ECG Signal Denoising for Wearable Devices

The ECG represents cardiac electrical activity recorded by electrodes placed at a distance from the heart, on the surface of the body. During a cardiac electrical activity recording session, artifacts like baseline wander, powerline interference, EMG noise and electrode motion artifact are present in addition to the ECG signal. These are unwanted signals that are mixed with the ECG signal and can lead to erroneus interpretations or diagnoses. Therefore, they must be eliminated from ECG signals using appropriate signal denoising techniques. The aim of this work is to find the best denoising technique for ECG signal in a given context: the proposed denoising technique must be suitable to be applied with a real time approach to real noisy signals acquired with a wearable device. Initially, some state-of-the-art ECG denoising techniques are compared; subsequently, strengths and weaknesses of these techniques are identified and attempts are made to optimize them for the given task. The employed dataset consists of 81 two-leads ECG recordings from 46 healthy subjects with a total duration of approximately 595 hours, collected using L.I.F.E company’s wearable vest Healer R2. The metrics identified to evaluate the quality of ECG signals before and after preprocessing are ECGSQI and basSQI. Additionally, computational time is considered to assess the applicability of the denoising algorithms in real-time situations. Visual inspection is also used to assess the morphological alterations of the reconstructed signals. The following SoA techniques have been reproduced and compared: two cascaded IIR filters, EMD implemented according to Blanco-Velasco algorithm and DWT algorithm. In a second phase, to reduce the computational cost of the EMD based denoising algorithm, two approaches are proposed for QRS complexes identification: one is to use a fixed-duration window, while the other is to use a window with a duration proportional to the RR interval. Different combinations of mother wavelet, thresholding rule and decomposition level are analyzed to find the optimal setting of the DWT based algorithm in the given context. Finally, the IIR filters are added upstream of the other techniques to try to improve their performance. According to the results obtained, the DWT based optimized algorithm with the upstream filters turns out to be the denoising technique with the highest performance for the given task. For the first channel the ECGSQI obtained is 0.841, the basSQI is 0.889; for the second channel the ECGSQI obtained is 0.833, the basSQI is 0.885. The obtained computational time is 25.8 ms.