A deep learning framework for electrocardiogram (ECG) super resolution and arrhythmia classification

The preferred sampling rate for recording the electrocardiogram (ECG) data as per industry standards is roughly 500 Hz. For the remote monitoring of patients, the transmission process has to be made power-efficient and less expensive. Therefore, the ECG should be transmitted in low resolution and reconstructed efficiently without losing crucial information on the receiver side for a proper diagnosis. This research proposes a deep learning–based scheme for reconstructing high-resolution ECG signals from its low-resolution version. Subsequently, a deep learning–based classifier is designed for cardiac arrhythmia (CA) classification.

In this work, we carefully designed two neural network architectures for ECG super resolution and cardiac arrhythmia classification, respectively. The first network, called ECG-SRCNN (ECG-super resolution convolutional neural network) is developed to reconstruct a high-resolution ECG signal from its low-resolution samples from the CPSC 2018 dataset. After obtaining a high-resolution signal, another 1D-CNN is designed for a nine-class cardiac arrhythmia classification.

The research work on ECG-SRCNN (ECG-super resolution convolutional neural network) gave an excellent arrhythmia classification for an upscaling ratio of 4 with a Pearson correlation coefficient of 97.5% in comparison to SAE (Stacked Autoencoder) and 1D-SRCNN ( with three convolution layers).

In conclusion, this research contributes to developing deep learning architecture for super resolution of ECG signal to enhance the arrhythmia classification from low-resolution ECG signals. The proposed scheme gave an accuracy of 99% on the CPSC 2018 dataset for nine different types of arrhythmia classification.