A Deep Learning Solution for Automatized Interpretation of 12-Lead ECGs

Alvaro Huerta Herraiz1, Arturo Martinez-Rodrigo2, José J Rieta3, Raul Alcaraz2
1Research Group in Electronic, Biomedical and Telecommunication Engineering, University of Castilla-La Mancha, Spain., 2University of Castilla-La Mancha, 3BioMIT.org, Universitat Politecnica Valencia


Background and Aim. A variety of algorithms for classification of rhythm abnormalities in ECGs have been proposed in last years. Although some have reported promising results, they have been mostly validated on short and nonpublic datasets, thus making their comparison difficult. Interestingly, PhysioNet/CINC Challenge 2020 provides an opportunity to compare these and other algorithms on a wide set of ECGs. The goal of this work is to introduce a novel method based on deep learning and beat-to-beat classification to participate in this competition.

Methods. Beats in all ECGs from the learning set were firstly segmented. Thus, R-peaks were detected from lead I and the mean RR interval was computed to define a window in which every single beat was get. Next, this window was applied to the remaining leads, consecutively concatenating single beats to generate a new signal, which was transformed into a 2D image through a continuous Wavelet transform. In this way, more than 130,000 images were obtained, which were used to train an 8-layer convolutional neural network (CNN) specifically designed to discern 9 different rhythm abnormalities. For an ECG-based validation, the probabilities of belonging to each class obtained by the CNN were averaged for all beats found in the recording. Finally, those classes presenting an average probability larger than 10% were assigned to the ECG.

Results. According to the Challenge guidelines, the algorithm was evaluated in terms of a class-weighted F-score and a generalization of the Jaccard measure. In average for all training signals, these metrics were 0.933 and 0.811, respectively. Regarding validation on the testing set, mean values for both performance indices were 0.654 and 0.372, respectively.

Conclusions. Although the proposed method has reported a notable ability to associate the dominant class to each ECG, detection of secondary abnormal rhythms in the recording should be improved.