Automatic prediction of arrhythmogenic right ventricular cardiomyopathy in electrocardiographic body surface mapping using Deep Learning

Arrhythmogenic right ventricular cardiomyopathy (ARVC) is a hereditary heart muscle disease, causing predisposition to ventricular arrhythmia and sudden cardiac death. The early detection of the disease and signs of progression is of great importance and can save lives. This work focuses on data processing, using Deep Learning (DL) to evaluate and analyze electrocardiographic signal collected through an innovative body surface mapping system on a small dataset of 40 patients, in order to formulate a criteria for the diagnosis of ARVC. Each patient underwent 10 minutes of recording with a novel non-invasive electrocardiographic body surface mapping (ECG-BSM) technique, that uses a 252-unipolar leads vest, and enables the recording of multiple electrograms from the whole thorax and a reconstruction of epicardial potentials. The main tasks of the DL techniques investigated in this work are the automation of the QRS dispersion detection, which is being researched to be one of the criteria for ARVC diagnosis, and the development of a classifier for the three types of patients: ARVC patients, gene carriers without a history of arrhythmia or structural cardiac changes and healthy controls. State-of-the-art automated ECG analysis methods, in particular classical methods based on traditional signal processing techniques, are limited in their application and are relatively difficult to implement as expertise in both signal processing and cardiology are required. Methods based on DL, on the other hand, have achieved unprecedented performance in initial studies in single-lead and 12-lead conventional ECG systems. These methods are able to learn to classify the information directly from the raw data and may not require expertise, once trained. Both these attributes and the ability for prediction of mortality even from normal ECGs suggest that these end-to-end models might be able to identify additional and even subtle electrocardiographic markers and abnormalities that are of interest and that are not being captured by traditional analysis methods, especially in the context of ECG-BSM techniques that may capture heart features that may not be visible on classical 12-lead ECG recording. The Deep Neural Network method used in this work is based on residual networks (ResNet18 and ResNet34) architectures, which have been successfully used for ECG abnormality detection in previous works. The first model developed in this work formulates a regression problem that takes as input a one-channel signal window that contains one heartbeat and targets the QRS interval duration. Whereas the second model aims at classifying the patients by taking as input a window of 252-channel signal. The results of the study show that, while the first model is not able to predict the QRS interval duration with a lead-by-lead analysis, the second model can classify with high accuracy the ARVC and control patients and with a good accuracy the gene-carrier patients. These findings demonstrate that an end-to-end deep learning approach has a high diagnostic performance similar to that of cardiologists and could potentially reduce the mortality of the ARVC disease by early diagnosis.