Key result
A stacked ensemble model using routine EMR variables identified heart failure with an ROC-AUC of 0.927, PR-AUC of 0.895, and accuracy of 0.856.
Why the study?
Early diagnosis of HF is challenging because specific symptoms are often ignored or misinterpreted, leading to delayed diagnosis. This study evaluated whether routinely available EMR variables can identify patients with existing HF.
Does a calibrated stacked ensemble machine learning model using 18 routinely collected EMR variables accurately identify existing heart failure and stratify risk in US adults compared to individual ML models and a baseline clinical model?
Observational (n=37,070)
Yes
Does a calibrated stacked ensemble machine learning model using 18 routinely collected EMR variables accurately identify existing heart failure and stratify risk in US adults compared to individual ML models and a baseline clinical model?
Answers are grounded in this paper and cite their sources.
A calibrated stacked machine learning model using 18 routine EMR variables can identify existing heart failure with high discrimination (ROC-AUC 0.927) without requiring specialized imaging or advanced biomarkers, offering a scalable approach for population screening.
Background: Heart failure (HF) is one of the most prevalent causes of morbidity, mortality, and healthcare expenditures, with approximately 7.7 million adults in the U.S. living with this condition. Early diagnosis of high-risk individuals has been a challenge, as HF-specific symptoms are often ignored or misinterpreted as normal aging, stress, or minor illnesses, leading to delayed diagnosis. This study evaluated whether routinely available electronic medical record (EMR) variables can identify patients with existing HF. Methods: We trained, tested, and evaluated several models, including logistic regression, SVM, KNN, random forest, XGBoost, MLP, and a custom stacked ensemble using stratified 5-fold cross-validation and 70/30 hold-out split for HF identification on EMR data from the All of Us Research Program. This group consisted of 37,070 adults (13,577 HF; 23,493 non-HF). The predictors included readily available demographics, vital signs, laboratory values, common conditions, lifestyle, and a deprivation index. Preprocessing steps included IQR-based winsorization, median imputation, one-hot encoding, and QuantileTransformer. Predicted probabilities were calibrated and adjusted to an assumed 2.5% prevalence. Permutation importance was used to rank features. Results: The stacked model obtained an ROC-AUC of 0.927 (95% CI 0.922–0.932), a PR-AUC of 0.895, and an accuracy of 0.856 in the test set, similar to XGBoost alone (0.929) and above an age + sex + atrial fibrillation baseline (0.808). At 2.5% prevalence, the top 10% of predicted risk contained 74.7% of HF cases. Atrial fibrillation, age, hypertensive disorder, anemia, and obesity were the most influential features. A secondary multiclass experiment (No-HF, HF with reduced ejection fraction, and HF with preserved ejection fraction) showed lower discrimination results (macro-AUC 0.91) with limited recall for HFrEF (0.37) and HFpEF (0.50). Conclusions: Routine EMR variables can identify existing HF with high discrimination. External and prospective validation is needed before clinical use.
No takes yet. Share an insight, caveat, or question.
Ali et al. (2026) conducted an observational in Heart Failure (n=37,070). Stacked ensemble machine learning model using 18 routinely collected EMR variables vs. Individual ML models (XGBoost, Random Forest, SVM, Logistic Regression, KNN, MLP) and an age + sex + atrial fibrillation baseline model was evaluated on Binary identification of existing heart failure in the 30% hold-out test set (ROC-AUC 0.927, 95% CI 0.922-0.932). A stacked ensemble model using routine EMR variables identified heart failure with an ROC-AUC of 0.927, PR-AUC of 0.895, and accuracy of 0.856.
Synapse has enriched 10 related papers. Consider them for comparative context: