Electrocardiogram-Based Abnormal Heartbeat Classification: A Deep Learning Approach for Arrhythmia Detection

By: Aditya Kendre

Heart arrhythmias are irregular rhythms in heartbeats that affect 3 million people worldwide every year. Due to the increasing rate of ECGs recording for diagnosis, it is now possible to devolve autonomous AI driven systems to identify arrhythmias in ECGs. A convolutional neural network (CNN) was developed and trained, to achieve accuracy on par or better than cardiologists in identifying arrhythmias in ECGs. The CNN not only surpassed the accuracy of cardiologists in identifying atrial fibrillation, but also achieved an overall top accuracy of 99%, and a constant accuracy of 96%.

The key to achieving such success is due to the large annotated dataset (PhysioNet), and data augmentation techniques. Originally, training a shallow CNN with few parameters were thought to create less complexly in learning, and make the CNN faster in training. Doing that merely did the opposite, the model did not learn fast, as the CNN started to overfit to the training data. Adding data augmentation not only fixed the issue of overfitting, but also increased the dataset size; conversely, this increased the time the CNN took to train.

Leave a Reply