Background Heart rate variability (HRV), an indicator of autonomic nervous system activity, is widely recognized for assessing heart failure prognosis. N-terminal pro-B-type natriuretic peptide (NT-proBNP) plays a crucial role in the diagnosis, risk stratification, and prognosis evaluation of heart failure, but its measurement relies on invasive procedures and lacks dynamic monitoring capabilities.

Objective To explore the correlation between wearable HRV parameters and NT-proBNP levels and validate their feasibility as a non-invasive risk stratification tool.

Methods The 24-hour wearable device physiological data and clinical indicators were collected from acute heart failure patients admitted to the Department of Cardiology, West China Hospital, Sichuan University, from January 2021 to December 2022. Post preprocessing, HRV parameters were extracted, and a threshold of 3 500 pmol/L was established for NT-proBNP levels. Spearman correlation was used to analyze the association between HRV and NT proBNP, and the efficacy of five machine learning models including Logistic Regression, K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Random Forest and XGBoost to predict poor prognosis (NT proBNP ≥ 3 500 pmol/L) was evaluated, and the characteristic contribution was quantified by the snap value.

Results A total of 51 heart failure patients were included, with 87 electrocardiographic datasets collected. Among the HRV parameters, SDNN, SD2, VLF, and 7 other indexes showed significant negative correlations with NT-proBNP levels (r: -0.390 to -0.371, P < 0.001). Patients with NT-proBNP≥3 500 pmol/L had significantly lower HRV parameters, including SDNN (MIQR: 51.10 38.50-67.20 ms vs 77.95 54.45-95.50 ms, P < 0.001), SD2 (MIQR: 68.30 52.90-93.90 ms vs 108.00 76.20-132.47 ms, P=0.003), VLF (MIQR: 18.82 5.84-59.61 mHz vs 59.36 33.70-116.90 mHz, P=0.002), ULF (MIQR: 6.30 1.99-18.02 mHz vs 18.60 10.05-34.09 mHz, P=0.001). Among the machine learning models, Logistic Regression demonstrated the best classification performance (AUC=0.830, 95% CI: 0.760 - 0.890). SHAP analysis revealed that SD2 and LF/HF contributed the most to the model's classification.

Conclusion Wearable HRV parameters are significantly correlated with NT-proBNP levels and can effectively differentiate high and low NT-proBNP groups using machine learning models, offering a new strategy for non-invasive dynamic monitoring and acute decompensation risk prediction in heart failure patients.