Integration of Clinical Indicators and Multiple Machine Learning Algorithms for Prognostic Evaluation in Sepsis Patients with Different BMI: Model Construction and Validation

Objective Sepsis is a leading cause of intensive care unit (ICU) mortality, with body mass index (BMI) contributing to prognostic heterogeneity through the so-called "obesity paradox." This study aimed to develep and validate a BMI-stratified prognostic model for 28-day mortality in sepsis patients by integrating clinical indicators with multiple machine learning (ML) algorithms, and to explore BMI-specific predictive patterns. Methods We conducted a retrospective analysis of 654 sepsis patients admitted between August 2022–August 2024. Demographic (age, gender), anthropometric (weight/height→BMI), vital signs (heart rate, respiratory rate, SpO₂), laboratory (CRP, PCT, D-Dimer, PT, APTT), and severity scores (SOFA, APACHE II, GCS) were collected. Patients were stratified into three BMI groups: underweight (BMI < 18.5 kg/m², n = 98), normal weight (BMI 18.5–23.9 kg/m², n = 276), and overweight/obese (BMI ≥ 24 kg/m², n = 280). Eight ML algorithms were employed: Regularized Logistic Regression (LR-L1/L2), Random Forest (RF), Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), CatBoost, Support Vector Machine (SVM) with RBF kernel, and Multi-Layer Perceptron (MLP). Feature engineering comprised recursive feature elimination (RFE), multiple imputation for missing values, and outlier handling using the interquartile rang (IQR)method. Model performance was evaluated via 10-fold cross-validation (CV) and external validation (15% of cohort) using AUC, sensitivity, specificity, accuracy, and F1-score. SHAP (SHapley Additive exPlanations) and permutation importance were used for interpretability. Subgroup analyses compared model performance across BMI strata. Results The 28-day mortality rate was highest among underweight patients (45.9%), followed by those with normal weight (25.3%) and overweight/obese patients (18.7%), a trend consistent with the obesity paradox (p < 0.001). The RFE method identified a set of 13 key predictors: BMI group, SOFA score, APACHE II score, PCT, CRP, D-Dimer, initial heart rate, SpO₂, mechanical ventilation duration, age, and underlying disease. Among ML models, CatBoost demonstrated the best overall performance (training set: AUC = 0.90, 95% CI: 0.87–0.93; sensitivity = 0.86; specificity = 0.84; accuracy = 0.85; external validation: AUC = 0.88, 95% CI: 0.82–0.93). Subgroup analysis revealed: (1) Underweight group: XGBoost performed best (AUC = 0.87) with SOFA score and D-Dimer as top predictors; (2) Normal weight group: LightGBM was optimal (AUC = 0.86) driven by APACHE II score and PCT; (3) Overweight/obese group: CatBoost outperformed (AUC = 0.89) with BMI, CRP, and mechanical ventilation duration as key features. SHAP analysis revealed taht a SOFA score > 10 was consistently associated with a significantly increased mortality risk across all BMI groups, while BMI 24–28 kg/m² was protective only in patients aged ≥ 65 years. Conclusion The CatBoost model integrating clinical indicators and BMI stratification exhibits robust performance for 28-day mortality prediction in 654 sepsis patients. BMI-specific ML models and SHAP-based interpretability provide personalized risk stratification, supporting BMI-tailored sepsis management.