A Survey of Kurtosis Optimization Schemes for MISO Source Separation and Equalization - Advances in Heuristic Signal Processing and Applications - page 210

Fig. 9.6 Atrial activity extraction in atrial fibrillation ECGs based on kurtosis optimization. To p :a
5-second segment of ECG lead V1 from an atrial fibrillation patient. Middle : atrial signal contribu-
tion to lead V1 estimated by the combined FastICA-SOBI technique [ 18 ] from the 12-lead ECG.
Bottom : atrial signal contribution to lead V1 estimated by RobustICA-f from the 12-lead ECG. For
reference, gray lines in the background show the original signal recorded in lead V1. Amplitudes
are in mV. The corresponding frequency spectra are shown in Fig. 9.7
in the frequency domain (Fig. 9.7 ). Visual results are validated by quantitative per-
formance measures such as spectral concentration, an atrial signal quality index
proposed in [ 18 ]. Results on a whole AF ECG database confirm RobustICA-f's im-
proved atrial signal estimation performance [ 73 , 74 ]. Source extraction techniques
for artifact suppression in ECG recordings are discussed at length in [ 67 ].
9.6 Conclusions
Over the last two decades, kurtosis has become one of the most popular con-
trasts for blind source separation and equalization in linear channels. In combina-
tion with MISO filtering structures for single-source extraction and suitable defla-
tion schemes, local optimizers of kurtosis lead to satisfactory source estimation in
ideal model conditions. Despite the lack of closed-form solutions, its mathematical
tractability and computational convenience have spurred the development of a rich
variety of cost-efficient iterative methods for optimizing this contrast, mostly based
on gradient and Newton updates. Some of the most representative of these algo-
rithms have been reviewed in this chapter. A selection of ready-to-use MATLAB TM
implementations can be found in the “Online Material” section below.
Space limitations have precluded the treatment of important issues such as filter-
order selection, more elaborate deflation strategies, in-depth analysis of finite sam-