Fetal ECG extraction using a novel polynomial network and non-linear BSS method

Long-term monitoring of fetal heart is vital for early clinical diagnosis and timely treatment of fetus. However, the collected signals from the abdomen of the pregnant woman are always corrupted by many interfering sources, most significantly affected by the maternal electrocardiograph (ECG). Similar with blind source separation (BSS), fetal ECG extraction can be considered as the separation of fetal ECG from the recordings. This letter proposes a novel multi-layer polynomial network to extract the fetal ECG waveform, where the non-linear mixing process is approximated with multiple polynomials. The network starts with the first-order polynomial layer, and further requirements to decrease the error between the output and target signal is achieved by increasing the number of layers. The proposed approach turns the inverse transformation of traditional BSS into a model learning problem, which is then solved by a simple convex optimization algorithm. Compared with the state-of-the-art BSS methods, experimental results show that the proposed method has better performance.