Joint spatial feature adaption and confident pseudo-label selection for cross-subject motor imagery EEG signals classification

Abstract

Motor imagery electroencephalograph (MI-EEG) classification plays an important role in noninvasive brain-computer interfaces (BCIs). However, the distribution shifts among different subjects make a major challenge to build classification models. Due to temporally-varying and spatially-coupling characteristics of MI-EEG data, recent methods have suffered from incomplete feature representations and the accumulation of incorrect pseudo-labels, even lower efficiency. To address these issues, the paper proposes a novel method for cross-subject MI-EEG classification, namely Joint spatial Feature Adaptation and Confident Pseudo-label Selection (JFACPS). JFACPS extracts joint spatial feature representations from two perspectives upon the aligned MI-EEG samples, where the spatio-temporal filtering features are extracted upon Euclidean space and the tangent space mapping features are extracted upon Riemannian space. Then, the joint spatial features are incorporated into a discriminative pseudo-labeling framework for feature adaptation. Among them, the samples with large differences in confidence between the highest and second-highest predictions are selected for adaptation. Meanwhile, a novel classifier is introduced to initialize more accurate pseudo-labels with high confidence during the first iteration of feature adaptation. We systematically conducted the experiments on two benchmark MI-EEG datasets, and the classification performance of JFACPS surpasses several state-of-the-art methods. Moreover, ablation studies also demonstrated the significance for both joint spatial feature and confident pseudo-label selection. Based on the parameter insensitivity experiments, our JFACPS method provides a novel calibration option for new subjects participating in MI-BCIs.