Explainable Retrospective Sepsis Classification Based on Elixhauser Comorbidity Groups Using Machine and Deep...

Area of Science:

• Computational biology
• Medical informatics
• Machine learning in healthcare

Background:

• Sepsis presents heterogeneously in patients.
• Comorbidities can significantly alter physiological signs of sepsis.
• Understanding these variations is crucial for accurate sepsis detection.

Purpose of the Study:

• To investigate if comorbidity-specific modeling improves retrospective sepsis discrimination.
• To compare the performance of various machine learning algorithms across different patient subgroups.
• To identify key features contributing to sepsis prediction in diverse patient populations.

Main Methods:

• Analysis of MIMIC-IV data across 27 Elixhauser-defined subgroups.
• Utilized Random Forest, XGBoost, CatBoost, LightGBM, AdaBoost, multilayer perceptron, TabNet, and FT-Transformer models.
• Employed SMOTE for class imbalance and SHAP for feature attribution.

Main Results:

• Tree-based models achieved ROC-AUC > 0.80 in multiple subgroups.
• Neural networks exhibited variable performance with trade-offs in discrimination and recall.
• Red cell distribution width, age, and magnesium were consistently important predictors.

Conclusions:

• Sepsis discrimination varies significantly based on patient subgroups and the chosen machine learning model.
• Current findings reflect retrospective analysis and require external validation for prospective clinical use.
• Further calibration and decision-analytic evaluation are necessary before clinical implementation.