Common Spatial Pattern with Feature Scaling (FSc-CSP) for Motor Imagery Classification

Abstract

Brain-Computer Interface (BCI) is a way to translate human thoughts into computer commands. One of the most popular BCI type is Electroencephalography (EEG)-based BCI, where motor imagery is considered one of the most effective ways. Previously, to extract useful information, various filters are introduced, such as spatial, temporal, and spectral filtering. A spatial filtering algorithm called Common Spatial Pattern (CSP) was developed and known to have excellent performance, especially in motor imagery for BCI application. In general, there are several approaches in improving CSP such as regularization approach, analytic approach, and frequency band selection. In general, the existing techniques for band selection is either to select or reject the band by ignoring the importance of the band. For example, Binary Particle Search Optimization Common Spatial Pattern (BPSO-CSP) was proposed to choose multiple possible best bands to be used in processing the data. In this paper, we propose an algorithm called Feature Scaling Common Spatial Pattern (FSc-CSP) to overcome the problem of feature selection. Instead of selecting features, the proposed algorithm employs a feature scaling system to scale the importance of each band by using Genetic Algorithm (GA) altogether with Extreme Learning Machine (ELM) as classifier, with 1 signifying the most important bands, declining until 0 for the unused bands, as opposed to the 1 and 0 selection system used in BPSO-CSP. Conducted experiments show that by employing feature scaling, better results can be achieved especially compared to vanilla CSP and feature selection with 100 hidden nodes in three from five BCI Competition III datasets IVa, namely aa, aw and ay, with around 5–8% better results compared to vanilla CSP and feature selection.