A Weighted Combination Method of Multiple K-Nearest Neighbor Classifiers for EEG-Based Cognitive Task Classification

Abstract

The effectiveness of wireless electroencephalograph (EEG) sensor-based medical or Brain Computer Interface applications largely depends on how to classify EEG signals as accurately as possible. Empirical studies show that EEG channels respond differently to different cognitive tasks. Thus, to effectively classify cognitive tasks, the channel cognitive sensitivity should be taken into account during the classification process. In this paper, we propose a weighted combination of multiple-nearest neighbor approach for cognitive task classification. Each EEG channel is assigned a weight that reflects its sensitivity to the cognitive task space. First, for each channel, a nearest neighbor algorithm is performed and an output is produced. To combine all the channel outputs, a modified aggregation method is utilized such that the weights assigned to the channels are accommodated. Experimental work shows that the proposed technique achieved 96.7% classification accuracy utilizing all the available channels, and 96.4% and 92.7% classification accuracies utilizing only 70% and 60% of the available channels, respectively, for subject 1; and 99.4% classification accuracy utilizing all the available channels, and 98.9% and 97.6% classification accuracies utilizing only 70% and 60% of the available channels, respectively, for subject 2.