Development and External Validation of an AI-ECG Algorithm for Estimating Elevated Serum NT-proBNP Levels

Background N-terminal pro–B-type natriuretic peptide (NT-proBNP) is a cornerstone biomarker for the diagnosis and management of heart failure, but its use may be limited by the need for blood testing and laboratory infrastructure. Artificial intelligence (AI) applied to electrocardiograms (ECGs) may offer a widely accessible, non-invasive approach to estimate NT-proBNP levels. Methods We developed a convolutional neural network incorporating residual and attention-based layers to estimate NT-proBNP levels from standard 12-lead ECGs. The model was trained using 84,895 ECG–NT-proBNP pairs from 40,762 adult patients; 8,545 patients were held out for internal validation. The model generated a nine-level ECG-BNP score. External validation was conducted in 679 patients at two tertiary cardiovascular centers. Discrimination for prespecified thresholds (>250, >500, >1,000 pg/mL) was assessed by AUROC with 95% CIs; calibration and threshold-specific sensitivity, specificity, PPV/NPV were evaluated. Results In internal validation, the AI-ECG score showed a strong correlation with measured NT-proBNP levels (Spearman ρ=0.85, p<0.001) and high discrimination across thresholds (AUROC >0.92). In external validation, the model achieved AUROCs of 0.866 (95% CI 0.838–0.894) for >250 pg/mL, 0.882 (95% CI 0.857–0.907) for >500 pg/mL, and 0.885 (95% CI 0.859–0.914) for >1,000 pg/mL. Performance was consistent across key clinical subgroups. Conclusions An AI-enabled ECG model can identify patients with elevated NT-proBNP levels with good accuracy in both internal and external validation cohorts. This approach may help identify patients who should undergo confirmatory NT-proBNP testing, particularly when biomarker testing is delayed, unavailable, or not routinely performed. Prospective studies are warranted to define whether this strategy provides incremental clinical value and can be integrated into clinical pathways.