A multichannel nonlinear adaptive noise canceller based on generalized FLANN for fetal ECG extraction

作者：Ma, Yaping^*; Xiao, Yegui; Wei, Guo; Sun, Jinwei

来源：Measurement Science and Technology, 2016, 27(1): 015703.

DOI：10.1088/0957-0233/27/1/015703

• fetal electrocardiogram (FECG)
• functional link artificial neural network (FLANN)
• generalized FLANN (GFLANN)
• adaptive noise canceller (ANC)
• F1 measure

In this paper, a multichannel nonlinear adaptive noise canceller (ANC) based on the generalized functional link artificial neural network (FLANN, GFLANN) is proposed for fetal electrocardiogram (FECG) extraction. A FIR filter and a GFLANN are equipped in parallel in each reference channel to respectively approximate the linearity and nonlinearity between the maternal ECG (MECG) and the composite abdominal ECG (AECG). A fast scheme is also introduced to reduce the computational cost of the FLANN and the GFLANN. Two (2) sets of ECG time sequences, one synthetic and one real, are utilized to demonstrate the improved effectiveness of the proposed nonlinear ANC. The real dataset is derived from the Physionet non-invasive FECG database (PNIFECGDB) including 55 multichannel recordings taken from a pregnant woman. It contains two subdatasets that consist of 14 and 8 recordings, respectively, with each recording being 90 s long. Simulation results based on these two datasets reveal, on the whole, that the proposed ANC does enjoy higher capability to deal with nonlinearity between MECG and AECG as compared with previous ANCs in terms of fetal QRS (FQRS)-related statistics and morphology of the extracted FECG waveforms. In particular, for the second real subdataset, the F1-measure results produced by the PCA-based template subtraction (TSpca) technique and six (6) single-reference channel ANCs using LMS- and RLS-based FIR filters, Volterra filter, FLANN, GFLANN, and adaptive echo state neural network (ESNa) are 92.47%, 93.70%, 94.07%, 94.22%, 94.90%, 94.90%, and 95.46%, respectively. The same F1-measure statistical results from five (5) multi-reference channel ANCs (LMS- and RLS-based FIR filters, Volterra filter, FLANN, and GFLANN) for the second real subdataset turn out to be 94.08%, 94.29%, 94.68%, 94.91%, and 94.96%, respectively. These results indicate that the ESNa and GFLANN perform best, with the ESNa being slightly better than the GFLANN but about four times more computationally expensive than the GFLANN, which makes the GFLANN a good alternative for NI-FECG extraction.

• 出版日期2016-1
• 单位哈尔滨工业大学