Wyniki wyszukiwania - BazTech

Ograniczanie wyników

1 Biocybernetics and Biomedical Engineering

1 2024

Znaleziono wyników: 1

Liczba wyników na stronie

Wyniki wyszukiwania

Wyszukiwano:
w słowach kluczowych: over complete mixture

Sortuj według:

Ogranicz wyniki do:

Spatio-spectral independent component analysis for fetal ECG extraction from two-channel maternal abdominal signals

Kotas Marian P., AlShrouf Anwar M.

Biocybernetics and Biomedical Engineering

2024

Vol. 44, no. 1

247--263

Independent component analysis (ICA) is widely used to separate maternal and fetal electrocardiograms. However, it has become less effective due to the efforts to reduce the number of recording electrodes. To address this issue, we propose an extension of ICA that can extract the fetal electrocardiogram from only two maternal abdominal electric signals. We solve this problem by increasing the dimension of the observed signals using the method of delays, followed by spatio-spectral filtering to separate the source signals. By iteratively applying this approach, we can extract signals that are not separable using the original observations alone. These signals are then clustered to create signal subspaces corresponding to different sources, allowing for a rough reconstruction of signal components produced by these sources. This initial decomposition can then be refined by using the reconstructed components as new observations, extending the original ones, and applying ICA to this extended signal representation. Applied to two-channel maternal abdominal signals, the proposed method was able to extract 3 source signals (two maternal and one fetal), resulting in the achievement of the goal of over-complete blind source separation (BSS). Furthermore, the method enabled the successful detection of fetal QRS (fQRS) complexes in experiments on two datasets of real-world maternal abdominal signals. For the ADFECGDB dataset, the method reached the sensitivity, positive predictivity, and F1 score of 100%, 99.97%, and 99.98%, respectively, outperforming all reference methods. For the PREGNANCY dataset, the corresponding values were 98.95%, 98.92%, and 98.93%, second only to one reference method.