Improving performance of asynchronous BCI by using a collection of overlapping sub window models

Abstract

Asynchronous Brain Computer Interfaces (BCI) have become an interesting topic in the present days because they provide simulation of realistic usage of BCI. For asynchronous BCI, the computer has to discriminate not only differences among various imaginary tasks but also detect relax periods. Since the training phase for building a classification model is still synchronous (cue-based), the main challenge is to classify the EEG signal continuously with good accuracy on asynchronous (uncue-based). This paper addresses achieving better performance by using a collection of overlapping sub windows models. A model is referred to a primitive classification model which consists of common spatial patterns (CSP) with linear discriminant analysis (LDA). Each primitive model was trained with the corresponding sub window indexes. We had 3 collections of models: task1 vs. task2, task1 vs. relax, and task2 vs. relax. These binary classification results were then fused together with Mahalanobis distance to gain better performance. The results were measured by mean square error (MSE), and their performance is better compared to the primitive model. Furthermore, the results on the test set were comparable to the 3 leading scores of BCI Competition IV dataset 1.