Deep Recurrent Neural Networks for ECG Signal Denoising

We present a novel approach to denoise electrocardiographic signals (ECG), utilizing deep recurrent neural network built of Long-Short Term Memory (LSTM) units. The network is pretrained using synthetic data, generated by dynamic model ECG and fine-tuned with a real data from Physionet PDB database of ECG signals. The results show that a 10-layer DRNN has a mean squared error as low as 0.179 for denoising real signals with white noise of amplitude 0.2 mV, making it a viable alternative for other commonly used methods. We also investigate the impact of synthetic data on the network performance on real signals. Our results show that networks pretrained with synthetic data have better results than network trained with real data only, regardless of the training set size. We propose to explain this by means of the transfer learning framework and the analogy to human cognitive process.