From one to twelve: feasibility and clinical utility of deep learning-derived 12-lead ECGs for remote cardiac monitoring

Background Cardiovascular diseases (CVDs) remain the leading cause of global mortality. While the 12-lead electrocardiogram (ECG) is the clinical gold standard, wearable devices typically provide only single-lead monitoring (Lead I), which lacks the spatial resolution required to detect complex pathologies such as myocardial infarction or ST-segment changes. Objective This study proposes a deep learning framework capable of reconstructing a full 12-lead ECG from a single-lead input while preserving diagnostic integrity. Methods We developed TONet (Transformer-based Optimization Network), a multi-task framework evaluated on the PTB-XL dataset ( n = 21,837). The architecture integrates a Convolutional Neural Network (CNN) encoder to extract local morphological features and a Transformer encoder to capture global temporal dependencies across the cardiac cycle. To ensure clinical validity, the model was trained using a dual-objective loss function, simultaneously optimizing for signal reconstruction (MSE) and multi-label diagnostic classification (BCE) across five diagnostic super-classes. Results TONet achieved an overall Pearson Correlation Coefficient (PCC) reaching 0.673 on the independent test set. Clinical agreement assessment via Bland-Altman analysis revealed a negligible per-lead bias (|Bias| ≤ 0.0024 mV across all 11 reconstructed leads), indicating no systematic voltage drift. In terms of diagnostic utility, the reconstructed 12-lead signals achieved a macro-averaged AUROC of 0.821 (95% CI: 0.810–0.831), with per-class AUROCs of 0.874 (NORM), 0.770 (MI), 0.862 (STTC), 0.782 (CD), and 0.815 (HYP). To assess generalisability beyond a single cohort, we additionally performed an independent in-distribution validation on the Chapman-Shaoxing 12-lead ECG database ( n = 43,559 quality-controlled recordings), where TONet, retrained from scratch using the identical protocol, achieved an overall PCC of 0.612 (95% CI: 0.552–0.666) and substantially outperformed the linear-regression baseline (PCC = 0.498), confirming that the reconstruction benefit transfers to a second, hardware- and population-distinct cohort. Conclusion TONet demonstrates that a hybrid CNN–Transformer network with a multi-task reconstruction-plus-classification objective can recover morphologically faithful 12-lead surrogates from a single Lead I drawn from clinical PTB-XL recordings. We position this work as a regularized deep-learning framework for multi-lead reconstruction rather than a deployable clinical device: the reconstructed precordial leads are insufficient for fine-grained ST-elevation myocardial infarction (STEMI) screening, and external validation on true wearable recordings remains essential.