ECG signal denoising based on deep factor analysis

Abstract Objective In telemedicine, dynamic electrocardiogram (ECG) monitoring is important for preventing and diagnosing cardiovascular diseases. However, the interference of the external environment causes a large amount of noises in the dynamic ECG signal, affecting the subsequent automated analysis. Therefore, reduction of the noises in the ECG signal is particularly important. Approach This study proposed a novel ECG signal denoising algorithm based on the deep factor analysis. The major technical innovations include a layer-by-layer denoising deep neural network built based on the factor analysis, in which a top-down strategy is used to reconstruct the signal. The Gaussian-distribution noise can be effectively removed at each layer; and complex noises can be represented by the sum of Gaussian components, thus also removed by the proposed deep network. Moreover, the noise reduction of the network is further improved through a supervised fine-tuning of the parameters of the proposed deep network model, thus increasing the robustness of the whole system in clinical applications. Results The excellent performance of the proposed method has been verified on the MIT-BIH database, and the noise reduction results are evaluated using the signal-to-noise ratio and root mean square error. Significance First of all, the algorithm does not rely on the setting of the frequency domain information and the threshold. Secondly, the algorithm preserves useful information while removing noises from the ECG signal. Finally, a gradient descent algorithm is used to supervise the training of the network, which can learn and preserve the small waveform features in the ECG signal. The performance of noise reduction is outstanding.