Deep Learning-Based Multi-Lead ECG Reconstruction from Lead I with Metadata Integration and Uncertainty Estimation

We propose a dual-branch deep learning framework for reconstructing standard 12-lead electrocardiograms (ECGs) from a single-lead input. The model integrates waveform information from Lead I ECG signals with clinically interpretable metadata to enhance reconstruction fidelity and introduces predictive uncertainty estimation to improve interpretability and reliability. A publicly available dataset of 10,646 ECG records was utilized. The model combined Lead I signals with clinical metadata through two processing branches: a CNN-BiLSTM branch for time-series data and a fully connected branch for metadata. Monte Carlo dropout was applied during inference to generate uncertainty estimates. Reconstruction performance was evaluated using Pearson's correlation coefficient and root mean square error. Metadata consistently contributed to performance improvements, particularly in the QRS complexes and T-wave segments, and the proposed framework outperformed U-Net when metadata were included. Predictive uncertainty showed moderate to strong positive correlations with reconstruction errors, especially in the chest leads, and heatmaps revealed waveform regions with reduced reliability in arrhythmic and morphologically atypical cases. To the best of our knowledge, this is the first study to incorporate predictive uncertainty into ECG reconstruction. These findings suggest that combining waveform data with metadata and uncertainty quantification offers a promising approach for developing more trustworthy and clinically useful wearable ECG systems.