A CatBoost machine learning model using 14 variables — including B-type natriuretic peptide (BNP) — achieved 81% accuracy in predicting three-year mortality in patients with both heart failure and atrial fibrillation.
AUC 0.809CatBoost model accuracy for predicting three-year death risk in heart failure patients with atrial fibrillation, with BNP among the top 5 predictors
What the researchers found
The CatBoost machine learning model achieved the highest AUC of 0.809 for predicting three-year all-cause mortality in HF-AF patients, outperforming five other ML approaches. Of 558 patients, 215 (38.5%) reached the primary endpoint of death during median follow-up of 1,185 days.
SHAP analysis identified the top five predictive features: NYHA heart failure classification, absolute lymphocyte count (ALC), high-sensitivity C-reactive protein (hs-CRP), B-type natriuretic peptide (BNP), and age. Notable feature interactions were found between lymphocyte count and NYHA class, and between lymphocyte count and BNP, suggesting these biomarkers provide complementary prognostic information.
Why it matters
Heart failure with atrial fibrillation is extremely common and deadly, but predicting which patients face the highest risk is difficult using traditional clinical scores alone. BNP — a peptide biomarker already widely measured in heart failure — was confirmed as one of the strongest predictors, and its interaction with lymphocyte count suggests that combining peptide biomarkers with immune markers could improve risk stratification beyond what either achieves alone.
How the study worked
Retrospective cohort study of 558 HF-AF patients admitted in 2018 with median follow-up of 1,185 days. The Boruta algorithm and LASSO regression selected 14 key variables. Six ML models were trained using 10-fold cross-validation on a 70/30 train-test split, optimized via grid search, and evaluated across 12 performance metrics. SHAP analysis was used for model interpretation and feature interaction analysis.
What this study cannot tell us
This was a single-center retrospective study with a moderate sample size (558 patients), which limits generalizability. External validation in an independent cohort was not performed. The model was trained on 2018 admission data, and treatment patterns may have changed since then. The study did not compare the ML model's performance against established clinical risk scores (like CHA₂DS₂-VASc) head-to-head.
How to read the evidence
This is a single-center retrospective cohort study with internal validation only. While the machine learning methodology is rigorous, the lack of external validation and moderate sample size are significant limitations for clinical implementation.
When this study was published
Published in 2025, this is a very current study reflecting the growing application of machine learning to clinical risk prediction in cardiology.
The bigger picture
BNP and its related peptide NT-proBNP are already the cornerstone biomarkers in heart failure diagnosis and monitoring. This study reinforces BNP's prognostic value while showing that machine learning can extract more predictive power from routine clinical data — including interactions between BNP and immune markers — than traditional risk scores. As ML-based clinical decision tools become more common, peptide biomarkers like BNP will likely play an even larger role in personalized cardiovascular care.
Questions still open
- Does this CatBoost model maintain its predictive accuracy when validated in external cohorts from different hospitals and healthcare systems?
- Would adding serial BNP measurements over time improve prediction accuracy compared to the single baseline measurement used here?
- Could the BNP-lymphocyte interaction identified by SHAP analysis reflect a specific pathophysiological mechanism that could be therapeutically targeted?
Common questions
What is BNP and why is it important in heart failure?
How could this prediction model help patients?
Read the original research
Prediction of three-year all-cause mortality in patients with heart failure and atrial fibrillation using the CatBoost model.
BMC cardiovascular disorders, 25(1), 466
Citation
Wu, Jiacan; Tao, Guanghong; Xie, Siyuan; Yang, Han; Qi, Fenglin; Bao, Naiyue; Li, Zhuo; Chang, Guanglei; Xiao, Hua. (2025). Prediction of three-year all-cause mortality in patients with heart failure and atrial fibrillation using the CatBoost model.. BMC cardiovascular disorders, 25(1), 466. https://doi.org/10.1186/s12872-025-04928-w