On Modeling Electrocardiograms With Correlated Leads Using Deep Learning Methods

This paper presents an innovative mathematical model for generating 12lead electrocardiograms (ECGs), based on a fundamentally novel approach to accounting for spatial dependencies between leads. The primary scientific novelty lies in the development of a linear transformation method applied to a set of physiologically grounded basis signals representing projections of the heart’s electrical field, augmented with correlated noise that precisely mimics real clinical artifacts. Unlike traditional generative models (VAE, GAN), which operate as «black boxes," our model provides explicit control over the morphology of key waves (P, QRS, T) and strictly enforces physiological constraints, including Kirchhoff’s laws for limb leads. This ensures anatomical consistency across all 12 leads—a feat previously unachieved in similar studies. The model demonstrated high performance on the PhysioNet PTBXL dataset: MSE = 0.015, cosine similarity = 0.94, F1score = 0.88 for normal rhythms and 0.82 for arrhythmias. A key advantage of the model is its computational efficiency (generation time of 50 ms) and relatively low memory requirements (2.5 GB). Comparative analysis against stateoftheart generative models (VAE, GAN, Diffecg) confirmed the superiority of the proposed model in terms of interpretability, parameter control, and physiological plausibility of synthesized signals.