Cardiac Arrhythmia Recognition Using Transfer Learning with a Pre-trained DenseNet

Abstract

Recent findings demonstrated that deep neural networks carry out features extraction itself to identify the electrocardiography (ECG) pattern or cardiac arrhythmias from the ECG signals directly and provided good results compared to cardiologists in some cases. But, to face the challenge of huge volume of data to train such networks, transfer learning is a prospective mechanism where network is trained on a large dataset and learned experiences are transferred to a small volume target dataset. Therefore, we firstly extracted 78,999 ECG beats from MIT-BIH arrhythmia dataset and transformed into 2D RGB images and used as the inputs of the DenseNet. The DenseNet is initialized with the trained weights on ImageNet and fine-tuned with the extracted beat images. Optimization of the pre-trained DenseNet is performed with the aids of on-the-fly augmentation, weighted random sampler, and Adam optimizer. The performance of the pre-trained model is assessed by hold-out evaluation and stratified 5-fold cross-validation techniques along with early stopping feature. The achieved accuracy of identifying normal and four arrhythmias are of 98.90% and 100% for the hold-out and stratified 5-fold respectively. The effectiveness of the pre-trained model with the stratified 5-fold by transfer learning approach is surpassed compared to the state-of-art-the approaches and models, and also explicit the maximum generalization of imbalanced classes.