An end-to-end motion artifacts reduction method with 2D convolutional de-noising auto-encoders on ECG signals of wearable flexible biosensors

Background and objective: The signals from the wearable biosensors face a number of artifacts that restrict their accuracy in diagnosis, especially motion artifacts. ECG is a noninvasive and reliable tool for detecting and preventing heart diseases. In real-world scenarios, ECG signals are more susceptible to different categories of artifacts contamination, which leads to incorrect interpretation. So, in this study, a significant attention has been paid to de-noising of ECG signals for accurate analysis and disease diagnosis. To reconstruct the clean signals from the noisy ones, a de-noising auto-encoder (DAE) could be used. Methods: Two de-noising auto-encoders, named as 2D-CEDNet and 2D-SCEDNet, have been proposed in this study to remove the artifacts from the noisy ECG signals. A prepared dataset using flexible wearable smart T-shirt (which contains flexible bio-sensors) along with two publicly available datasets named as 'MIT-BIH Noise Stress Test Database (NSTDB)' and 'MIT-BIH Normal Sinus Rhythm Database (NSRDB) with NSTDB Noises' are used to verify the performance of these nets. The noisy ECG signals of all datasets are split into 10s, 5s, and 3s segments, which are transformed into their corresponding RGB (red, green, blue) images. These transformed images are used as the input of the proposed auto-encoders to test their performance. Results: The experimental results on the noisy ECG signals with various levels of SNR demonstrate that the proposed auto-encoders outperform the state-of-the-art auto-encoders with higher SNRimp and lower RMSE and PRD. Furthermore, the proposed auto-encoders compress the noisy signals to 128 times smaller than the original one. Conclusions: From the findings, we believe that the proposed auto-encoders have promising applications in wearable electronics and clinics.