Predicting success of atrial fibrillation ablation: comparing machine learning approaches of intracardiac electrograms

Abstract

Abstract Introduction Machine learning (ML) and Deep learning (DL) neural networks have been used to predict which patients will respond to atrial fibrillation (AF) ablation. However, the predictive results of DL are modest compared to traditional statistical models of clinical features with or without intracardiac electrograms. Purpose To test the hypothesis that optimized ML architectures will better identify patients who respond to AF ablation using intracardiac EGM than other DL architectures in our large registry of 320 consecutive patients. Methods Our patients had age 65.1±10.4Y, were 25% women, 61.6% non-paroxysmal, with intracardiac EGM recorded from multipolar catheters and 64-pole baskets (Fig A). Patients were propensity matched into those in whom ablation terminated by ablation (N=160, "Term") and those without termination (N=160, "Non-Term"). Using a proven DL architecture, ResNet, we systematically varied the depth (i.e., number of residual layers (D)) and width (i.e., number of nodes per layer (W)). We predicted termination (Fig C), presenting EGMs to training cohorts and testing in separate cohorts in a 6-fold cross-validation which we repeated 10 times for each fold with different randomization. Results Fig B illustrates that the prediction accuracy is generally negatively correlated with the network depth but positively correlated with the width. Specifically, the deep architecture (D=4, W=256, with 265M trainable parameters, blue-circled in Fig B) predicted with AUC = 0.61±0.08. However, a shallow-narrow architecture (D=1, W=16, green-circled in Fig B) achieved same AUC (0.61±0.09, p=0.65), while reducing the number of parameters by 97% compared to the DL architecture (79K vs. 265M). Additionally, a shallow-wide architecture (D=1, W=2048, red-circled in Fig B) achieved higher AUC (0.74±0.08; p<0.0001), while saving 44% of the parameters (148M vs 265M) and 26% training time compared to the deep architecture. Conclusion Shallow neural network classification of raw EGMs predicted ablation response with better accuracy than other architectures. Future work should determine if deep architectures ultimately ‘overfit’ detailed EGM data, which may explain their lower success.Figure