Robust EEG feature learning model based on an adaptive weight and pairwise-fused LASSO

Abstract

Epilepsy is a serious neurological disorder that affects almost 70 million people worldwide. Electroencephalography (EEG), as an important research tool for epilepsy detection, has been studied extensively in the literature. In this paper, an adaptive feature learning model for EEG recordings based on an adaptive weight and pairwise-fused LASSO (AWPF-LASSO) is proposed. The Fourier spectra of EEG recordings are used as the original features. Our main innovation is that two highly flexible penalization terms are taken into account in the proposed model. On the one hand, an adaptive weight penalization, which considers the contribution of different features by multiplying each variable's corresponding regression coefficient by an adaptive weight, is introduced. On the other hand, considering that the concentration degree of EEG spectral features varies in different categories, the new model further introduces a pairwise-fused penalization, which enables group selection of variables by utilizing the correlation information of the data. Moreover, the new model can be solved by the coordinate descent algorithm with computational complexity of O(np), where n is the sample number and p is the feature number. In addition, the coordinate descent algorithm effectively ensures convergence. By applying the AWPF-LASSO model, distinctive EEG features that represent the essentials of the data can be extracted. Experiments show that the newly extracted features perform well under different classifiers. Furthermore, the proposed model is robust and can yield an accuracy of more than 98.5% even when EEG data are corrupted by white noise and EEG artifacts with different SNR levels.