Sparse representation based ECG heartbeat classification using convolutional neural networks

The cardiovascular disorders can be monitored using electrocardiogram (ECG or EKG). Since the interpretation of arrhythmias disorders is difficult by human, it is of great importance to develop automated models to help physicians classify arrhythmias in a real-time and accurate manner. The classification of ECGs signals is challenging as there are similarities among ECG heartbeats within inter-patient and intra-patient recordings, in addition to the noise resulted from ECGs' apparatus and settings. This paper presents a new model, which uses sparse representations of ECG signals and classifies them using a Convolutional Neural Network. An orthogonal matching pursuit method is implemented to produce appropriate sparse-represented signals based on an overcomplete Gabor dictionary. The sparsified (SR-ed) inputs are fed into the proposed network which consists of 5 successive double 1-D convolutional layers to classify 5 ECG categories. In contrast to traditional algorithms, this method does not require preprocessing such as feature extraction or selection. The MITBIH Arrhythmia database is used to evaluate the performance of the developed model. The model is tested using non-SR-ed, SR-ed and compressed SR-ed (CSR-ed) data for both imbalanced and balanced classes datasets. The experimental results show that the model accuracies of the SR-ed data reach up to 97.44% and 97.32% for the imbalanced and balanced datasets respectively. While for the non-SR-ed ones, the accuracies reach up to 98.84% and 98.69% for the imbalanced and balanced datasets respectively. Finally, as for the CSR-ed inputs, the accuracies reach up to 93.96% and 83.38% for the imbalanced and balanced datasets respectively.