Wave Masking (WM), a preprocessing technique, has been showed to improve the accuracy of electrocardiogram (ECG) reconstruction when using a linear regression (LR) model. This technique involved the selective masking of sections of the ECG to focus the regression model on certain features of the signal. However, the original WM method applies zero-padding to mask the signal may not be the optimal approach as alternative padding approaches have not been explored. Furthermore, since WM was originally adapted from image recognition techniques, ECG reconstruction may also benefit from incorporating previous samples called Temporal Dependency (TD) as in Long Short-Term Memory (LSTM) networks.
This study evaluates the performance of LR models in ECG reconstruction using various WM and TD variants and hybrid approaches. The preprocessing techniques analysed include A) WM with zero-padding (WMZ). B) WM with sigmoid function padding (WMS). C) WM with signal baseline padding (WMB). D) Incorporating previous samples into the signal (TD). E) Combination of WMZ and TD (WMZ_TD). F) Combination of WMS and TD (WMS_TD). G) Combination of WMB and TD (WMB_TD).
The analysis was conducted using 4,250 unique 10-second normal ECG recordings from the CODE 15% database, which were denoised and resampled to 500 Hz. A patient-wise validation was performed with 3500 ECGs used for training and 750 ECGs for testing. The different preprocessing techniques were evaluated using correlation and root mean square error (RMSE) metrics. The results indicate no significant performance differences between the tested techniques. This suggests that modifying the WM padding method—whether by using a sigmoid function or the signal baseline—does not improve the accuracy of the LR model. Additionally, incorporating previous samples as part of the input signal does not yield any notable enhancement in reconstruction performance.