Fetal ECG (FECG)
IntroductionNon-invasive FECG aims to provide ECG traces representing the fetal heart activity in early and late pregnancy comparable with those achievable on adult subjects. Even with all the limitations posed by the indirect recording [1], a morphological analysis can be carried out to identify fetal pathologies and set up a therapeutic intervention in time. In transabdominal potential recordings, the lowpower FECG is hidden in a mixture of several high-power sources, mainly ascribable to the maternal physiological interferences (ECG, EMG, respiration) and to the instrumental noises. It is well known that the original sources cannot be separated using traditional frequency domain filters due to the spectral overlap between the different sources [2]. The only methods giving recognized good results are based on adaptive filtering or Blind Source Separation (BSS) algorithms, the latter usually providing better results than the former at the expenses of an increased complexity [3].To cope with the lack of real-time ICA algorithms able to provide results with a quality comparable with the batch ones, we developed a block-on-line algorithm, OL-JADE, based on the famous JADE algorithm [4], able to give the same quality of the original batch algorithm joined with the tracking ability of an on-line one. Since this kind of processing is targeted for real-time clinic examinations, we also proposed a real-time implementation on a floating point DSP [5]. The critical issue of the algorithm is the permutation ambiguity (due to the mathematical formulation of the problem) since it may lead to a blockby-block scrambling of the separated sources. In this paper this problem, partially addressed in the formulation of the on-line algorithm, is studied on real mixtures, providing further information and possible solutions, along with comparisons with the original batch approach.
MethodsICA aims to find a linear transformation that minimizes the statistical dependence between the components of a random vector. When it represents the mixture of more unknown independent source signals, and the mixing process is also unknown, ICA can be used to find a good estimate of the sources, thus acting as a BSS algorithm.OL-JADE is a block-on-line version of the famous JADE [4] algorithm, which consists of a Second Order Statistics (SOS) stage providing centering and whitening of the original signal mixtures, followed by a Higher Order Statistics (HOS) stage. Whitening decorrelates and orthogonalizes the original mixtures, reducing the number of parameters to estimate, so that only a rotation, provided by the HOS stage, is then required to identify the independent sources [6]. JADE is a batch algorithm, working on a segment of data of interest containing enough statistical information on the independent components to perform the separation. This means that both the SOS and HOS stages are performed off-line. The HOS stage obtains higher-order independence by multiplying decorrelated mixtures by an orthogonal rotation ...