Beat-by-Beat ECG Monitoring from PPG Using Spatio-Temporal Information with WaveNet

The synthesis of the Electrocardiogram (ECG) from the Photoplethysmogram (PPG) is a hot research topic in the last decade. Also, PPG can be measured using PPG sensor connected with fingertips or even measured from RGB signals. The successful reconstruction of the ECG waveform from a PPG signal promises to enable continuous, comfortable, and diagnostically rich cardiac monitoring outside of traditional clinical settings. The main limitations of the current ECG reconstruction techniques are the use of invalid training datasets, unsuitable networks, and/or insufficient information. Therefore, in this paper, we propose to emphasize the cleanness of the datasets including spatial and temporal information with adapted WaveNet as an efficient network which is deep neural network model for generating raw audio waveforms. The adaptation includes the use of the network for regression rather than audio generation. Also, the architecture of the network is optimized for signal-to-signal regression. Two scenarios are designed and tested in this paper, namely, beat-to-beat regression fixed number of beats (three beats is tested) to beat. The effect of using the time interval information is evaluated as well, The experimental results demonstrated that using fixed number of beats is better than using single beat and better that using signal to signal ECG estimation. Also, it confirmed that the time interval information is essential for ECG monitoring. The proposed system can achieve a Pearson‘s correlation coefficient (r), Fréchet distance (FD), root mean square error (RMSE), percentage root mean square difference (PRD), mean absolute error (MAE) and standard deviation (SD) of 0.9831, 0.1174, 0.0373, 12.3207, 0.0262 and 0.0371, respectively which outperforms the existing state-of-the-arts ECG reconstruction algorithms. This research uses both of spatial and temporal information in beat-to-beat ECG reconstruction.