Explainable Retrospective Sepsis Classification Based on Elixhauser Comorbidity Groups Using Machine and Deep Learning on MIMIC-IV

Background/Objectives: Sepsis is clinically heterogeneous, and comorbidities may alter physiological patterns associated with sepsis status. This study examined whether comorbidity-specific modeling revealed differences in retrospective sepsis discrimination. Methods: MIMIC-IV data were analyzed across 27 Elixhauser-based analytical subgroups using Random Forest, XGBoost, CatBoost, LightGBM, AdaBoost, multilayer perceptron, TabNet, and FT-Transformer. Available neural-model scripts used a stratified 80/20 training–evaluation split. Class imbalance was addressed using training-partition SMOTE and model-specific loss or sampling settings. SHAP was used to describe model-attributed feature contributions. Results: Tree-based ensembles achieved ROC-AUC values above 0.80 in many subgroups, whereas neural models showed more variable performance and trade-offs between discrimination and sepsis-class recall. Red cell distribution width, age, and magnesium were frequently emphasized by the fitted models. Conclusions: The findings demonstrate subgroup- and model-specific differences in retrospective sepsis-status discrimination. Because a verified pre-sepsis feature-time boundary was not enforced, the results should not be interpreted as prospectively validated early prediction. External validation, calibration, and decision-analytic evaluation are required before clinical use.