Random Forest and Filter Bank Common Spatial Patterns for EEG-Based Motor Imagery Classification

We propose using filter bank common spatial pattern (FBCSP) feature extraction algorithm, and random forest (RF) technique for classification of EEG motor imagery signals. FBCSP algorithm allows extracting features and dealing with subject variability by automatic selection of frequency bands. Performing random forest in the classification avoid the use of feature selection step, since RF combine a bagging (bootstrap aggregation) and a random selection of features. We evaluate our system on the dataset 2b of the Brain-Computer Interface BCI Competition IV. The proposed method is promising since it has outperformed the results obtained in BCI Competition for some subjects in term of accuracy and kappa.