Toward Patient-Specific Prediction of Ablation Strategies for Atrial Fibrillation Using Deep Learning

• DOI: 10.3389/fphys.2021.674106
• Publication Date: 2021-05-26T06:53:33Z
• Authors: Marica Muffoletto; Ahmed Qureshi; Aya Zeidan; Laila Muizniece; Xiao Fu; Jichao Zhao; Aditi Roy; Paul A. Bates; Oleg Aslanidi
• Keywords: Physiology; computational modelling; 610; deep learning; patient imaging; 620; 3. Good health; 03 medical and health sciences; 0302 clinical medicine; catheter ablation; QP1-981; atrial fibrillation; classification algorithm

Atrial fibrillation (AF) is a common cardiac arrhythmia that affects 1% of the population worldwide and associated with high levels morbidity mortality. Catheter ablation (CA) has become one first line treatments for AF, but its success rates are suboptimal, especially in case persistent AF. Computational approaches have shown promise predicting CA strategy using simulations atrial models, as well applying deep learning to images. We propose novel approach combines image-based computational modelling atria classifiers trained on patient-specific which can be used assist therapy selection. Therefore, we convolutional neural network (CNN) combination (i) 122 tissue images obtained by unfolding patient LGE-MRI datasets, (ii) 157 additional synthetic derived from data enhance training dataset, (iii) outcomes 558 terminate several AF scenarios corresponding models. Four CNN were this dataset balanced techniques predict three strategies images: pulmonary vein isolation (PVI), rotor-based (Rotor) fibrosis-based (Fibro). The accuracy these ranged 96.22 97.69%, while validation was 78.68 86.50%. After training, applied an unseen holdout test set images, results compared respective simulations. highest rate observed correct prediction Rotor Fibro (100%), whereas PVI class predicted 33.33% cases. In conclusion, study provides proof-of-concept networks learn MRI datasets image-derived models providing technology tailoring patient.