Unsupervised Anomaly Detection For Identifying Arrhythmogenic Rhythms In Atrioventricular Block Hearts Using Deep Convolutional Autoencoders

Atrioventricular block (AVB) is a condition that can cause delayed or absent impulse conduction and increase the risk of arrhythmic complications such as atrial fibrillation (AF), premature ventricular contraction (PVC), and ventricular tachycardia (VT). Therefore, long-term monitoring of the heart's AVB rhythms and the ability to identify and predict abnormal rhythms is crucial for clinical decision-making in AVB patients. However, current tools to analyze irregular AVB rhythms and identify anomalies are limited. While supervised deep learning algorithms have been effective for anomaly detection during regular ordered sinus rhythm, they lack practicality for analyzing irregular AVB ECG rhythms. To address this challenge, we developed an unsupervised anomaly detection approach using deep convolutional autoencoder (CAE) models. We created three different CAE models consisting of either residual blocks (ResBlocks) or standard and transpose convolutional layers (TransConv1Ds) to identify anomalies in AVB ECG data. We acquired 30 days of continuous ECG data from seven porcine AVB hearts with complete AV node blockage and optimized the models accordingly. Both ResBlocks- and TransConv1Ds-based CAE models demonstrated high sensitivity and specificity to identify AF, PVC, and VT compared to a PCA-based traditional autoencoder model, with a true positive rate of 92-99% and a true negative rate of 85-92%. The ResBlocks-based CAE models showed high robustness across different ECG data acquired at different days and various porcine models compared to a TransConv1D-based model. As a case application of our anomaly detection approach, we applied the CAE to evaluate the efficacy and toxicity of TBX18-based biological pacemakers for treating AVB hearts. Our CAE analysis revealed that injecting these pacemakers at the ventricular high septum increased the occurrence of AF and PVC compared to the control group, suggesting that the rhythms from the biological pacemakers competed with the ventricular rhythms of the AVB heart. Moreover, the CAE demonstrated high sensitivity, specificity, and robustness in identifying anomalies from human AVB ECG rhythms, indicating the potential for unsupervised algorithms to improve AVB management and treatment.