Background: Ablation procedures targeting Localized Atrial Tachycardia (AT) can be drastically facilitated if the origin of the abnormality is located based on Electrocardiogram (ECG) signals. Previous work has shown that P-wave morphology can help make such a prediction. This study aims to develop a Machine Learning algorithm that detects the likely origin of Localized AT based on features extracted from P-waves and signals recorded in the Coronary Sinus (CS).
Methods: A total of 240 rhythms recorded during ablation procedures were studied. For every rhythm, P-waves were detected in 9 of the 12 ECG leads using a Deep Learning model based on the U-net architecture. Wave polarity was determined by comparing it to templates, and the most frequent polarity in every lead was retained as a feature for the given rhythm. Other wave morphology characteristics, such as peak magnitude and area under the curve, and the activation delays of the 4 CS probes, were also considered. A machine learning algorithm was trained using the resulting data to classify rhythms into areas in the heart corresponding to the focus of the AT. Feature selection techniques were implemented to reduce the size of the dataset and model hyperparameters were tuned using Bayesian Optimization.
Results: 4 types of anatomic divisions of the heart were tested: Left, Right, Septal and Left Lateral. For each division, a model was trained on the 240 rhythms and tested on a separate dataset of 40 rhythms. Our model gave a Kappa Score of 0.55 and an overall accuracy of 74%.
Conclusion: This work lays the foundation for future endeavors in the field of machine learning applied to cardiology. ECG and CS signals can be analyzed to determine the origin of Localized AT. Such predictive models can be deployed in the operating rooms to assist physicians.