HRAE-Stack: A Leakage-Controlled, Calibrated, and Explainable Stacking Framework for ICU Mortality Prediction

Authors

• Kareem Hameed Khalaf, Faculty of Electrical and Computer Engineering, University of Tabriz, Iran
• A. Moallemi Khiavi, Faculty of Electrical and Computer Engineering, University of Tabriz, Iran
• Dhafar Hamed Abd, College of Computer Science and Information Technology, University of Anbar, Ramadi, IraqFollow

Abstract

ICU mortality prediction requires more than good discriminative ability; the development process should minimize optimistic bias and produce clinically meaningful probabilities. In this work, we introduce HRAE-Stack, a heterogeneous stacking method that combines fold-dependent preprocessing, consensus feature selection, four tree-based models, entropy-modulated out-of-fold meta-features, Bayesian ridge averaging, and probability calibration. We estimated all transformations and model selection decisions within development folds and fixed them before the model is evaluated on the unobserved held-out partition. Mean receiver operating characteristic–area under the curve (ROC-AUC) was 0.974 ± 0.018, and the mean Brier score was 0.045 ± 0.032 for 10-fold cross-validation. ROC-AUC for the held-out partition (n = 553) was 0.979, and 531 patients (96.0%) were correctly classified. Ablation studies showed that each architectural component does not have a universally positive impact; rather, metric-dependent trade-offs were derived. Meanwhile, stress tests indicated that the performance degraded more significantly under additive noise than under missing data or feature permutations. The calibration slope for the held-out partition is close to unity, but calibration in the large suggests continued underprediction. Therefore, the main contribution of this work is a controlled development and multidimensional evaluation framework, and not a novel machine learning algorithm. Independent validation of the proposed method is needed before further clinical use.

Recommended Citation

Khalaf, Kareem Hameed; Khiavi, A. Moallemi; and Abd, Dhafar Hamed (2026) "HRAE-Stack: A Leakage-Controlled, Calibrated, and Explainable Stacking Framework for ICU Mortality Prediction," Mesopotamian Journal of Artificial Intelligence in Healthcare: Vol. 4: Iss. 1, Article 4.