PointSenseNet: A Hierarchical Point-Cloud-Based Reconstruction Framework with Application to Localized Inverse...

Cardiac excitation mapping based on intracardiac ECG is essential for identifying arrhythmogenic substrates in persistent atrial fibrillation, yet existing methods are limited by spatial resolution and the requirement of electrode-tissue contact. This study aims to develop a high-resolution localized non-contact cardiac mapping method.

We propose a hierarchical point-cloud deep learning model, PointSenseNet (PSN), designed to address sparse signal reconstruction problems exemplified by ECG mapping. Using a cardiac electrophysiology simulation model, we evaluated PSN across multiple non-contact distances and compared it with the conventional inverse ECG mapping method. In addition, we investigated the effectiveness of the virtual intermediate point cloud layer incorporated into PSN.

Across various excitation patterns, PSN demonstrated improved predictive accuracy and spatial resolution compared with the conventional method. At a non-contact distance of 10 mm, PSN successfully reconstructed critical structures such as conduction block regions and achieved a relative error (normalized membrane potential) below 10% and activation time error below 6 ms. The inclusion of the intermediate point cloud layer further improved performance across different non-contact distances.

PSN demonstrates promising performance for non-contact ECG mapping, achieving higher predictive accuracy than that of current methods. Owing to its strong generalization capability, PSN also holds potential for broader application to sparse signal reconstruction tasks beyond ECG mapping.

These findings have the potential to substantially reduce the procedural burden associated with contact-based mapping in future clinical applications while providing more precise mapping information that may facilitate the development of novel ablation strategies.