Classification of mechanisms underlying cardiac arrhythmias by deep learning

Abstract

This work presents the development and comparison of classifiers based on both traditional classifiers and more advanced deep learning techniques to identify different types of bioelectric propagation in cardiac tissue, which are the mechanisms underlying arrhythmias. The classification uses energy maps obtained from the magnetic field generated by three possible bioelectric models of propagations through simulated cardiac cells (classes), consisting of a linear path, a reentrant circuit around an inexcitable obstacle, and a spiral trajectory. Different classifiers were implemented and compared in regard to their accuracies, aiming to associate each energy map to its respective class. These classifiers were evaluated both on a test set composed by images without noise and on a more challenging “noise test set,” composed by images meant to test the generalization powers of the models in the presence of noise. A Monte Carlo cross-validation approach was used to evaluate the averaged behavior of the classifiers over 10 runs. The classifiers compared include conventional decision support tools as decision trees and random forests, a simple multi-layer perceptron, and deep learning architectures such as a simple Convolutional Neural Network, ResNet50, and InceptionV3. These last two had the best performances on the noise test set, with averaged accuracies of 0.852 and 0.827 over 10 runs, and standard deviations of 0.036 and 0.048, respectively. The use of handcrafted features, specifically histograms of oriented gradients, was also shown to substantially improve results for the classifiers that do not use convolutions, to the point of surpassing the results of the deeper neural networks. The comparative performance of the studied classifiers elucidates some of the properties of this classification problem. ResNet50 and InceptionV3 seem to be promising candidate models for initial tests with real-life data. The study of handcrafted features might also prove useful, especially if the availability of real images is scarce. The use of this automated classification may contribute to the efficacy of diagnostics in a future use of biomagnetic mappings in a medical-hospital environment.