JMIR Publications

ABSTRACT

Background:

Sepsis-associated liver injury is a serious complication of sepsis, contributing to increased mortality and morbidity. Early identification of sepsis-associated liver injury is critical for improving patient outcomes, yet the heterogeneity of sepsis makes timely diagnosis challenging. Machine learning offers promising solutions for predicting adverse outcomes in sepsis patients.

Objective:

we aim to develop an explainable machine learning model, incorporating stacking techniques, to predict the occurrence of liver injury in sepsis patients.

Methods:

This study utilized data from two large ICU databases (MIMIC-IV and eICU) to develop and evaluate machine learning models for predicting sepsis -associated liver injury in critically ill patients. Nine different Machine learning models, including decision trees, random forests, and XGBoost, were employed. A stacking ensemble method was implemented to integrate the strengths of individual models. Model performance was assessed using ROC-AUC, PR-AUC, and other relevant metrics, with SHAP values used to interpret feature importance.

Results:

The LightGBM, XGBoost, and random forest models demonstrated superior performance in both internal and external validation sets. The stacking model further improved prediction accuracy, achieving the highest ROC-AUC scores. Key predictors identified include total bilirubin, lactate, and prothrombin time, with SHAP analysis confirming their significant contribution to Sepsis-associated liver injury prediction.

Conclusions:

The stacking ensemble model developed in this study provides accurate and robust predictions of sepsis-associated liver injury in sepsis patients, with potential clinical utility for early intervention and personalized treatment strategies.