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

Area of Science:

• Cardiovascular Electrophysiology
• Artificial Intelligence in Medicine
• Signal Processing

Background:

• Reducing noise in electrophysiological (EP) signals is critical for accurate cardiac diagnosis, mapping, and ablation.
• Traditional denoising methods often fall short, impacting clinical decision-making.

Purpose of the Study:

• To evaluate the efficacy of generative AI, specifically a β-Variational Autoencoder (β-VAE) model, in denoising intra-ventricular monophasic action potential (MAP) signals.
• To compare the performance of the β-VAE model against traditional denoising techniques.

Main Methods:

• A β-VAE model was trained on 5706 time series of intra-ventricular MAP signals from patients with ischemic cardiomyopathy.
• The model's denoising performance was assessed against various noise types, including EP noise, and compared to established baseline methods.

Main Results:

• The β-VAE model achieved superior denoising performance, with a Pearson's Correlation of 0.967 ± 0.009, compared to the best baseline at 0.879 ± 0.022.
• The model effectively reduced diverse noise types, notably EP noise, in single cardiac beats.
• Generative AI learned essential signal features without manual annotation, outperforming current state-of-the-art techniques.

Conclusions:

• Generative AI, particularly the β-VAE model, offers a powerful solution for eliminating noise in EP signals.
• This technology can enhance diagnostic accuracy and treatment efficacy for various heart rhythm disorders, especially in complex cases like arrhythmias.