Machine learning algorithms for ICU mortality prediction in community-acquired pneumonia: A standardized validation and calibration study on the NACef cohort
Medical Data Mining · 2026 · 9 (3) · 17
DOI: 10.53388/mdm202609017 ↗Kurz gefasst
Fachartikel in der Zeitschrift Medical Data Mining, 2026. Verfasst von Hanane Belmouss und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Background
Mortality in intensive care due to community-acquired pneumonia remains high. Although machine learning models have demonstrated promising predictive performance, standardized validation and explicit calibration assessment across independent clinical settings remain limited.
Methods
This study systematically reviewed machine learning algorithms for mortality prediction in community-acquired pneumonia. A search across six databases identified 241 records, of which seven met the eligibility criteria. Standalone algorithms meeting predefined selection criteria were subsequently implemented within a standardized validation framework and evaluated on the NACef cohort (n = 764; 163 deaths, 21.2%). Model performance was assessed using nested cross-validation and independent hold-out testing, with evaluation of both discrimination and calibration metrics.
Results
Three standalone algorithms, XGBoost, LightGBM, and Logistic Regression, were selected for standardized implementation and evaluation. In nested cross-validation, all models achieved mean AUC values above 0.90. On the independent hold-out test set, uncalibrated AUC values were 0.941 for XGBoost, 0.933 for LightGBM, and 0.915 for Logistic Regression, with small absolute differences across models. Recall ranged from 0.818 (LightGBM) to 0.969 (Logistic Regression), while precision ranged from 0.615 (Logistic Regression) to 0.658 (LightGBM). Calibration analysis indicated probability misalignment before recalibration. Platt scaling and isotonic regression improved calibration metrics, with isotonic regression achieving the lowest expected calibration error while maintaining comparable discrimination.
Conclusion
Logistic Regression, XGBoost, and LightGBM demonstrated comparable discrimination, and calibration improved probability reliability, underscoring the importance of harmonized validation and explicit calibration assessment in clinical machine learning research.
Zitieren
APA
Belmouss, H., & Aamouche, A. (2026). Machine learning algorithms for ICU mortality prediction in community-acquired pneumonia: A standardized validation and calibration study on the NACef cohort. Medical Data Mining, 9(3), 17. https://doi.org/10.53388/mdm202609017
Vancouver
Belmouss H, Aamouche A. Machine learning algorithms for ICU mortality prediction in community-acquired pneumonia: A standardized validation and calibration study on the NACef cohort. Medical Data Mining. 2026;9(3):17. doi:10.53388/mdm202609017
BibTeX
@article{Belmouss2026,
author = {Belmouss, Hanane and Aamouche, Ahmed},
title = {Machine learning algorithms for ICU mortality prediction in community-acquired pneumonia: A standardized validation and calibration study on the NACef cohort},
journal = {Medical Data Mining},
year = {2026},
volume = {9},
number = {3},
pages = {17},
doi = {10.53388/mdm202609017}
}
EUVIMED