Non-invasive methods for estimating the fetal electrocardiogram (FECG) by using ECG recordings from the abdomen of the mother are of great interest in order to carry out less complicated and safer fetal monitoring compared to the invasive methods. In this thesis, independent component analysis (ICA) is used to extract morphologically accurate fetal ECG, by estimating the maximally independent sources from the observed mixture.
A constrained ICA (cICA) algorithm is derived, which is based on constraining the ICA optimization problem such that the estimated sources must be correlated with some reference signals.
Different reference signals are generated, ranging from pulse signals at the QRS locations to templates generated from e.g. maternal ECGs. A hyper-parameter search is conducted, where different step-sizes, correlations thresh- olds and combinations of maternal and fetal reference signals are tested. Using the parameters giving the best results, the cICA algorithm is tested on a synthetic dataset, where its performance is compared to other extraction algorithms such as template subtraction PCA as well as other ICA methods such as FastICA and Infomax.
From this test, it is clear that cICA can do accurate extraction if the true FECGs are used as reference, but performs less well if other references are used, and is comparable to classical ICA methods in these cases. None of the tested algorithms were able to extract morphologic features (QT- interval and T/QRS ratio) from real ECG mixtures. It was found that cICA converges very fast, but to a local minimum, which is due to the non-convex nature of the problem. It can therefore be concluded that constrained ICA will not provide morphologically accurate FECG if the true FECGs are not provided.