Semi-Supervised Learning for ECG Classification

Abstract

We present an approach for automatic cardiac abnormality detection using two leads ECG. This approach was developed in the context of the Physionet/Computing in Cardiology Challenge 2021. Our model is decomposed into an Encoder and a Decoder. It is a huge neural network model with more than 36 million parameters. Although the Challenge training dataset consists of more than 88 thousand annotated ECGs, our model is extremely prone to overfitting to the training data. The encoder is a convolution neural network followed by three transformer encoder blocks. The decoder is a transformer encoder block followed by a feed forward neural network. To reduce the overfitting, we pretrain the Encoder in a semi-supervised way on three tasks. Given an ECG segment, L1, the first task is to detect the QRS on L1; the second task is to predict the ECG shape on an ECG segment, L2 following L1, given the QRS location on $L_{2}$; the third task is to predict the number of samples, after $L_{1}$ , before the next QRS. The Decoder weights were firstly estimated with the frozen Endoder pre-trained parameters and then the whole model parameters were fine-tunned. Our team, named matFCT, received a challenge score of 0.43 on the official test dataset. However, we were unable to qualify for ranking because we weren't able to submit the preprint to the Computing in Cardiology Conference before the deadline.