical intervention and mitigate mortality. While numerous machine learning models demonstrate formidable accuracy in sepsis prediction, they lack adequate interpretability. Merely a handful of studies have endeavored to interpret sepsis prediction models, albeit with extremely limited interpretable capacity. In clinical practice, doctors prefer to ascertain alterations in predictions according to perturbations in features. This study develops an interpretable sepsis prediction model based on counterfactual inference. Subsequent to model selection, XGBoost is utilized for prediction. SHAP furnishes global and local feature importance rankings. Counterfactual inference investigates variations in predictions under feature perturbations, proffering insights into potential treatment regimens. This readily interpretable model endowed with pragmatic utility stands to benefit sepsis management through provision of accurate and interpretable predictions.