A Hybrid AutoML Ensemble Integrating Conventional Learners and Gradient-Boosting Models for Multi-Outcome Prediction in ICU Patients with Pseudomonas aeruginosa

Carbapenem-resistant Pseudomonas aeruginosa (PA) jeopardises intensive-care patients worldwide. We developed a real-time, interpretable hybrid automated machine learning (AutoML) ensemble to predict multiple outcomes. A retrospective cohort of 847 ICU admissions with PA (2018–2024) underwent VTF–MI–L1 feature selection; XGBoost, LightGBM, CatBoost, random forests and linear/logistic regressors were ensembled via bagging, voting, stacking and boosting. Nested five-fold cross-validation evaluated performance (AUC for classification; MSE, RMSE, MAE and R² for regression); SHAP explained predictions, and inference latency was recorded. Across four regression endpoints—carbapenem-resistance rate (CRR), average CRR of the last two isolates (CRR-PA-Last2), ICU length of stay (ICU-LOS) and time from ICU admission to death (ICU-Death interval)—XGBoost regressor (XGB-R) performed best (mean MSE = 9.76 × 10³, RMSE = 64.11, MAE = 25.24, R² = 0.77; mean Friedman rank = 1.95). For classification, the Voting Classifier achieved the highest AUC (0.842) for in-hospital mortality (IHM), whereas the LightGBM classifier led for antimicrobial susceptibility of the last PA isolate before discharge (LastPaAST, AUC = 0.981). SHAP highlighted age, cumulative carbapenem exposure, the durations of mechanical ventilation (MV-days), central venous catheterisation (CVC-days) and urinary catheterisation (UC-days) as key contributors. All top models produced predictions in < 50 ms, supporting bedside antimicrobial-stewardship and infection-control decisions; multicentre prospective validation is warranted.