Several Key Methods Of Noninvasive Fetal Electrocardiogram Adaptive Extraction

Fetal electrocardiogram（FECG）monitoring in perinatal stage has been attracting people’s attention along with the advance of biomedical instrument development tech-nology and strengthening consciousness of eugenic and superior nurture.Therefore,the corresponding FECG signal processing technology has gradually become one of research hot spots in the field of biomedical signal processing.FECG is regarded as an objec-tive indicator of fetal cardiac electrophysiological activity and can distinguish the subtle changes of fetal heart beats.As a kind of effective means of FECG monitoring,it can reflect the state of fetus health during pregnancy and is helpful for early diagnosis of fe-tal intrauterine hypoxia,congenital heart disease,etc.in both pregnancy and intrapartum stage.Moreover,FECG can reflect the whole heart activity as compared with traditional doppler ultrasound technique.Therefore,extracting clear FECG signal is the primary task of FECG monitoring,which will contribute to clinical medical diagnosis.However,firstly,the maternal ECG（MECG）component within the composite ab-dominal ECG（AECG）is the distorted version of MECG through the nonlinear transmis-sion path.Secondly,for multiple AECG signals,the traditional adaptive noise canceller（ANC）adopts single-channel AECG to extract FECG,which may lead to the absence of some information on FECG and the decline of fetal QRS（FQRS）detection accuracy.Thirdly,a few MECG component,EMG and power line interference still reside in FECG signal produced by ANC,which brings poor FECG signal-to-noise ratio（SNR）and is unfavorable for its morphological analysis.These issues become key points of FECG extraction.This work focuses on several key methods of adaptive FECG extraction.It contains three aspects including approximation of nonlinearity from the MECG to distorted MECG within AECG,optimization of ANC with multiple abdominal channels,and denoising of the extracted FECG.The main contributions of this work are summarized as follows:（1）A nonlinear ANC based on the generalized functional link artificial neural net-work（FLANN,GFLANN）is proposed for FECG extraction.These conventional ANCs based on FIR filters,Volterra filter or FLANN were respectively used to identify the non-linearity between the MECG and the composite AECG.But these methods present limited ability of suppressing maternal component.To this end,considering the expansion order and cross term of kernel function,a GFLANN-based nonlinear ANC is designed,where an FIR filter and a GFLANN are equipped in parallel in each reference channel to respec-tively approximate the linearity and nonlinearity between MECG and composite AECG.This new approach can further reduce and eliminate the MECG component and hence achieve clear FECG.In every reference channel,the coefficients of the linear FIR filter and the weights of nonlinear GFLANN are updated by LMS algorithm and the steepest descent method respectively.Both simulated database and PNIFECG database were used to valuated the performance of our proposed technique in terms of the correlation coeffi-cients between the original FECG and extracted one and the statistical indices with respect to the detected FQRS.For the subdataset A,the F₁-measure produced by our proposed technique reaches 97.9%.In addition,the proposed technique has low computation com-plexity and is favorable to portable and low power-consuming measurement applications.（2）An novel adaptive nonlinear ANC structure with both multiple abdominal chan-nels and multiple maternal ones is proposed for FECG extraction.The conventional ANC usually uses single abdominal channel,which will result in the absence of some infor-mation on the extracted FECG and further affect the accuracy of FQRS detection.To solve this problem,this paper tries to reasonably increase the number abdominal channels and combine them by using of a linear combination（LC）,and then the optimal input of the multiple abdominal is obtained.Based on the principle of ANC for FECG extrac-tion,a novel ANC structure of FECG extraction is proposed equipped with both multiple abdominal channels and multiple maternal ones simultaneously.The LC connecting the primary channels is updated by a constrained recursive least square（RLS）algorithm.Ex-perimental results on both the PNIFECG database and DaISy database demonstrate the effectiveness of the proposed technique in the aspects of the statistical indices about FQRS detection and the visual FECG.For the subdataset A,the F1-measure result obtained by the proposed method amounts to 99.0%.This novel structure has the advantages of higher FQRS detection accuracy and relatively low complexity,which promotes its application in the situation of long-term FECG monitoring.（3）A novel adaptive LMS-based Fourier analyzer is proposed for post-denoising the extracted FECG signal.The afore-mentioned techniques focus on extracting FECG signal by suppressing MECG component within AECG.However,a small amount of MECG components,EMG and other noises still reside in the extracted FECG,which need to be further denoised to obtain clearer FECG waveform aiming at future morphological analysis.From the spectrum of extracted FECG,we found that it is of non-stationarity and pseudo periodicity and can be represented by a sum of a number of dominant low-frequency discrete sine and cosine signals.Based on these characteristics,the inverse dis-crete Fourier transform（IDFT）may be used to denoise the extracted signal by selecting proper spectrum threshold,and further reconstruct clear FECG.However,this technique has the contradiction between preserving signal local characteristics and restraining noise,and it is difficult to deal with the amplitude and periodic changes of nonstationary FECG.To this end,a novel post-denoising technique is proposed based on adaptive Fourier an-alyzer,whic is implemented by simultaneously tracking the amplitude and periodicity of the dominant FECG components.Simulations show that the new method posses excel-lent denoising performance,especially when FQRS wave and MQRS wave are close or overlapped,the proposed method can still restore the FQRS wave.Experimental results with clinical PNIFECG and DaISy database show that the proposed method is superior to the IDFT in terms of the visual FECG and two SNRs based on eigenvalue analysis and mutual relation.In particular,for DaISy database,the above two SNRs obtained from the proposed method improve 6.25dB and 6.73dB respectively.Furthermore,this new method is implemented adaptively and presents better denoising performance in case of low-SNR noisy signal as compared with the IDFT approach,which is potential for clinical application.