Probabilistic modeling approach for interpretable inference and prediction with data for sepsis diagnosis

Abstract

Sepsis is a serious disease that can cause death. It is important to predict sepsis within the early stages after the presence of sepsis symptoms. In this paper, a new probabilistic modeling approach is used to establish classifiers for sepsis diagnosis. This approach is characterized by unique strong interpretability, which is reflected in three aspects: (1) evidence acquisition based on likelihood analysis, (2) probabilistic rule-based inference, and (3) parameters optimization using machine learning algorithms. Four-fold cross-validation is used to train and validate classifiers established by the new approach and alternative ones. Results show that in terms of classification capability, the classifier established by the new approach generally performs better than the majority of alternative classifiers for sepsis diagnosis, and close to the best one. As the classifier also features an inherent interpretability, it can be used as a tool for supporting diagnostic decision-making in sepsis diagnosis.