L2G-ECG: Learning to Generate Missing Leads in ECG Signals using Adversarial Autoencoder

An electrocardiogram (ECG) is a clinically accepted record used for the diagnosis of cardiovascular diseases. The standard guideline to monitor heart health is to employ 12-lead ECG signals. Yet, in certain trauma patients or ones with physical injury, it is not possible to place electrodes at prescribed locations. In this paper, we propose an adversarial learning-based method to generate multiple missing leads in ECG signals using a lesser number of available leads. A convolution neural network-based encoder-decoder-like network is trained to generate the missing lead signals, guided by a Visual Turing Test discriminator to enhance signal realism. We have experimentally demonstrated it on a publicly available Physionet Challenge 2021 dataset consisting of subjects from 3 different nationalities and genetic predispositions, including healthy individuals and those with multiple comorbidities across 29 different arrhythmia pathologies. We have exhaustively checked the ability to generate 12-lead ECG from 1,$\dots$, 7 available leads, with generalization assessed on a held-out test set. Quantitative analysis of the synthesized ECG signals is done by computing Mean Square Error (MSE), Structural Similarity Measure (SSIM), information-based metrics, and morphological feature comparison of real and generated Lead-II signals. Our study shows that as available leads increase from 1-7, average $AUP_{S}$ values improved from 0.81 to 0.98 and $AUP_{M}$ decreased from 30.96$\times 10^{-5}$ to 1.18$\times 10^{-5}$. Notably, 12-lead ECG reconstruction is achievable with just 2 limb and 3 chest leads, supported by low MSE, high SSIM, minimal information loss, and strong Lead-II feature generation, highlighting the model's real-world applicability.