A Stepwise Discriminant Analysis and FBCSP Feature Selection Strategy for EEG MI Recognition

Abstract

Accurate decoding of brain intentions is a pivotal technology within Brain-Computer Interface (BCI) systems that rely on Motor Imagery (MI). The effective extraction of information features plays a critical role in the precise decoding of these brain intentions. However, there exists significant individual and environmental variability in signals, and the sensitivity of EEG signals from different subjects also varies, imposing higher demands on both feature exploration and accurate decoding. To address these challenges, we employ adaptive sliding time windows and a stepwise discriminant analysis strategy to selectively extract features obtained through the Filter Bank Common Spatial Pattern (FBCSP). This entails the identification of an optimal feature combination tailored to specific patients, thereby mitigating individual differences and environmental variations. Initially, adaptive sliding time windows are applied to segment electroencephalogram (EEG) data for different subjects, followed by FBCSP for feature extraction. Subsequently, a stepwise discriminant analysis (SDA) incorporating prior knowledge is employed for optimal feature selection, effectively and adaptively identifying the best feature combination for specific subjects. The proposed method is evaluated using two publicly available datasets, the EEG recognition accuracy for Dataset A is 98.47%, and for Dataset B, it is 95.2%. In comparison to current publicly reported research results (utilizing Power Spectral Density (PSD) + Support Vector Machine (SVM) methods) for Dataset A, the proposed method improves MI recognition accuracy by 25.37%. For Dataset B, compared to current publicly reported results (FBCNet method), the proposed method improves MI recognition accuracy by 26.4%. The experimental results underscore the method's broad applicability, scalability, and substantial value for promotion and application.