Transfer Learning for Improved Classification of Drivers in Atrial Fibrillation.

"Drivers" are theorized mechanisms for persistent atrial fibrillation. Machine learning algorithms have been used to identify drivers, but the small size of current driver datasets limits their performance. We hypothesized that pretraining with unsupervised learning on a large dataset of unlabeled electrograms would improve classifier accuracy on a smaller driver dataset. In this study, we used a SimCLR-based framework to pretrain a residual neural network on a dataset of 113K unlabeled 64-electrode measurements and found weighted testing accuracy to improve over a non-pretrained network (78.6±3.9% vs 71.9±3.3%). This lays ground for development of superior driver detection algorithms and supports use of transfer learning for other datasets of endocardial electrograms.

Effects of Biventricular Pacing Locations on Anti-Tachycardia Pacing Success in a Patient-Specific Model.

Patients with drug-refractory ventricular tachycardia (VT) often undergo implantation of a cardiac defibrillator (ICD). While life-saving, shock from an ICD can be traumatic. To combat the need for defibrillation, ICDs come equipped with low-energy pacing protocols. These anti-tachycardia pacing (ATP) methods are conventionally delivered from a lead inserted at the apex of the right ventricle (RV) with limited success. Recent studies have shown the promise of biventricular leads placed in the left ventricle (LV) for ATP delivery. This study tested the hypothesis that stimulating ATP from multiple biventricular locations will improve termination rates in a patient-specific computational model. VT was first induced in the model, followed by ATP delivery from 1-4 biventricular stimulus sites. We found that combining stimulation sites does not alter termination success so long as a critical stimulus site is included. Combining the RV stimulus site with any combination of LV sites did not affect ATP success except for one case. Including the RV site may allow biventricular ATP to be a robust approach across different scar distributions without affecting the efficacy of other stimulation sites. Combining sites may increase the likelihood of including a critical stimulus site when such information cannot be ascertained.

Deep Learning Based Prediction of Atrial Fibrillation Disease Progression with Endocardial Electrograms in a Canine Model.

OBJECTIVE: We sought to determine whether electrical patterns in endocardial wavefronts contained elements specific to atrial fibrillation (AF) disease progression. METHODS: A canine paced model (n=7, female mongrel, 29±2 kg) of persistent AF was endocardially mapped with a 64-electrode basket catheter during periods of AF at 1 month, 3 month, and 6 months post-implant of stimulator. A 50-layer residual network was then trained to map half-second electrogram samples to their source timepoint. RESULTS: The trained network achieved final validation and testing accuracies of 51.6 and 48.5% respectively. Per class F1 scores were 24%, 59%, and 53% for 1 month, 3 month, and 6 month inputs from the testing dataset. CONCLUSION: Differentiation of AF based on its time progression was shown to be feasible with a deep learning method. This is promising for differentiating treatment based on disease progression though low accuracy with earlier timepoints may be an obstacle to identifying nascent AF.