Random forest classification for p300 based brain computer interface applications

One of the most successful types of brain computer interfaces (BCI) is based on the P300 evoked potential (EP) elicited by oddball type of paradigms. Given a particular paradigm the main challenge is to obtain an efficient and robust classification. This paper proposes the use of Random Forest (RF), a tree based ensemble learning method providing state-of-the-art generalization performance, for P300 BCI classification. The performance of the proposed method is compared to both the most commonly used classifiers for this problem: the support vector machine (SVM), and the step-wise linear discriminant analysis (SWLDA); and to two state-of-the-art methods: the multiple convolutional neural networks (MCNN) and the ensemble support vector machine (ESVM). The proposed method has been evaluated on two public available BCI datasets: the BCI competition dataset II for healthy subjects and the image driven paradigm dataset for disabled subjects. The proposed method demonstrated a significant improvement in classification accuracy on both datasets.