Standard 12-lead ECG acquisition requires 10 electrodes whose correct placement is technically demanding, prone to poor skin contact, motion artefacts, and inadvertent displacement. Those factors limit routine use especially in ambulatory, mobile cardiac monitoring settings whose goal is to perform continuous arrythmia detection and classification. We hypothesize that reducing electrode count from 10 to 7 (removing 3 precordial electrodes) and synthesizing the 3 resulting missing leads (V2, V4, V5) with generative deep learning techniques can preserve downstream arrhythmia classification performance.
Using the PhysioNet 2020 Challenge dataset, a 7-electrode configuration provides 5 directly measured leads plus 4 derived leads; the remaining 3 precordial leads were reconstructed via (1) a Conditional Generative Adversarial Network (GAN) with adversarial plus L1 loss, and (2) a Diffusion model with state-space backbone (200 timesteps). The reconstructed 12-lead signals feed a state-of-the-art arrhythmia classifier for a 27-cardinality multi-label diagnosis. Performance is measured using the PhysioNet Challenge metric, as well as AUROC, accuracy, and F1-score compared against real leads and zero-filled baselines.
Our results show how, at optimal configurations (200 timesteps for Diffusion), both generative approaches significantly outperformed zero-filling. Diffusion achieved a Challenge metric of 0.424 (AUROC 0.947, Accuracy 0.961, F1 0.508), while GAN reached 0.397 (AUROC 0.942, Accuracy 0.960, F1 0.490), versus 0.307 for zero-filled baseline. Real leads achieved 0.476. Crucially, GAN inference was 925× faster (0.18s vs 166.56s per batch), enabling real-time deployment. While Diffusion offered higher fidelity, GAN provided a superior speed-accuracy trade-off for clinical applications.
Generative models substantially improve arrhythmia classification over naive zero-filling when leads are reduced. The choice between GAN and Diffusion depends on application constraints: GAN enables real-time monitoring with modest accuracy loss, while Diffusion provides high-fidelity results when latency is acceptable. This study paves the way for future work on reducing diffusion latency, improving generation accuracy, and randomizing missing leads.