Regularised Encoder-Decoder Architecture for Anomaly Detection in ECG Time Signals

Abstract

Electrocardiography(ECG) is a procedure to record the electrical activity of the human heart. The recorded time series ECG signal are often used by medical professionals to detect any arrhythmia the subject may have. Work has been done to automate the task by modelling the problem as anomaly detection using encoder-decoder based techniques, training on just normal data and using distribution of loss to predict normal or abnormal data. We argue that normal encoder-decoder with just reconstruction loss suffers from two problem: 1. Latent vector is not smooth and continuous, which might lead to memorising signals 2. Network is prone to outliers as mean squared error is used for reconstruction loss. We propose a regularised encoder-decoder based architecture with KL divergence as regulariser for latent vector which solves the above two problem. The regulariser will enforce the network to minimise the distance between latent vector distribution and normal distribution, hence making latent vector smooth and continuous, at the same time as diverse as possible. We have experimented with various architectures such as Multilayer Perceptrons, Recurrent Neural Networks, Long Short Term Memory Networks, 1D Convolutional Neural Networks for encoder and decoder and found that regularised network outperforms normal network in all the cases. More specifically our regularised network with F1-score of 0.90 outperformed the current state of art for ECG anomaly detection which uses Long Short Term Memory networks for both encoder and decoder, resulting in F1-score of 0.88. As a result we present a regularise encoder-decoder network in this paper which outperforms current techniques for anomaly detection on ECG data.