Transformer-based Dual-ECG model classified heart failure status changes with AUROC 0.89 and accuracy 87%, correlating BNP prediction with r=0.77.
Does a Transformer-based deep learning model analyzing paired 12-lead ECGs accurately classify heart failure status changes in adult patients?
A novel AI-driven dual-ECG model can accurately detect longitudinal changes in heart failure status, offering a potential non-invasive alternative to serial BNP testing.
Absolute Event Rate: 0% vs 0%
Abstract Background Heart failure (HF) hospitalizations are associated with poor prognosis, underscoring the necessity for early detection and longitudinal monitoring. Deep learning (DL) models have demonstrated utility in detecting HF from electrocardiograms (ECGs); however, their ability to track HF status changes over time remains uncertain. A scalable, non-invasive, AI-driven methodology may facilitate improved HF management across diverse clinical settings. Purpose This study aimed to develop a Transformer-based DL model capable of analyzing paired 12-lead ECGs recorded at two distinct time points to classify HF deterioration, improvement, or no change, thereby providing an advanced tool for HF monitoring. Methods This retrospective study encompassed 30,171 ECGs from 6,531 adult patients in a single-center cohort, randomly allocated into training (70%), validation (15%), and test (15%) datasets. HF deterioration was defined as a ≥40% increase in BNP with a follow-up BNP ≥100 pg/mL. Improvement was defined as a ≥40% decrease in BNP with a baseline BNP ≥100 pg/mL. Cases that did not meet these criteria, including those with baseline BNP 100 pg/mL despite a ≥40% decrease, were classified as no change. Thresholds were established based on clinical relevance and prior evidence(1). A Transformer-based DL model (Dual-ECG HF Trajectory Model, DEHT) was developed to evaluate temporal ECG waveform variations and detect HF status changes. Model interpretability was enhanced using Gradient-weighted Class Activation Mapping (Grad-CAM) to identify salient ECG features informing predictions. Results The cohort had a mean age of 64.6 ± 15.4 years, a median BNP of 66.3 pg/mL (IQR: 24.6–175.1), and a mean left ventricular ejection fraction of 59.2% ± 12.2%. The proposed model achieved high accuracy in classifying HF status changes (AUROC: 0.89, 95% CI: 0.88–0.90; accuracy: 0.87, 95% CI: 0.86–0.88), demonstrating robust performance even in patients with HFrEF or BNP ≥100 pg/mL. Furthermore, the model effectively estimated BNP levels from a single 12-lead ECG, with a correlation coefficient of 0.77 (95% CI: 0.74–0.79). Grad-CAM analysis identified QRS complex alterations, particularly changes in QRS duration and amplitude, as principal determinants of model predictions. Conclusions The Dual-ECG deep learning model demonstrated potential in detecting HF status changes, offering a non-invasive adjunct to BNP testing. By leveraging temporal ECG dynamics, this approach has the potential to support continuous HF monitoring and clinical decision-making. Preliminary external validation on an independent dataset is in progress, with further validation planned through multicenter trials. Future directions include integration into clinical workflows and development of AI-driven ECG monitoring systems for real-time, home-based HF management, with implications for reducing hospital readmissions.Dual-ECG AI Model for HF Progression Performance of DEHT model
Nishihara et al. (Sat,) reported a other. Transformer-based Dual-ECG model classified heart failure status changes with AUROC 0.89 and accuracy 87%, correlating BNP prediction with r=0.77.