System for adaptive extraction of non-invasive fetal electrocardiogram

Abstract

This study aimed to find the most suitable combination of adaptive and non-adaptive methods for extraction of non-invasive fetal electrocardiogram (NI-fECG) using signals recorded from the mother's abdomen. Among the nine methods considered, the combination of independent component analysis (ICA), fast transversal filter (FTF), and complementary ensemble empirical mode decomposition with adaptive noise (CEEMDAN) proved to be the most effective for the extraction of fECG from abdominal recordings. This combined method was suitable due to both being effective in extracting fECG and being less computationally complex. Further, so far, FTF and CEEMDAN methods have not been extensively tested for fECG extraction, and in particular, have not been examined as a hybrid method. The ICA-FTF-CEEMDAN hybrid algorithm was tested on two patient databases: Fetal Electrocardiograms, Direct and Abdominal with Reference Heartbeats Annotations (FECGDARHA) and PhysioNet Challenge 2013. The evaluation of the accuracy of fQRS complexes detection was performed using the following parameters: accuracy (ACC), sensitivity (SE), positive predictive value (PPV), and F1 score. The fetal heart rate (fHR) determination accuracy was evaluated using Bland-Altman plots and fHR traces. When testing on the FECGDARHA database, average values of ACC = 92.98%, SE = 95.33%, PPV = 96.4% and F1 = 95.86% for detection fQRS were achieved. The error in estimating the fHR was -1.02 +/- 7.02 (mu +/- 1.96 sigma) bpm. When testing on the Challenge 2013 database, average values of ACC = 78.47%, SE = 82.06%, PPV = 87.90% and F1 = 84.62% for fQRS detection were achieved, and the error in estimating the fHR was -6.62 +/- 10.33 (mu +/- 1.96 sigma) bpm. In addition, a non-invasive morphological analysis (ST analysis) was performed on the records from the FECGDARHA database, which was accurate in 7 of 12 records with values of mu < 0.03 and values of +/- 1.96 sigma < 0.04.