A Non-Invasive Fetal Electrocardiogram Extraction Algorithm Based on ICA Neural Network

Ye, Yalan; Yao, Xun; Zhang, Zhilin; Mo, Quanyi

doi:10.1109/icbbe.2007.210

Cited by 2 publications

(1 citation statement)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Most of papers [11][12][13] propose using models based on multi-layer perceptron (MLP). This model is static and does take into account timevarying features of maternal and fetal mixing process.…”

Section: Introductionmentioning

confidence: 99%

Blind Separation of Abdominal Electrocardiogram Sources through Dynamic Neural Network

Devyatykh¹,

Gerget²

2016

Proceedings of the 2016 Conference on Information Technologies in Science, Management, Social Sphere and Medicine

View full text Add to dashboard Cite

Abstract-Cardiovascular system of the fetus is biological critical infrastructure. Fetal electrocardiogram and its characteristics such as heart ratio, waveform and dynamic behavior overall include vital information about health state, development and possible deviations from normal fetation. Thus fetal heart ratio monitoring is mandatory during pregnancy. Widespread Doppler ultrasound diagnostics can guarantee obtaining accurate results but is not suitable for long-term monitoring. Non-invasive fetal electrocardiography proposes to extract fetal signal from maternal abdominal electrocardiogram. This approach is applicable for long-term monitoring, but because of amplitude of maternal R-peaks is significantly larger than fetal it is a challenge to extract fetal signal. In this paper we propose using dynamic neural networks for extracting fetal components and demonstrate its advantages compared to blind source separation though independent component analysis. The training algorithm is a combination of backpropagation through time and resilient propagation. The proposed approach accuracy of R-peak detection is 97%. Statistical analysis proved that developed algorithm can process even non-stationary signals with loss of accuracy and no additional training is required.

show abstract

Section: Introductionmentioning

confidence: 99%