Abstract
One of the challenges in brain computer interface (BCI) research is achieving subject independence. A second challenge arises from the reduced accuracy as the number of classes increases. Recognizing that ensemble and multimodal classifiers improve upon single mode models, we introduce an ensemble-based approach for subject-independent BCI (SIBCI), and present its results on a range of fusion methods at different input and output stages. Using Support Vector Machine (SVM) classifiers, a range of different EEG features (modalities) were derived and ranked based on their subject-invariance. An ensemble of SVM classifiers, each operating on one input modality and combining right and left hemisphere data, were fused using error correcting codes and majority vote. Single mode classifiers provided accuracies ranging from slightly better than chance (25% for a four-class model), to 56.9%. By fusing classifier votes, our multimodal ensemble was able to achieve a subject independent classification accuracy of up to 70.6% on a four quasi movement test dataset; a 24% improvement over our best single-mode classifier.
