Computational Paradigms for Heart Arrhythmia Detection: Leveraging Neural Networks

The electrocardiogram (ECG) captures a person’s heart activity and is, therefore, one of the best ways to diagnose heart arrhythmia. Artificial intelligence can quickly learn and process a huge amount of data, the properties of which could simplify the process of diagnosing any disease. In this work, we created 4 convolutional neural network models for determining 6 types of heart arrhythmias classes based on two kinds of heartbeat images of ECG signals. The goal was to investigate the effects of different types of input data and model structure on the success of the prediction. We worked with a publicly available MIT-BIH database of arrhythmias. The division of ECG records into heartbeats was realized using our R peak detection method, which was able to detect up to 97.39 % of all R peaks. Each heartbeat was imaged as a representation in the time domain and then in the time-frequency domain using a short-time Fourier transform. The best model, trained over time-domain representations, consisted of 5 convolutional and 3 pooling layers and achieved a classification accuracy of 98.5 %.