Extrapolation of Ventricular Activation Times From Sparse Electroanatomical Data Using Graph Convolutional Neural...

Area of Science:

• Computational electrophysiology
• Artificial intelligence in medicine
• Cardiac electrophysiology

Background:

• Electroanatomic mapping is crucial for ventricular tachycardia assessment but faces challenges in achieving high resolution.
• Current interpolation methods struggle to reconstruct complete biventricular activation times.
• Accurate and dense mapping is essential for effective tachycardia treatment.

Purpose of the Study:

• To investigate the use of graph convolutional neural networks (GCNNs) for estimating biventricular activation times from sparse electroanatomic measurements.
• To develop an AI-driven method that overcomes limitations of traditional mapping techniques.
• To improve the accuracy and efficiency of cardiac activation mapping.

Main Methods:

• Training a GCNN on over 15,000 synthetic ventricular depolarization patterns generated by a computational model.
• Utilizing diverse anatomical geometries and simulated physiological conditions (scar, conduction velocity variations).
• Validating the GCNN on independent synthetic data and experimental datasets (porcine and swine hearts).

Main Results:

• The GCNN accurately reconstructed biventricular activation times in synthetic data with a mean absolute error (MAE) of 3.9 ms ± 4.2 ms at 1 sample/cm².
• Experimental data showed MAE < 10 ms, even with limited input measurements.
• A model-guided measurement strategy reduced data requirements by 40% while maintaining accuracy.

Conclusions:

• GCNNs offer a powerful tool for estimating biventricular activation times, surpassing traditional methods in accuracy and speed.
• The developed AI approach significantly enhances electroanatomic mapping for ventricular tachycardia.
• Real-time uncertainty estimation enables optimized data acquisition, improving clinical workflow efficiency.