Improved Interpretability Without Performance Reduction in a Sepsis Prediction Risk Score

TL;DR

This paper introduces STEWS, an interpretable sepsis prediction score that performs as well as a machine learning model.

Contribution

The novel contribution is converting an ML model into an interpretable integer score without sacrificing performance.

Findings

STEWs was shown to be significantly equivalent to logistic regression in sepsis prediction.

The conversion of ML models into interpretable forms does not reduce predictive power.

Abstract

Objective: Sepsis is a life-threatening response to infection and a major cause of hospital mortality. Machine learning (ML) models have demonstrated better sepsis prediction performance than integer risk scores but are less widely used in clinical settings, in part due to lower interpretability. This study aimed to improve the interpretability of an ML-based model without reducing its performance in non-ICU sepsis prediction. Methods: A logistic regression model was trained to predict sepsis onset and then converted into a more interpretable integer point system, STEWS, using its regression coefficients. We compared STEWS with the logistic regression model using PPV at 90% sensitivity. Results: STEWS was significantly equivalent to logistic regression using the two one-sided tests procedure (0.051 vs. 0.051; p = 0.004). Conclusions: STEWS demonstrated equivalent performance to a comparable logistic regression model for non-ICU sepsis prediction, suggesting that converting ML models into more interpretable forms does not necessarily reduce predictive power.

Figures