Evaluating ECGs and Diagnosing Various Heart Conditions via Convolutional Neural Networks

Overview

Electrocardiograms, or ECGs, are widely used technologies to record electrical signals from the heart to diagnose patients with various medical conditions. Currently, a shockingly high percentage of primary care physicians and ER workers are unable to accurately read ECGs. A review of 78 articles found that physician accuracy of ECG interpretation is at 54%. Additionally, many lower-income countries have restricted access to cardiologists, and cannot always afford accurate ECG readings. Opening access to cheaper ECG diagnoses would greatly benefit patients around the world. Machine learning is growing in popularity and is continuously being used to solve similar problems. In this study, a Convolutional Neural Network is trained on a dataset containing over 20,000 ECG tracings to determine whether a patient is sick or healthy, returning an F1 score of 0.823. In addition, the neural network was also trained to determine which class of condition the patient had: myocardial infarction, ST/T change, conduction disturbance, or hypertrophy. This study showed an accuracy of about 70%. With accuracies as high as 72%, future work can be done on these models to improve performance. These algorithms can be implemented in clinical settings to provide an accessible, affordable, and accurate method for evaluating ECGs. The results indicate that ECG analysis based on CNNs generalizes accurately to 12-lead exams, taking the technology one step closer to standard clinical practice.