Poster U11 - Deep Learning-Enhanced Cardiac Microphysiological Systems: Studying In Vitro and In Vivo Reentry Arrhythmia

Introduction: Arrhythmia can be caused by two major tissue-level mechanisms: reentry and focal activities, which disrupt the normal sinus rhythm of the heart. Understanding these mechanisms and their contributions to arrhythmias in individual patients is critical for developing effective treatments. To this end, patient-specific in vitro human models, based on induced pluripotent stem cell-derived cardiomyocytes (iPSC-CMs), have been developed to replicate these arrhythmogenic activities and identify pathophysiological mechanisms specific to individual patients. Moreover, recent advancements in in vitro and in vivo electrophysiology (EP) mapping technology have enabled the capturing of patient-specific arrhythmogenic mechanisms, generating a vast amount of EP videos. However, current methods and systems for analyzing these EP videos are limited in their ability to perform high-throughput analysis, and interpretations of clinical EP videos can be subjective and vary among clinicians.

Materials and

Methods: To address these limitations, we introduce cEP-Net, a novel deep learning system designed to address the limitations in analyzing cardiac EP activities. The system consists of a cardiac microphysiological model to replicate reentry and focal activities in vitro and a deep learning-based EP video classifier. The classifier features a two-stream architecture, including spatial and temporal streams, each equipped with a two-hidden-layer two-dimensional CNN framework and three fully connected layers. This design enables efficient analysis of EP videos, categorizing arrhythmogenic activities into four types: no or localized activities, normal linear propagation, focal activity, and reentrant activity.

The preprocessing pipeline begins with raw EP image frames acquired by an Optical Mapping System (OMS). These frames are enhanced through normalization, noise thresholding, and average-pooling, followed by optical flow to highlight calcium reentrant wave propagation. We trained cEP-Net using over 111,000 annotated EP videos generated by our cardiac microphysiological system and evaluated its performance in both in vitro and in vivo settings.

In vitro testing assessed cEP-Net's capability to capture reentry formation and evolution in cardiac tissues from the first day after seeding. The classifier also detected transient focal activities in tissues expressing the pacemaker ion channel HCN4. Additionally, cEP-Net was used to assess Flecainide-induced vulnerability of human cardiac tissues to reentry. For in vivo applications, cEP-Net was tested to detect EP propagation singularities in patients with atrial fibrillation.

Results, Conclusions, and Discussions: Our pretrained cEP-Net demonstrated impressive performance metrics, achieving a high Kappa statistic of 91.0% and a concordance rate of 94.7% across all categories, indicating the system's robust reliability in classifying electrophysiological activities. We successfully visualized in vitro EP videos using UMAP, clustering them into four distinct categories, which further validated the classifier's effectiveness. cEP-Net proved capable of capturing critical phenomena such as reentry development and evolution during early in vitro cardiac tissue formation, as well as the initiation of focal activities in tissues expressing HCN4, a pacemaker protein.

Additionally, we leveraged the pretrained cEP-Net to quantify the arrhythmogenic risks posed by cardiac agents. Notably, cEP-Net effectively identified Flecainide-induced vulnerability to reentry. Flecainide significantly shifted the vulnerable window to a longer action potential duration and increased the width of the vulnerable window for reentry, consequently raising the probability of arrhythmia. These results underscore its potential in drug safety testing.

Beyond in vitro applications, we applied cEP-Net to analyze in vivo real-time electroanatomical mapping EP videos, focusing on phase singularity identification in atrial fibrillation (AF) hearts. The system accurately captured the electrophysiological patterns characteristic of persistent AF, demonstrating its utility in clinical settings for understanding and managing complex arrhythmias. Using cEP-Net, we analyzed the differences between hearts with persistent AF and those with paroxysmal AF. Hearts with persistent AF displayed a higher number of both focal activities and reentries in the pulmonary vein and the posterior box of the left atrium compared to hearts with paroxysmal AF.

Overall, cEP-Net excels in high-fidelity classification and visualization of EP activities, offering valuable insights into the mechanisms of arrhythmogenesis and the impact of therapeutic interventions. Using a transfer learning algorithm, cEP-Net successfully captured in vivo abnormalities. This comprehensive approach highlights cEP-Net's potential as a powerful tool for both research and clinical applications in cardiac electrophysiology.

Acknowledgements (Optional): This research has been performed with the support of the American Heart Association Career Development Award #857583, a grant from National Heart Lung and Blood Institute 1R01HL161366-01A1, and cooperative research funds provided by SUMMIT Co., LTD, and the Korea Research Institute of Chemical Technology. S.-J.P. was funded by Georgia Institute of Technology and Emory University School of Medicine.