An Attention Based Neural Architecture for Arrhythmia Detection and Classification from ECG Signals

Arrhythmia is ubiquitous worldwide and cardiologists tend to provide solutions from the recent advancements in medicine. Detecting arrhythmia from ECG signals is considered a standard approach and hence, automating this process would aid the diagnosis by providing fast, cost-efficient, and accurate solutions at scale. This is executed by extracting the definite properties from the individual patterns collected from Electrocardiography (ECG) signals causing arrhythmia. In this era of applied intelligence, automated detection and diagnostic solutions are widely used for their spontaneous and robust solutions. In this research, our contributions are two-fold. Firstly, the Dual-Tree Complex Wavelet Transform (DT-CWT) method is implied to overhaul shift-invariance and aids signal reconstruction to extract significant features. Next, A neural attention mechanism is implied to capture temporal patterns from the extracted features of the ECG signal to discriminate distinct classes of arrhythmia and is trained end-to-end with the finest parameters. To ensure that the model’s generalizability, a set of five train-test variants are implied. The proposed model attains the highest accuracy of 98.5% for classifying 8 variants of arrhythmia on the MIT-BIH dataset. To test the resilience of the model, the unseen (test) samples are increased by 5x and the deviation in accuracy score and MSE was 0.12% and 0.1% respectively. Further, to assess the diagnostic model performance, AUC-ROC curves are plotted. At every test level, the proposed model is capable of generalizing new samples and leverages the advantage to develop a real-world application. As a note, this research is the first attempt to provide neural attention in arrhythmia classification using MIT-BIH ECG signals data with state-of-the-art performance.