Machine Learning Model Improves Prediction of Heart Failure Risk in CKD

Researchers of a new multicenter study appearing in the Journal of the American Heart Association have developed and validated a machine learning–based model that significantly improved the prediction of heart failure (HF) risk in patients with chronic kidney disease (CKD), offering a potential tool for earlier intervention in a high-risk population.¹

Heart failure is one of the most serious and common cardiovascular complications among patients with CKD, driven by overlapping biological mechanisms between the heart and kidneys.² Despite this elevated risk, current HF prediction models are largely derived from the general population and often fail to accurately assess risk in patients with CKD.

To address this gap, the researchers created a model they say may be able to predict HF using just 9 variables. While the group noted that that further validation in broader populations is needed before widespread adoption, they highlighted the potential of the model to offer a more precise and tailored approach to identifying patients at risk of HF. Their model has been translated into a web-based calculator designed for clinical use, allowing physicians to estimate a patient’s 5-year risk of heart failure using easily accessible data inputs.

To create their model, the researchers leveraged data from more than 52,000 patients with CKD in the China Renal Data System, then validated the model in nearly 22,000 patients from independent Chinese hospitals and more than 3,000 participants from the UK Biobank, providing a diverse, international validation cohort.

The study focused on predicting the 5-year risk of new-onset HF. The researchers evaluated multiple machine learning approaches, including random forest, neural networks, support vector machines, and decision trees. Among these, an extreme gradient boosting (XGBoost) model consistently demonstrated the strongest performance.

In internal validation, the XGBoost model achieved an area under the curve (AUC) of 0.905, indicating high predictive accuracy. This performance remained strong in external validation cohorts, with an AUC of 0.879 in Chinese populations and 0.851 in the UK Biobank cohort.

Importantly, the model outperformed traditional risk scores, including the widely used Atherosclerosis Risk in Communities (ARIC) model. Existing models showed declining accuracy as kidney function worsened, highlighting their limitations in CKD populations. By contrast, the machine learning approach maintained consistent performance across varying levels of kidney function.

To enhance clinical usability, researchers simplified the model to include just 9 key variables, all routinely available in clinical practice. They included age, kidney function measures such as estimated glomerular filtration rate (eGFR) and urine albumin-to-creatinine ratio (ACR), cardiac biomarkers like NT-proBNP, and clinical history factors such as coronary heart disease, atrial fibrillation, and medication use.

“Our study identified age, preexisting cardiac disease, and the use of antihypertensive medication as significant predictors of the risk of incident HF,” wrote the researchers. “These findings align with established predictors in the ARIC-HF risk model, which emphasizes age, hypertension, and cardiovascular comorbidities as critical contributors to HF pathogenesis.”

The importance of these variables reflects the interconnected nature of heart and kidney disease, noted the group. For example, reduced eGFR and elevated albuminuria signal worsening kidney damage and are strongly associated with increased HF risk. Similarly, elevated NT-proBNP levels indicate cardiac stress and may detect early, subclinical heart dysfunction before symptoms emerge.

Visualization tools used in the study, including Shapley additive explanations (SHAP), further illustrated how these variables contribute to individual risk. Higher levels of NT-proBNP and albuminuria, older age, and pre-existing cardiovascular disease all significantly increase predicted HF risk, while better kidney function is associated with lower risk.

References

1. Lu Y, Chen J, Zhou S, et al. Development and validation of a machine-learning model for incident heart failure prediction in chronic kidney disease: a multicenter cohort study. J Am Heart Assoc. Published online March 18, 2026. doi:10.1161/JAHA.125.046156

2. Xanthopoulos A, Papamichail A, Briasoulis A, et al. Heart failure in patients with chronic kidney disease. J Clin Med. 2023;12(18):6105. doi:10.3390/jcm12186105