Multi-channel delineation of intracardiac electrograms for arrhythmia substrate analysis using implicitly regularized convolutional neural network with wide receptive field

Abstract

Objective Automated segmentation of intracardiac electrograms and extraction of fundamental cycle length intervals is crucial for reproducible arrhythmia substrate analysis conducted during electrophysiology procedures. The objective of this study is to develop a robust, computationally efficient end-to-end model for the precise electrogram multi-channel delineation using a highly imbalanced dataset. Methods A temporal deep convolutional neural network (CNN) based on the UNet architecture incorporating convolutional layers of varying dilation rates was implicitly regularized through data augmentations (DAs), a domain specific Tversky loss function, and distinct labelling strategies for segments comprising atrial fibrillation (AF). An exploratory study utilizing Bayesian search was conducted to optimize architectural and loss function hyperparameters. The impact of dilated convolutions, data augmentations, and labelling strategies on the performance and generalization capability was assessed through an ablation study. The performance of different models was evaluated using a cross-validation procedure and two independent test datasets derived from two separate patient cohorts containing 326, 84, and 97 electrograms encompassing sinus rhythms, abnormal complexes during ongoing tachycardias, and stimulation protocols. Results A UNet model with optimized loss hyperparameters, a dilated receptive field, and atrial fibrillation (AF) annotated as a positive class (D-UNet-L) achieved an average Srensen-Dice coefficient (SDC) of 84.9% on recordings with regular atrial beats across test datasets, surpassing the performance of models without loss optimization (81.5%), without dilated kernels (81.3%), and with inversed AF labelling (77.5%). Notably, the highest average accuracy (Acc) of 95.8% for AF recordings was obtained by a model trained on negatively assigned AF segments, outperforming D-UNet-L (88.9%), the model without loss optimization (81.5%), and the model without dilated kernels (81.3%). The reference D-UNet-L model exhibited overall root-mean-square errors of 8.3 and 9.0ms across test datasets. Additionally, 61.5% and 20.8% of delineations exhibited absolute errors below 5ms and 10ms, respectively. Disabling data augmentation (DAs) resulted in a 2.7% decrease in validation SDC and a 5.3% increase in training SDC.“ Conclusion Generalization capability across independent datasets was improved by employing exponentially weighted Tversky loss. The model's segmentation performance on longer sequences with atrial fibrillation was improved by incorporating dilated convolution kernels. Noise-aware and morphology data augmentations effectively mitigated overfitting potential in a limited training dataset. Label noise introduced by annotating atrial fibrillation sequences into a positive class strengthened regularization of the model, particularly in its ability to identify regular beats. However, it also negatively impacted performance on F-waves.Objective Automated segmentation of intracardiac electrograms and extraction of fundamental cycle length intervals is crucial for reproducible arrhythmia substrate analysis conducted during electrophysiology procedures. The objective of this study is to develop a robust, computationally efficient end-to-end model for the precise electrogram multi-channel delineation using a highly imbalanced dataset. Methods A temporal deep convolutional neural network (CNN) based on the UNet architecture incorporating convolutional layers of varying dilation rates was implicitly regularized through data augmentations (DAs), a domain specific Tversky loss function, and distinct labelling strategies for segments comprising atrial fibrillation (AF). An exploratory study utilizing Bayesian search was conducted to optimize architectural and loss function hyperparameters. The impact of dilated convolutions, data augmentations, and labelling strategies on the performance and generalization capability was assessed through an ablation study. The performance of different models was evaluated using a cross-validation procedure and two independent test datasets derived from two separate patient cohorts containing 326, 84, and 97 electrograms encompassing sinus rhythms, abnormal complexes during ongoing tachycardias, and stimulation protocols. Results A UNet model with optimized loss hyperparameters, a dilated receptive field, and atrial fibrillation (AF) annotated as a positive class (D-UNet-L) achieved an average Srensen-Dice coefficient (SDC) of 84.9% on recordings with regular atrial beats across test datasets, surpassing the performance of models without loss optimization (81.5%), without dilated kernels (81.3%), and with inversed AF labelling (77.5%). Notably, the highest average accuracy (Acc) of 95.8% for AF recordings was obtained by a model trained on negatively assigned AF segments, outperforming D-UNet-L (88.9%), the model without loss optimization (81.5%), and the model without dilated kernels (81.3%). The reference D-UNet-L model exhibited overall root-mean-square errors of 8.3 and 9.0ms across test datasets. Additionally, 61.5% and 20.8% of delineations exhibited absolute errors below 5ms and 10ms, respectively. Disabling data augmentation (DAs) resulted in a 2.7% decrease in validation SDC and a 5.3% increase in training SDC.“ Conclusion Generalization capability across independent datasets was improved by employing exponentially weighted Tversky loss. The model's segmentation performance on longer sequences with atrial fibrillation was improved by incorporating dilated convolution kernels. Noise-aware and morphology data augmentations effectively mitigated overfitting potential in a limited training dataset. Label noise introduced by annotating atrial fibrillation sequences into a positive class strengthened regularization of the model, particularly in its ability to identify regular beats. However, it also negatively impacted performance on F-waves.