Simultaneous Denoising and Reconstruction of 12-Lead Electrocardiogram

Promising results have been shown in previous studies for the reconstruction of 12-lead electrocardiogram (ECG) using machine learning techniques. Reliable results have also been shown for denoising ECG with machine learning tools. These two tasks (denoising and reconstruction) are typically handled separately, but this does not have to be. This paper aims to construct 5 denoised ECG leads from a subset of 3 noisy ECG leads. It also aims to analyse the performance of the models over 5 different noise levels. 30s ECG signals were acquired from 80 patients in PTB database. They were denoised and resampled to 500Hz. Different levels of gaussian noise (OdB, 5dB, 10dB, 15dB, 20dB) were added to lead I, II and V3. Linear regression (LR) and long short-term memory (LSTM) models were trained at the different noise levels to predict denoised leads V1, V2, V4, V5, and V6. 80% of the data was used to train the model and 20% was used to test. The average values of Pearson correlation (r), root-mean-square error (RMSE) and signal-to-noise ratio (SNR) against the baseline at Odb were 0.79±0.12, 0.12±0.05mV, and 4.26±2.41dB for LR and 0.87±0.15, 0.10±0.05mV, and 6.42±3.34dB for LSTM respectively. At 20dB they were 0.91±0.10, 0.08±0.05mV, and 7.73±3.99dB respectively for both models. Both models showed similar performance in higher input SNR. As the input SNR reduced, LSTM model showed a more stable performance. At higher input SNR, LR is recommended, whereas at lower input SNR, LSTM is the better option.