An EEMD-ICA Approach to Enhancing Artifact Rejection for Noisy Multivariate Neural Data

As neural data are generally noisy, artifact rejection is crucial for data preprocessing. It has long been a grand research challenge for an approach which is able: 1) to remove the artifacts and 2) to avoid loss or disruption of the structural information at the same time, thus the risk of introducing bias to data interpretation may be minimized. In this study, an approach (namely EEMD-ICA) was proposed to first decompose multivariate neural data that are possibly noisy into intrinsic mode functions (IMFs) using ensemble empirical mode decomposition (EEMD). Independent component analysis (ICA) was then applied to the IMFs to separate the artifactual components. The approach was tested against the classical ICA and the automatic wavelet ICA (AWICA) methods, which were dominant methods for artifact rejection. In order to evaluate the effectiveness of the proposed approach in handling neural data possibly with intensive noises, experiments on artifact removal were performed using semi-simulated data mixed with a variety of noises. Experimental results indicate that the proposed approach continuously outperforms the counterparts in terms of both normalized mean square error (NMSE) and Structure SIMilarity (SSIM). The superiority becomes even greater with the decrease of SNR in all cases, e.g., SSIM of the EEMD-ICA can almost double that of AWICA and triple that of ICA. To further examine the potentials of the approach in sophisticated applications, the approach together with the counterparts were used to preprocess a real-life epileptic EEG with absence seizure. Experiments were carried out with the focus on characterizing the dynamics of the data after artifact rejection, i.e., distinguishing seizure-free, pre-seizure and seizure states. Using multi-scale permutation entropy to extract feature and linear discriminant analysis for classification, the EEMD-ICA performed the best for classifying the states (87.4%, about 4.1% and 8.7% higher than that of AWICA and ICA respectively), which was closest to the results of the manually selected dataset (89.7%).