The research team at Cho Ray Hospital enrolled consecutive patients from October 2024 through June 2025. Patients qualified if they were at least 18 years old and met the 2021 European Society of Cardiology criteria for acute heart failure. The study excluded people with isolated right ventricular failure, those who declined participation, and cases in which survival could not be established. All candidate information was obtained at admission or during the first 24 hours, before the outcome occurred, creating a defined prediction time and limiting temporal leakage. The primary endpoint was all-cause death during the index hospitalization; 53 of the 572 participants, or 9.3%, died in hospital.

The dataset covered demographics, medical history, comorbidities, vital signs, laboratory measurements, electrocardiographic findings and echocardiographic characteristics. Predictors included respiratory failure, cardiogenic shock, blood pressure, renal function, electrolyte concentrations, cardiac biomarkers and imaging measurements. Missing values were imputed with a random-forest-based iterative algorithm, and variables with more than 20% missing data were removed; systolic pulmonary artery pressure was excluded on that basis. After preprocessing, 31 candidate predictors remained. The data were divided into a training cohort of 429 patients and an independent test cohort of 143, with all preprocessing fitted on the training data and then applied unchanged to the test set.

Researchers trained six algorithms: logistic regression, elastic-net logistic regression, random forest, XGBoost, LightGBM and CatBoost. Because deaths were relatively uncommon, synthetic minority oversampling technique, known as SMOTE, was applied only within each cross-validation training fold, while class weighting was used where supported. Hyperparameters were selected through searches constrained to relatively shallow trees, reduced learning rates, subsampling and minimum-leaf limits. The analysis included stratified 10-fold cross-validation, repeated 10-by-10 cross-validation, nested cross-validation during tuning and 1,000 bootstrap resamples for optimism correction. Decision thresholds were derived solely from training data and then applied to the test set. The study followed the TRIPOD + AI reporting statement.

On the independent test set, random forest had the highest observed discrimination, with a 95% confidence interval of 0.757 to 0.922, and an optimism-corrected AUC of 0.81 from repeated cross-validation and bootstrap analysis. Logistic regression reached 0.795, elastic net 0.794, CatBoost 0.786, XGBoost 0.763 and LightGBM 0.694. The GWTG-HF score reached 0.817 and had a precision-recall AUC of 0.33, compared with 0.31 for random forest. Pairwise DeLong tests found no significant AUC difference between any machine-learning model and GWTG-HF, with all p-values above 0.05. Random forest’s difference from the established score was 0.029 and its DeLong p-value was 0.713, but those calculations rested on just 13 test-set deaths. The authors therefore emphasized that statistical non-significance in this underpowered comparison cannot establish equivalence.

At thresholds selected from training data, random forest achieved 84.6% sensitivity, 70.8% specificity and 72.0% accuracy, while XGBoost recorded the highest accuracy at 74.1% and specificity at 73.8%, with sensitivity of 76.9%. The GWTG-HF score produced 92.3% sensitivity, 67.7% specificity and 69.9% accuracy. Precision was low for every approach, including 0.224 for random forest and 0.222 for GWTG-HF, partly reflecting the low 9.1% mortality prevalence in the test set. Wide confidence intervals around the classification estimates added to the uncertainty. Decision-curve analysis nevertheless suggested that both random forest and GWTG-HF offered more net benefit than treating all or no patients as high risk across threshold probabilities of roughly 0.05 to 0.30, with GWTG-HF producing the greatest net benefit over most of that range.

Calibration exposed a major limitation of the machine-learning predictions. CatBoost had the numerically lowest Brier score, 0.085, and random forest had the calibration slope closest to one at 1.087, but the scaled Brier scores were low or negative for every machine-learning model. Their calibration intercepts ranged from −1.00 to −1.91, indicating systematic overestimation of predicted risk. The investigators associated this pattern with aggressive imbalance correction and noted that reducing or omitting over-correction and applying post hoc recalibration improved calibration without materially changing discrimination. The GWTG-HF score was better calibrated, with a Brier score of 0.073, scaled Brier score of 0.117, slope of 1.633 and intercept of −0.095. These findings separate the ability of a model to order patients by risk from the reliability of its absolute probability estimates.

To make the leading model interpretable, the researchers used SHAP, or SHapley Additive exPlanations, which decomposes each prediction into the contributions of individual variables. The most influential features included acute decompensated heart failure, respiratory failure, NT-proBNP, high-sensitivity troponin I and left ventricular dilation. Direction was important: higher NT-proBNP, higher troponin I, respiratory failure and cardiogenic shock increased predicted mortality risk, whereas an acute decompensated heart failure phenotype and left ventricular dilation were associated with lower predicted risk. The latter pattern was consistent with the descriptive data, in which decompensation was less common among nonsurvivors while de novo and shock presentations made up a larger share of fatal cases. These SHAP findings describe associations learned by the model, not causal effects, and the model combined markers of acute hemodynamic decompensation, myocardial injury, congestion and end-organ dysfunction.

Several design features strengthen the study, including prospective enrollment, ascertainment of outcomes, a prespecified prediction time, prevention of preprocessing leakage and evaluation of discrimination, calibration and decision utility. Important constraints remain. Only 53 deaths occurred, with roughly 40 in the development cohort and 13 in the test cohort. The resulting events-per-variable ratio was approximately 1.7 overall and 1.3 in the development data, well below conventional guidance and enough to make model stability difficult to establish. The work was also limited to one tertiary center and a nine-month enrollment period. Internal validation and optimism correction cannot substitute for external testing across hospitals, populations and time periods. In addition, all patients were classified as non-Black when calculating the race component of GWTG-HF, which was considered appropriate for this Vietnamese cohort but restricts transportability to more ethnically diverse populations.

The practical message is measured rather than algorithm-centric: explainable machine learning produced clinically coherent risk rankings in a prospective Southeast Asian acute-heart-failure cohort, but its advantage over a conventional clinical score remains unproven. The random-forest result is promising, particularly because its SHAP explanations linked predictions to recognizable severity markers, yet the low event count and miscalibrated probabilities prevent the use of its raw outputs for individual decisions. Before implementation in an electronic health record, the models would require recalibration, multicenter external validation, temporal validation, assessment of clinician usability and alert burden, appropriate governance and prospective evaluation of clinical impact. Until then, the findings support continued study of explainable models as possible complements to established risk tools rather than as ready replacements.

Subject of Research: Prospective evaluation of explainable machine-learning models for predicting in-hospital mortality in acute heart failure

Article Title: Explainable machine learning for predicting in-hospital mortality in acute heart failure: a prospective single-center cohort study from Vietnam

Article References: Hoang, S. V., Nguyen, K. M., Nguyen, T. N., & Do, H. N. (2026). Explainable machine learning for predicting in-hospital mortality in acute heart failure: a prospective single-center cohort study from Vietnam. Frontiers in Cardiovascular Medicine, 13, Article 1927124. https://doi.org/10.3389/fcvm.2026.1927124

Image Credits: AI Generated

DOI: 10.3389/fcvm.2026.1927124

Keywords: acute heart failure, in-hospital mortality, explainable AI, machine learning, random forest, risk prediction, SHAP explanations, model calibration, clinical score comparison, Vietnam, Explainable, machine