Classification of 12-Lead Electrocardiograms using Convolutional Autoencoders and Transfer Learning

Sardar Ansari1, Christopher Gillies1, Kevin Ward1, Hamid Ghanbari2
1Emergency Medicine, University of Michigan, 2Internal Medicine-Cardiology, University of Michigan


Abstract

Introduction: Deep convolutional neural networks have gained popularity in the past decade and have been applied to several medical applications for signal and image processing. However, supervised training of a deep model requires a large labeled dataset. While unlabeled data is often abundant in most medical centers, labeling is often very time-consuming and burdensome. As a result, only a small subset of the data is usually annotated by clinicians for training.

Methods: We propose to use a hybrid approach, starting with unsupervised training of a convolutional autoencoder to learn the intrinsic features of the ECG signals, such as different presentations of the P, Q, R, S and T waves. The hidden layer of the autoencoder learns a compact and low-dimensional representation of the input ECGs by reducing the redundancy in the raw data and with minimal loss of information. The autoencoder is trained on a large dataset of ~2 million 12-lead ECGs from the University of Michigan’s Section of Electrophysiology. Once trained, the encoder acts as a transformation, mapping the high-dimensional input to a low-dimensional vector which can be used as input to a convolutional neural network or a conventional machine learning algorithm. Given the reduction in the size of the input space, the supervised training can be performed using a small dataset, such as the public dataset provided by the Challenge.

Results: The preliminary model was trained on our data and Physionet’s public dataset. Then it was tested on Physionet’s public dataset using a simple split approach resulting in an AUROC of 0.917, AUPRC of 0.791, accuracy of 0.961, F1-score of 0.801, F2-score of 0.8 and G2-score of 0.603. Testing on the hidden dataset yielded F2 and G2 scores of 0.765 and 0.545, respectively. These promising results demonstrate the potential for a solution combining convolutional autoencoders and transfer learning.