Automated Methods for Classification and Digitization of ECG Images from CVD Patients

Electrocardiography (ECG) is an essential tool for understanding the electrophysiology of the heart and its electrical activity in response to cardiovascular diseases (CVD). However, a physical paper-based format is common, which poses difficulties in the presence of millions of patients’ data, reducing global accessibility in cardiac assessment. We propose a simple approach to digitize physical ECGs (images or papers) and a novel deep learning to classify the CVD condition (normal or abnormal) using a convolutional neural network (CNN). Digitization simply involves removing the background and reading pixels on an image mask to identify a signal. On the other hand, the CNN includes feed-forward blocks to train on automatically extracted image features. Our team, VitalRhythms, achieved a challenge classification score (F-measure) of 0.694 during the unofficial phase and an overall 10-fold cross-validation score of 0.766. Moreover, the reconstruction score (SNR) was -18.10, with a 10-fold cross-validation SNR of -13.20, but our official phase training submissions were unsuccessful. This study paves the way towards implementing sophisticated deep learning tools for the purpose of digitizing paper-based ECG and aiding the assessment of cardiovascular diseases and, thus, simplifying cardiac care in the presence of big patient data.