Fetal Movement Cancellation in Abdominal Electrocardiogram Recordings Using Signal-to-Signal Translation.

Abstract

This study addresses the cancellation of fetal movement in abdominal electrocardiogram (AECG) recordings through deep neural networks. For this purpose, a generative signal-to-signal translation model consisting of two coupled generators is employed to discover the relations between fetal movement-contaminated and clean AECG recordings. The model is trained on the fetal ECG synthetic database (FECGSYNDB) which provides AECG recordings from 10 pregnancies along with their ground-truth maternal and fetal ECG signals. The signals are initially segmented into 4-second segments and then fed into the network for denoising. It is demonstrated that the signal-to-signal translation method can reconstruct clean AECG signals with average mean-absolute-error (MAE), root-mean-square deviation (RMSD), and Pearson correlation coefficient (PCC) of 0.099, 0.124, and 99.12% respectively, between clean and denoised AECG signals. Furthermore, the robustness of the method to low signal-to-noise ratio (SNR) input values is shown by an RMSD range of (0.047, 0.352) for SNR values within the range of (-3, 3) dB. Clinical Relevance- The proposed framework allows for the denoising of abdominal ECG signals for non-invasive fetal heart rate monitoring. The approach is accurate due to the use of advanced neural network techniques.