Fetal Electrocardiogram Extraction Utilizing Deep Long Short-Term Memory Network

The fetal electrocardiogram(FECG)provides important information about the health condition of the fetus.By monitoring the FECG during the perinatal period,physicians are able to diagnose early fetal hypoxia,fetal distress,congenital heart malformations,neonatal arrhythmia,intrauterine growth retardation and other diseases in pregnancy and delivery,which reduce fetal morbidity and mortality.Nowadays the most widely used technique for fetal monitoring in clinics is cardiotocography(CTG).However,CTG is not able to provide reliable information of instantaneous fetal heart rate(FHR)variability.FECG is a more promising technique better than CTG because it has the potential to provide high-precise information of FHR and valuable status information of fetal heart.Therefore,it is of great practical and theoretical significance to research on how to obtain clear and complete fetal ECG signals.The non-invasive FECG extraction methods are divided into two categories:combined source(CS)methods and abdominal electrode-sourced(AES)methods.CS methods require both chest and abdominal electrodes to obtain maternal electrocardiogram(MECG)and abdominal signals.And AES methods require only abdominal electrodes to obtain maternal abdominal signals,which are more convenient for clinical application.In this paper,the FECG extraction method of CS utilizing deep long short-term memory(Deep LSTM)network is proposed.Firstly,a Deep LSTM network model for FECG extraction is constructed.Then with the MECG as reference signal,the MECG component in maternal abdominal signals is estimated utilizing Deep LSTM.At last,the clear FECG is obtained by subtracting the estimated MECG component from the abdominal signal.The proposed FECG extraction method utilizing Deep LSTM is validated by the experiments on both synthetic and real ECG signals.The experimental results indicate that the proposed FECG extraction method utilizing Deep LSTM is effective and better than some conventional noninvasive FECG extraction methods respectively based on least square support vector machine(LSSVM),adaptive neuro-fuzzy inference system(ANFIS)and v-support vector regression(v-SVR).For AES signals containing only maternal abdominal signals,the FECG extraction method of AES utilizing blind source separation based on time-frequency distributions(TFBSS)combined with Deep LSTM is proposed.Firstly,the preliminary estimation of MECG and FECG is obtained by TFBSS.Then with the estimated MECG as reference signal,the MECG component in the preliminarily estimated FECG is estimated utilizing Deep LSTM.Finally,the clear FECG is obtained by removing the estimated MECG component.The real data are adopted to validate the proposed FECG extraction method utilizing TFBSS combined with Deep LSTM.The experiment results show that the FECG extraction method utilizing TFBSS combined with Deep LSTM is better than FECG extraction methods respectively based on TFBSS,Fast ICA,Robust ICA,Fast ICA combined with Deep LSTM and Robust ICA combined with Deep LSTM.