Abstract. Ocular artifacts are the most important form of interferences in EEG signals. Before analyzed, EEG signals should be pretreated by removal of ocular artifacts. CICA is an excellent approach to separate the desired source signals. But, the choice of reference signals is crucial. In this paper, we adopted CICA to separate ocular artifact from EEG, using a different method from Lu to build the reference signals, which can avoid the subjectivity during the operation. It was proved to be effective.
IntroductionElectroencephalogram (EEG) analysis has very important significance for clinical diagnosis. But EEG signals are very random and faint, which are easily interfered by multifarious artifacts. Ocular artifacts are the most important form of interferences in EEG signals. So, before analyzed, EEG signals should be pretreated by removal of ocular artifacts. To obtain these artifacts, one powerful technique is blind source separation (BSS) [1,2], which simultaneously separates all source signals. More recently, independent component analysis(ICA) is proved to be an excellent BSS method [3,4].However, in most cases, what we need are not all source signals but just a few ones. For such cases, constrained independent component analysis (CICA) [5,6], which was built up based on fast ICA [7,8], is a more suitable method. It can avoid heavy computational load and costing of time.Because ocular artifacts only present to several EEG channel signals, such as channel fp1, fp2 and fpz, cICA is a good idea to separate them. CICA is actually used to form a constrained optimization problem maximizing a new objective function subject to the additional constraints that the extracted ICs are the closest to the corresponding reference signals. And an efficient adaptive algorithm, Lagrange multipliers method [9], can be adopted to solve this constrained optimization problem. It is worth mentioning that C.J.James has successfully removed ocular artifacts from EEG using cICA [10]. The method he adopted to build reference signals is to set up square pulse over the region of interest with a zero reference elsewhere. But just as he has illuminated, this method is too subjective, because the shape of the reference signal may influence the output result in the sense that artifacts of slightly different morphology may be extracted for different reference morphology. In addition, when using a correlation measure of closeness, the phase of a reference must be closely matched to that of the desired source signal, which has to be done with great workload by repeatedly applying cICA to data with the reference shifted by one sample to cover one period of the signal of interest [10].