Cardiac Abnormality Detection Based on an Ensemble Voting of Single-Lead Classifier Predictions

• Main Authors: Aublin, Pierre, Ammar, Mouin Ben, Achache, Nina, Benahmed, Melina, El Hichami, Aymane, Barret, Michel, Fix, Jeremy, Oster, Julien
• Format: Conference Proceeding
• Language: English
• Subjects: Band-pass filters; Computational modeling; Computer architecture; Convolution; Deep learning; Electrocardiography; Neural networks
• ISSN: 2325-887X
• DOI: 10.23919/CinC53138.2021.9662824

We developed a fully deep learning model to identify cardiac abnormalities from ECGs for the PhysioNet/CinC 2021 Challenge. Decision on different lead subsets was based as an average voting of all available single-lead predictions. ECG signals were bandpass filtered between 0.5 and 120 Hz, resampled at 250 Hz, cropped to 10 seconds and normalized (zero-mean, unit-variance). The neural network architecture consisted of fifteen blocks. Most blocks consisted in one-dimensional convolution followed by rectified linear unit activation, batch normalization, and dropout layers. Twelve blocks also contained a squeeze and excitation module. A global max pooling layer allowed for the extraction 512 features for each signal. Those features were inputted in fully connected MLP with two hidden layers with leaky rectified linear activation and linked to the outputs through a sigmoid activation. Our team (iadi-ecg) obtained scores of 0.48, 0.47, 0.47, 0.47, 0.46 on the twelve, six, four, three, two lead versions of the hidden challenge test set, resulting in final ranking between the 11 th, and 12th out of 39 teams). The suggested approach had difficulties to generalize well on the hidden test set, and future works will aim at developing an hybrid model, as we assume that hand-crafted features might help for generalization purpose. The proposed technique demonstrated its ability to classify ECGs even when only two leads were available.