Deep learning formulation of electrocardiographic imaging integrating image and signal information with data-driven...

Area of Science:

• Biomedical Engineering
• Computational Cardiology
• Artificial Intelligence in Medicine

Background:

• Electrocardiographic imaging (ECGI) non-invasively maps heart electrical activity using body surface potentials (BSP).
• The inverse problem in ECGI is mathematically ill-posed, posing significant challenges for accurate cardiac electrical mapping.
• Advancements in artificial intelligence offer potential solutions to overcome ECGI's inherent difficulties.

Purpose of the Study:

• To propose a deep learning (DL) formulation for ECGI to learn the statistical relationship between BSP and cardiac activation.
• To develop a novel ECGI method using Conditional Variational Autoencoders and deep generative neural networks.
• To evaluate the accuracy and efficacy of the proposed DL-based ECGI method.

Main Methods:

• Simulated cardiac activation maps and BSP data on six distinct cardiac anatomies.
• Trained a Conditional Variational Autoencoder model on 5000 activation maps across five cardiac anatomies.
• Tested the model on a new patient anatomy with 200 activation maps, predicting 10 distinct activation maps per BSP data point.

Main Results:

• The DL-based ECGI method successfully generated volumetric cardiac activation maps with high accuracy on simulated data.
• Achieved a mean absolute error of 9.40 ms with a standard deviation of 2.16 ms on the testing dataset.
• The probabilistic nature of the model allowed for the prediction of multiple distinct activation maps, reflecting data uncertainty.

Conclusions:

• The proposed DL formulation of ECGI effectively integrates imaging information for estimating cardiac electrical activity from BSP.
• The method inherently accounts for spatio-temporal correlations within the data, enhancing accuracy.
• This AI-driven approach holds promise for improving the clinical utility and diagnostic power of ECGI.