ELECTROCARDIOGRAM CLASSIFICATION BY COMBINING WAVELET TRANSFORMS AND CONVOLUTIONAL NEURAL NETWORKS

The electrocardiogram (ECG) is a powerful tool for diagnosis and screening of many heart-related disorders which are of primary concern and thus need to be correctly diagnosed as early as possible. Its characteristic signal holds abundant information, primarily regarding electrical functioning but also related to functional and structural elements of the heart, some of which are evident just by visual inspection, but others are hidden, even to an expert eye. In this work, the wavelet transform was used to access the frequency domain and highlight potentially crucial features. It is well suited for non-stationary signals with dynamic components like the ECG while keeping at the same time a good temporal resolution, differently from the more common Fourier transform for example. Once obtained the transformed ECG representations, called scalograms, an artificial intelligence algorithm was developed for the automated classification of pathological signals. In particular, a convolutional neural network (CNN), well suited for images like scalograms, used supervised deep learning to classify data without explicitly selecting features. The data were retrieved from the publicly available datasets provided by PhysioNet, containing almost one-hundred-thousand labelled real-world ECG, collected through four different contents. Due to the extremely unbalanced nature of the classes representing a realistic distribution of heart-related disorders, the model was trained only on a limited selection of four of them, while the multi-label problem of having more than one disorder associated with some of the signals was simplified to be only multi-class. The end-to-end architecture was built to input one scalogram for each lead as an image channel in a hyperspectral fashion; doing so, it was also possible to design four more models with a reduced number of leads, respectively with six, four, three, and two leads. These models were cross-validated and achieved similar performances in terms of accuracy (86% average), precision (83% average), and sensitivity (82% average) across the different classes and models. These results show the validity of the approach for ECG classification, in addition to a certain redundancy in the different lead's information with the five models almost superimposable. Overall, wavelet transform provides a useful representation of the ECG, which can be combined with CNN to effectively diagnose different disorders, even when a reduced number of leads is used.