Enhancing mortality prediction in septic ICU patients with malignancies through ensemble machine learning: a...

Sepsis remains a leading cause of mortality in the intensive care unit (ICU), and patients with underlying malignancies are disproportionately affected. Conventional severity scores frequently provide inadequate prognostic discrimination in this vulnerable subgroup. We aimed to develop an ensemble machine learning model to predict 28-day ICU and in-hospital mortality in septic patients with malignancies using comprehensive ICU data. In this retrospective study, 4312 adult septic patients with malignancy were extracted from the MIMIC-IV database. Clinical data from the first ICU day were analyzed. Nine machine learning algorithms were assessed across five feature sets derived from different selection methods (LASSO, mRMR, Relief, Boruta). Top models were combined via a blending ensemble with Gaussian Naive Bayes as the meta-learner. Platt scaling was applied for probability calibration. Model performance was evaluated using discrimination (AUC), calibration (Brier score, calibration curves), clinical utility (decision curve analysis), and interpretability (SHAP analysis). Detailed calibration, DCA, and SHAP figures are provided in the Supplementary Materials. LASSO feature selection performed best. The final blending ensemble model achieved the highest predictive AUCs for 28-day ICU mortality (0.753) and in-hospital mortality (0.777), outperforming individual models and traditional scores (SOFA, OASIS, APSIII, SAPSII). Platt scaling improved calibration, reducing Brier scores from 0.218 to 0.184 (ICU 28-day) and from 0.194 to 0.163 (in-hospital). Decision curve analysis confirmed positive net benefit across clinically relevant threshold probabilities. SHAP analysis identified APSIII, BUN, PO2, and OASIS as top predictors. Survival analysis showed significant differences between model-stratified risk groups (log-rank p < 0.0001). In septic ICU patients with malignancies, the blending ensemble model achieved modest but statistically significant improvements in mortality prediction over conventional severity scores. The model offers a promising tool for early risk stratification, and prospective validation together with integration into clinical workflows is warranted.