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

Pierre Aublin1, Julien Oster1, Mouin Ben Ammar2, Jérémy Fix3, Michel Barret4
1INSERM, 2ENSTA, 3CentraleSupélec, 4International Research Lab Georgia Tech - CNRS IRL 2958 CentraleSupélec


Abstract

To tackle the 2021 PhysioNet/Cinc challenge, we (iadiecg team) proposed a deep learning model for the classification of a single-lead ECG signal. Decision on the reduced lead-set was taken as an average voting of the available single lead-based predictions. Single lead ECG signals were resampled at 250 Hz, bandpass filtered between 0.5 and 120Hz using a 3rd order Butterworth filter, and normalized (zero mean and unit variance). The neural network architecture consisted of 15 blocks, which include a one-dimensional convolutional layer, followed by rectified linear unit activation, batch normalization, and dropout layers. Between consecutive blocks, squeeze and excitation layers were introduced. A final global max pooling layer extracted 512 features for each signal, which were inputted in two fully connected layers with leaky rectified linear activation units. Training was performed by minimising a custom loss function, which combined a dice loss and binary cross entropy, using a Stochastic Gradient Descent with a cyclic update of the learning rate, with a Batch size of 64 over 50 epochs. A 5-fold cross-fold validation scheme was used for training and evaluating the model locally. Each fold contained recordings from all the databases, but were stratified by abnormality and subject (all leads from the same recording were included in the same fold). Using the challenge metric, an average score of 0.657, 0.643, 0.642, 0.639, 0.629 was obtained for the 12, 6, 4, 3, 2 lead datasets respectively for the cross validation, with the corresponding submitted entry scores of 0.586, 0.577, 0.574, 0.572, and 0.563. Final scores reflect a good level of performance for the detection of cardiac abnormalities, which could be further improved with optimised hyperparameters (in the loss) or by incorporating hand-crafted features or pre-training with a representation learning approach.