Explainable machine learning for predicting in-hospital mortality in acute heart failure: a prospective single-center cohort study from Vietnam
- 1Department of Internal Medicine, School of Medicine, University of Medicine and Pharmacy at Ho Chi Minh City
- 2Department of Cardiology, Cho Ray Hospital
- 3Department of Cardiology, Nhan Dan Gia Dinh Hospital
Abstract
Background: Accurate risk stratification is essential in acute heart failure (AHF), which carries substantial in-hospital mortality. The Get With The Guidelines–Heart Failure (GWTG-HF) score is widely used, but its performance may vary across populations, and prospective data from Southeast Asia are scarce. We developed and internally validated explainable machine-learning (ML) models for in-hospital mortality in a Vietnamese AHF cohort and compared them with the GWTG-HF score.
Methods: We prospectively enrolled 572 consecutive patients admitted with AHF (October 2024–June 2025). Missing data were imputed with a random forest–based algorithm, and class imbalance handled with SMOTE within training folds. Six algorithms (logistic regression, elastic net, random forest, XGBoost, LightGBM, CatBoost) were trained with stratified cross-validation, with optimism correction, and evaluated on an independent 25% test set. We assessed discrimination (ROC-AUC, PR-AUC), calibration (calibration slope/intercept, Brier and scaled Brier scores), classification, and clinical utility (decision-curve analysis); thresholds were derived from training data only. Pairwise AUC comparisons used the DeLong test, and interpretation used SHAP. Reporting followed TRIPOD + AI.
Results: Fifty-three patients (9.3%) died in hospital. Random forest had the highest discrimination (AUC 0.846, 95% CI 0.757–0.922), followed by the GWTG-HF score (0.817), logistic regression (0.795), elastic net (0.794), CatBoost (0.786), XGBoost (0.763), and LightGBM (0.694). No ML model significantly outperformed the GWTG-HF score (all DeLong p > 0.05), a comparison limited by only 13 test-set events. Although CatBoost had the numerically lowest Brier score (0.085), scaled Brier scores were low for all models and calibration intercepts were uniformly negative, indicating over-estimation of risk from aggressive imbalance correction; predicted probabilities therefore require recalibration. SHAP identified respiratory failure, NT-proBNP, high-sensitivity troponin I (risk-increasing) and an ADHF phenotype and left ventricular dilation (associated with lower predicted risk) as the most influential features.
Conclusions: Explainable ML, led by random forest, achieved good discrimination but did not significantly outperform the GWTG-HF score, most plausibly reflecting limited statistical power rather than true equivalence. With miscalibrated probabilities requiring recalibration, explainable ML is best viewed as a complementary risk-ranking tool; external, multicenter, and temporal validation are required before clinical use.
Related
- Explainable machine learning and deep learning models for predicting in-hospital mortality in burn patients
- A machine learning-driven and explainable prediction model for in-hospital mortality in patients with diabetes mellitus and congestive heart failure: integrating XGBoost and SHAP approaches.
- Development and validation of an interpretable machine learning model for predicting in-hospital mortality among critically ill patients with liver cirrhosis
- Mortality prediction in patients with acute myocardial infarction and prior ischemic stroke: a prospective cohort study with explainable machine learning and SHAP analysis
- Development and temporal validation of an explainable machine learning model for predicting in-hospital diabetic foot ulcer in patients with type 2 diabetes using inflammatory–nutritional composite indices
BibTeX
@article{hoang2026explainable,
title = {{Explainable machine learning for predicting in-hospital mortality in acute heart failure: a prospective single-center cohort study from Vietnam}},
author = {Sy Van Hoang and Kha Minh Nguyen and Tai Nhat Nguyen and Huy Nhat Do},
journal = {Frontiers in Cardiovascular Medicine},
year = {2026},
month = oct,
doi = {10.3389/fcvm.2026.1927124},
url = {https://www.frontiersin.org/articles/10.3389/fcvm.2026.1927124/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Explainable machine learning for predicting in-hospital mortality in acute heart failure: a prospective single-center cohort study from Vietnam
AU - Sy Van Hoang
AU - Kha Minh Nguyen
AU - Tai Nhat Nguyen
AU - Huy Nhat Do
PY - 2026
DA - 2026/10/21
JO - Frontiers in Cardiovascular Medicine
DO - 10.3389/fcvm.2026.1927124
UR - https://www.frontiersin.org/articles/10.3389/fcvm.2026.1927124/full
AB - Background: Accurate risk stratification is essential in acute heart failure (AHF), which carries substantial in-hospital mortality. The Get With The Guidelines–Heart Failure (GWTG-HF) score is widely used, but its performance may vary across populations, and prospective data from Southeast Asia are scarce. We developed and internally validated explainable machine-learning (ML) models for in-hospital mortality in a Vietnamese AHF cohort and compared them with the GWTG-HF score. Methods: We prospectively enrolled 572 consecutive patients admitted with AHF (October 2024–June 2025). Missing data were imputed with a random forest–based algorithm, and class imbalance handled with SMOTE within training folds. Six algorithms (logistic regression, elastic net, random forest, XGBoost, LightGBM, CatBoost) were trained with stratified cross-validation, with optimism correction, and evaluated on an independent 25% test set. We assessed discrimination (ROC-AUC, PR-AUC), calibration (calibration slope/intercept, Brier and scaled Brier scores), classification, and clinical utility (decision-curve analysis); thresholds were derived from training data only. Pairwise AUC comparisons used the DeLong test, and interpretation used SHAP. Reporting followed TRIPOD + AI. Results: Fifty-three patients (9.3%) died in hospital. Random forest had the highest discrimination (AUC 0.846, 95% CI 0.757–0.922), followed by the GWTG-HF score (0.817), logistic regression (0.795), elastic net (0.794), CatBoost (0.786), XGBoost (0.763), and LightGBM (0.694). No ML model significantly outperformed the GWTG-HF score (all DeLong p > 0.05), a comparison limited by only 13 test-set events. Although CatBoost had the numerically lowest Brier score (0.085), scaled Brier scores were low for all models and calibration intercepts were uniformly negative, indicating over-estimation of risk from aggressive imbalance correction; predicted probabilities therefore require recalibration. SHAP identified respiratory failure, NT-proBNP, high-sensitivity troponin I (risk-increasing) and an ADHF phenotype and left ventricular dilation (associated with lower predicted risk) as the most influential features. Conclusions: Explainable ML, led by random forest, achieved good discrimination but did not significantly outperform the GWTG-HF score, most plausibly reflecting limited statistical power rather than true equivalence. With miscalibrated probabilities requiring recalibration, explainable ML is best viewed as a complementary risk-ranking tool; external, multicenter, and temporal validation are required before clinical use.
SN - 2297-055X
ER -