An Automatic 12-Lead ECG Interpreter Based on Neural Network Trained on Combination of Signal, Physiological and Encoded Features

Dishant Beniwal, Divyansh Srivastava, Pooja Aggarwal, Ashish Sahani
IIT Ropar


Interpretation of 12-lead ECG is key to diagnosis of electrical abnormalities of the heart. However, manual interpretation of the ECG signals is time-consuming, and requires skilled personnel with a high degree of training. Conventional signal processing techniques have limited success in this. In this paper, we explore convolutional neural networks trained on combined signal features, physiological features, encoded features to make more accurate diagnosis. For this we used the data provided under the CINC 2020 challenge which has ECG of different lengths varying from 6 s to 60 s with nine labels. The data set is imbalanced in terms of samples available per label.

First we extract 6 s data frames by using a moving window with length of 6 second and overlap of 4 s. This creates a total of 60000 data sets. A balanced subset with 2500 sets per label is picked up for training. This data is high-pass filtered and normalized between -1 and 1. An autoencoder is trained to extract 20 bottleneck features. Following ECG features are extracted - R-R interval, PR interval, QT interval, P-R amplitude ratio, R-T amplitude ratio and heart rate variability. Signal features namely - power in 1-20 Hz, 20-50 Hz, 50-100 Hz, 100-150 Hz, 150 Hz - 250 Hz, skew, kurtosis, complexity measures are extracted. All these features are then used to train a multi layer neural network.

A model, incorporating partial implementation of the above approach, was trained on provided data which produced following cross-validation scores: AUROC (0.905), AUPRC (0.871), Accuracy (0.978), F-measure (0.904), Fβ (0.914), Gβ (0.777). Implementation of the same model on the hidden test set of the competition produced the following scores: Fβ (0.554) and Gβ (0.337) with a geometric mean of (0.432).