Abstract 4366733: Deep Learning-Based BNP Classification from Single-Lead ECG

Background: Brain natriuretic peptide (BNP) is a key heart failure biomarker. Single-lead electrocardiograms (ECGs) from wearable devices offer valuable diagnostic and prognostic insights. We developed a deep learning model to predict BNP levels from single-lead ECGs for rapid, non-invasive screening. Methods: Using a large dataset of 528,654 first-lead (Lead I) ECG images from 237,686 unique patients at Houston Methodist Health System, each paired with a corresponding BNP value obtained within 24 hours, we trained a modified VGG16 convolutional neural network (CNN). The dataset included 236,160 records with BNP <100 pg/mL, 105,382 with BNP 100–399 pg/mL, 41,618 with BNP 400–899 pg/mL, and 39,782 with BNP ≥900 pg/mL. The model was designed for multilabel stratification into four clinically relevant categories: <100, 100–399, 400–899, and ≥900 pg/mL. Data were split by unique patients into training/validation (n=131,974) and testing (n=105,712) cohorts. We also evaluated the association between ECG-predicted BNP categories and incident heart failure (HF) in a subset of patients without preexisting HF. Results: In the testing dataset, the model demonstrated robust performance, with a macro-average AUC of 0.87 ± 0.03 across all classification thresholds using Youden's J statistic. For the prediction of severe BNP elevation (≥900 pg/mL), the model achieved a sensitivity of 0.83 and a specificity of 0.73. The overall accuracy for distinguishing normal (<100 pg/mL) from elevated BNP levels was 0.76. Among 6,278 patients without prior HF, ECG-predicted BNP 100–399 was associated with a hazard ratio (HR) of 2.10 (95% CI: 1.74–2.53), BNP 400–899 with an HR of 2.58 (1.92–3.45), and BNP ≥900 with an HR of 3.92 (3.39–4.54) for incident HF independently of age/sex (all compared to BNP <100 (Figure)). A model including age/sex/ECG-BNP (c-index 0.74 [0.73-0.76]) outperforms age/sex/BNP (c-index 0.70 [0.68-0.71]) for prediction of incident HF (P<0.001). Conclusion: Our novel deep learning-facilitated ECG-BNP model has high discrimination of BNP categories using single-lead ECG tracings. ECG-BNP outperforms blood BNP measurements in predicting HF. This approach offers a scalable, non-invasive tool for HF risk stratification and monitoring that could extend diagnostic access in both clinical and remote care settings, including integration with consumer-grade devices such as smartwatches that capture Lead I ECGs.