Can Generative AI Learn Physiological Waveform Morphologies? A Study on Denoising Intracardiac Signals in Ischemic Cardiomyopathy.

Abstract

Reducing electrophysiological (EP) signal noise is essential for diagnosis, mapping, and ablation, yet traditional approaches are suboptimal. This study tests the hypothesis that generative artificial intelligence (AI), specifically Variational Autoencoders (VAEs), can effectively denoise these signals by forming robust internal representations of 'clean' signals. Utilizing a dataset of 5706 time series from 42 patients with ischemic cardiomyopathy at risk of cardiac sudden death, we set out to apply a β-VAE model to denoise and reconstruct intra-ventricular monophasic action potential (MAP) signals, which have verifiable morphology. The β-VAE model is evaluated against various noise types, including EP noise, demonstrating superior denoising performance compared to traditional methods (Pearson's Correlation of denoised vs original of 0.967 ± 0.009 for our proposed model vs 0.879 ± 0.022 for the best performing baseline). Results indicate that the model effectively reduces a wide array of noise types, particularly EP noise. We conclude that generative AI provides powerful tools that can eliminate diverse sources of noise in single beats by learning essential signal features without manual annotation, outperforming state-of-the-art denoising techniques.Clinical Relevance- The proposed β-VAE model's ability to effectively denoise and reconstruct intracardiac signals, particularly in the challenging context of arrhythmias, can significantly enhance diagnostic accuracy across a variety of heart rhythm disorders and improve treatment efficacy.