Deep Adaptive Denoising Auto-Encoder Networks for ECG Noise Cancellation via Time-Frequency Domain

In this paper, we study the performance of a deep adaptive denoising auto-encoder network (DeepADAENet) for electrocardiogram (ECG) signal noise cancelation in the time-frequency domain for practical use cases. In order to achieve a higher resolution in distinguishing the noise from valuable data, the fractional Stockwell transform (FrST) is exploited to convert the ECG to the time-frequency image. The magnitude of the time-frequency version of the ECG is noise-canceled using DeepADAENet. Then, inverse FrST is utilized to return the denoised time-frequency ECG into the time domain. Furthermore, we use the MIT-BID Apnea-ECG database (APNEA-ECG) for preparing the dataset due to various physiologies and records compared with other ECG databases. Moreover, muscle artifacts (MA), baseline wander (BW), and electrode motion (EM) from the MIT-BID Noise Stress Test Database (NSTDB) are utilized to make noisy this clean dataset. The ECG signals recorded by non-clinical devices contain more noise than clinical recording. Accordingly, by changing the coefficient and frequency of noise resources, we attempt to close the simulated noisy signal to reality. Results reveal the excellent performance of DeepADAENet compared with similar work in terms of signal-to-noise ratio (SNR), root mean square error (RMSE), and percent root mean square difference (PRD).