Neurodynamic non-invasive fetal electrocardiogram extraction

Fetal electrocardiography in contrary to adult is not that well represented in publications, yet circulatory system of the fetus is probably the most valuable and crucial biological infrastructure. Fetal heart ratio, form of QRS-wave and dynamics of cardiovascular system activity allow estimating fetus state, maturity, possibilities of heart abnormality occasion. This information can be received with guaranteed accuracy through Doppler-ultrasound procedure, however duration of such kind of monitoring is limited. Fetal electrocardiogram is an obvious source of information about fetal heart activity. However, because of low signal-to-noise ratio and prevailing of maternal component, non-invasive ways of acquiring this signal do not guarantee absolute accuracy. Problems of non-invasive electrocardiography demand complex mathematical approaches because maternal and fetal R-peaks overlap in time and frequency domains and have similar morphological structure of heart waves. In this paper we propose approach for extracting fetal electrocardiography from abdominal signal, which is based on dynamic neural network. The common problem for both dynamic and deep learning is caused by linearity of backpropagation and thus vanishing or exploding of gradients occurs. We proposed resilient propagation through time approach that unites training based on sign of derivative and parallel unfolding. We compared developed algorithm with blind source separation through independent component analysis and noted several important advantages that our model delivers - accuracy does not depend on: length of signal; amount of independent channels.