Blind adaptive filtering for extraction of fetal ECG from maternal abdominal ECG

This research presents a novel method and results for automatically extracting beat-wise fetal ECG (fECG), including RR, QS, PR and QT intervals, from maternal surface electrodes early in the pregnancy down to week twenty and possibly earlier.; Extraction is based on a two-stage Blind Adaptive Filtering (BAF) approach that overcomes the theoretical limitations in applying conventional Blind Source Separation (BSS) methods based on ICA (Independent Component Analysis) or AMUSE (Algorithm for Multiple Unknown Signals Extraction) to this signal extraction problem. The limited number of channels, the non-stationary observations, and the considerable channel noise in the real patient data make the BSS algorithms inapplicable to the fECG extraction problem contained herein.; The proposed BAF approach has two stages. Stage I (BAF-I) employs a harmonic filter array to extract the fECG frequency domain features, while the Stage II (BAF-II) incorporates a specially designed LAMSTAR neural network to utilize the fECG time domain features. The intersection of these fECG features from two different domains provides a possibility to locate the fECG signal and further to extract it. Additionally, an extension of this system allowed the identification of the following items: QRS complex, P wave, T wave, PR duration, QT duration, and the QS duration.; Extraction via the above BAF approach has succeeded in all real data sets available. These data sets were recorded from the abdominal skin of the pregnant mothers including the gestation week twenty, the earliest for which we obtained raw maternal ECG recordings. In every case beat-wise un-averaged fECGs were isolated. Furthermore, our results matched well (788 of 802 events) to Doppler data that corresponds to the same 20th week data set used by us and to which we never had direct access.; On the basis of the quality and clarity of the extracted fECG signals, there appears to be no limitations that will prevent success at earlier gestation weeks down to possibly week seventeen. However, at the time of this writing, no raw data earlier than week twenty is available.