An approach of neural network based fetal ECG extraction

In this paper, we describe an adaptive method to separate fetal ECG from composite ECG that consists of both maternal and fetal ECGs by using ADALINE (adaptive linear network). The input signal is the maternal signal and the target signal is the composite signal. The network emulate maternal signal as closely as possible to abdominal signal thus only predict the maternal ECG in the abdominal ECG. The network error equals abdominal ECG minus maternal ECG, which is the fetal ECG. The characteristic that enables fetal extraction is due to correlation between maternal ECG signals with the abdominal ECG signal of pregnant woman. A GUI program is written in Matlab to detect the changes in extracted fetal ECG by different values of momentum, learning rate and initial weights used in the network. However, the learning rate, momentum and initial weights are adjusted until the results are reasonably well. It is found that filtering performs best by high learning rate, low momentum, and small initial weights.