All-in-one electrical atrial substrate indicators with deep anomaly detection

Background and Objective: Electrophysiological studies based on high-density catheter mapping have become a cornerstone in ablation procedures for managing atrial fibrillation. Since their introduction, extensive efforts have been devoted to analyzing intracardiac atrial electrograms (EGMs) to identify potential arrhythmogenic regions, often employing machine learning techniques. The aim is to investigate the potential of deep anomaly detection algorithms as either complements or replacements for established electrophysiological indicators used to characterize EGMs associated with arrhythmic substrate. Methods: We investigated three deep anomaly detection algorithms. Model outputs were used to generate consistent and robust scores for each signal in a completely unsupervised manner. We applied these techniques to 8 patients with a leave-one-out strategy. Results: Our numerical experiments show that: (i) higher anomaly scores are correlated with higher EGM fractionation and duration and lower voltage, (ii) thresholding anomaly score percentiles and standard indicator values produce consistent classifications, and (iii) morphology analysis is more robust compared to a stratification provided by single standard indicator, without the need for determining arbitrary thresholds. Conclusions: Our results demonstrate the effectiveness and robustness of deep anomaly detection algorithms in the characterization of anomalous cardiac EGMs. By providing an all-in-one method to assess pathological features, these NN models eliminate the limitations that arise from manually combining and visually comparing traditional indicators. Our electro-anatomical maps displaying anomaly scores could have significant implications for improving the accuracy and efficiency of ablation procedures aimed at managing cardiac arrhythmias, such as atrial fibrillation.