Impact of Contrast Functions in Fast-ICA on Twin ECG Separation

Fetal Electrocardiography (FECG) is a traditional method to measure fetal heart conduction signals during gestation. It can allow determination of fetal heart rate (FHR) along with amplitudes and timing of the FECG components, as these are indices for fetal health. FECG recording is problematic in the clinic due to the several interferences that corrupt the signal. It is recorded using simple electrodes placed on the mother’s abdomen and is a part of this abdominal ECG. This abdominal ECG also contains several interferences. First the omnipresent maternal ECG (MECG) is a huge source of interference. Other possible noise sources are respiration and muscle activity along with thermal/electronic noise and noise from electrode-skin contact. The problem becomes all the more difficult and complicated for twin fetuses that could have FECG signals of similar morphology, amplitudes and heart rates. We are interested in extracting twin FECG signals from abdominal ECG using Independent Component Analysis (ICA) techniques, a way of Blind Source Separation (BSS). Twin fetuses require close monitoring of their heart health to track for congenital heart problems. These are especially important with identical twins complicated with twin-twin transfusion syndrome. In this paper, we work with the Fast-ICA technique and propose and test two types of data-centric contrast functions (Poly-L and AbsPow). These are obtained when the underlying data pdf is modeled as an exponential power distribution. We test this with 3 types of abdominal data sets simulated, in-vivo (clinical) and ICA bench-marks with added Gaussian and muscle noise. We also compare performance across the standard fixed-point contrast functions in Fast-ICA using a normalized metric. We clearly show that Poly-L and AbsPow perform superior to the standard-ICA data-based Pearson method and works on par and in some cases better than standard ICA polynomial schemes like Pow3. When estimation was done over data acquisition time, it was clear that adequate data (10-30 mins) was needed for good performance. Performance metrics for data-centric contrast functions with order 3-4 (Poly-3 and AbsPow) had the best performance for certain cases with little increase in complexity. This supports further testing with a large database of twin gestation data with heart defects.