Early Prediction of Mortality for Septic Patients Visiting Emergency Room Based on Explainable Machine Learning: A...

Area of Science:

• Medical Informatics
• Critical Care Medicine
• Machine Learning

Background:

• Sepsis is a leading cause of hospital mortality worldwide.
• Early prediction of sepsis mortality is crucial for efficient resource allocation.
• Machine learning (ML) models can potentially improve early mortality prediction in sepsis patients.

Purpose of the Study:

• To construct and evaluate ML models for predicting sepsis patient mortality in emergency departments.
• To compare the predictive performance of ML models using clinical variables versus SOFA components.
• To identify key clinical variables influencing sepsis mortality prediction.

Main Methods:

• Prospective, multicenter cohort study of sepsis patients in emergency departments (September 2019 - December 2020).
• Six ML models (logistic regression, SVM, random forest, XGBoost, LGBM, CatBoost) were trained using fivefold cross-validation.
• Compared 44 baseline clinical variables against six SOFA components (PF, PLT, bilirubin, cardiovascular, GCS, creatinine) for mortality prediction.

Main Results:

• CatBoost achieved the highest AUC (0.800) using clinical variables; XGBoost achieved the highest AUC (0.678) using SOFA components.
• Albumin, lactate, blood urea nitrogen, and international normalization ratio significantly impacted predictions.
• PaO2/FiO2 (PF) and platelet (PLT) counts were significant SOFA-related predictors.

Conclusions:

• ML models demonstrate good performance in predicting sepsis mortality.
• Baseline clinical variables provide more accurate early predictions than SOFA components.
• The study identified significant variables impacting sepsis mortality prediction, aiding clinical decision-making.