Hybrid Signature-based Approach to ECG Classification

Hamid Khandahari and Adam Mahdi
University of Oxford


Abstract

Introduction. The electrocardiogram (ECG) is key in monitoring the heart's electrical activity and identifying problems with the heart rhythm. It can help to tell if an individual is having a heart attack or had a heart attack in the past and thus facilitating the choice of an appropriate treatment. However, manual interpretation of the ECG requires expert knowledge and is time-consuming. The PhysioNet/Computing in Cardiology Challenge 2020 asks to design an algorithm to automatically identify cardiac abnormalities using 12-lead ECG recording.

Methods. The initial training set comes from China Physiological Signal Challenge 2018 and consists of 6,877 (male 3,699, female 3,178) 12-lead ECG recordings of various length (6-60 seconds), sampled at 500 Hz. Each ECG recording has at least one label from a set of nine possibilities including normal sinus rhythm and eight abnormal rhythms such as atrial fibrillation or first-degree atrioventricular block. In this early development of the algorithm we considered only lead II, which usually gives a good view of the P-wave and is commonly used to record the rhythm strip. As part of the preprocessing, the ECG signals were standardised (by making them zero-mean and unit standard deviation) and made of equal 10,000 sample length (either through slicing or concatenating). Our model was an 18-layer, 1-dimensional convolutional neural network with max-pooling and dropout to provide regularisation.

Results and Conclusion. Our initial and unique entry had an F2 score of 0.509 and G2 score of 0.301 resulting in the geometric mean of 0.391. Next, we aim to develop a hybrid model combining deep learning with a signature-based regression model. Recall that the signature can be thought of as a collection of summary statistics that determine a path constructed from patients time-series (here based on ECG signals).