Key result
XGBoost models predict in-hospital mortality with ~0.85 AUROC, outperforming transferred cross-endpoint models.
Why the study?
Patients with heart failure face substantial morbidity and mortality. This study aimed to build machine learning models for predicting in-hospital or post-discharge mortality in acute decompensated heart failure and explore whether models trained on one mortality endpoint transfer to the other.
Do machine learning models (XGBoost and logistic regression) accurately predict in-hospital and 1-month post-discharge mortality in patients hospitalized with acute decompensated heart failure, and do models trained on one mortality endpoint transfer effectively to the other?
Cohort (n=45,157)
Yes
Do machine learning models (XGBoost and logistic regression) accurately predict in-hospital and 1-month post-discharge mortality in patients hospitalized with acute decompensated heart failure, and do models trained on one mortality endpoint transfer effectively to the other?
In patients hospitalized for acute decompensated heart failure, XGBoost models strongly predict in-hospital mortality (AUROC 0.85) and moderately predict 30-day post-discharge mortality (AUROC 0.70), with cross-endpoint transferability demonstrating partially shared prognostic signals but confirming the necessity of endpoint-specific models.
Answers are grounded in this paper and cite their sources.
Endpoint-specific models may enhance HF mortality prediction; leaves open cross-endpoint transferability and generalizability pending prospective validation.
Aims Patients with heart failure (HF) face substantial morbidity and mortality. This study aimed to build machine learning (ML) models for predicting in-hospital or post-discharge mortality in acute decompensated HF and explore whether models trained on one mortality endpoint transfer to the other. Methods Two retrospective HF cohorts were analyzed: 40,265 patient encounters from 15 hospitals in the United States (USA) with in-hospital mortality (3.6%) as the endpoint, and 4,892 encounters discharged alive from five hospitals in Greece (GRC) with 1-month all-cause mortality (9.2%) as the endpoint. ML models were trained using either all available features in each cohort (full models) or only variables shared between cohorts (restricted models), enabling both within-endpoint internal validation and cross-endpoint transferability assessment to be performed. Results During within-endpoint internal validation, discrimination was strong for the full USA models and moderate for the full GRC models, with extreme gradient boosting (XGBoost) reaching an area under the receiver operating characteristic curve of 0.85 and 0.70, respectively. Advanced age and elevated admission blood urea nitrogen were the most consistent predictors. In the cross-endpoint transferability analysis, restricted XGBoost models applied to the opposite cohort retained above-chance discrimination but were significantly outperformed on the same test set by the models developed for that cohort’s endpoint. Conclusion In two international cohorts of patients with acute decompensated HF, post-discharge mortality in the GRC cohort was more challenging to predict than in-hospital mortality in the USA cohort. XGBoost models developed for one mortality endpoint retained a degree of discriminative ability on the other endpoint, suggesting a partially shared prognostic signal between in-hospital and post-discharge mortality, while confirming the expected superiority of endpoint-specific models.
No takes yet. Share an insight, caveat, or question.
Kanavos et al. (2026) conducted a cohort in Acute decompensated heart failure (n=45,157). Machine learning (XGBoost and logistic regression) risk prediction models vs. Cross-endpoint transferred models and Get With The Guidelines-Heart Failure (GWTG-HF) risk score was evaluated on In-hospital mortality (USA cohort) and 1-month post-discharge all-cause mortality (Greece cohort) (AUROC 0.85 (USA in-hospital) and AUROC 0.70 (Greece 1-month post-discharge)). XGBoost models predicted in-hospital mortality in the USA cohort and 1-month post-discharge mortality in the Greek cohort, outperforming cross-endpoint transferred models.
Synapse has enriched 10 related papers. Consider them for comparative context: