Application of self-adaptive multiple-kernel extreme learning machine to improve MI-BCI performance of subjects with BCI illiteracy.

• It is difficult to extract common and recognizable features from BCI illiterate subjects who do not exhibit typical brain activity. Inspired by the kernel trick, we proposed a multiple kernel extreme learning machine framework, which can map the input features to multi-dimensional nonlinear spaces as much as possible, so as to increase the separability probability of the features of BCI illiteracy. • We summarize the following two conclusions: the more kernels were used, the better the BCI performance is, especially for the BCI illiteracy; non-sparse multiple kernel learning can usually outperform the sparse form. As mentioned above we proposed the linear combination of four kernels in non-sparse form. • We employed differential evolution (DE) to find the optimal initial parameters of the classifier for the purpose of obtaining the optimal classifier suitable for the current system. • In comparison with the state-of-the-art classifiers, the experimental results showed that the performance of our method outperformed other control methods, especially for the BCI illiterate subjects. Electroencephalography (EEG)-based brain-computer interface (BCI) allows interactions between the brain and the external world. However, some potential subjects have "BCI illiteracy": they cannot control BCI devices effectively. Usually, these subjects do not exhibit typical brain activity, and there are large variabilities between subjects. It is difficult to extract common and recognizable features from these subjects using current methods. Inspired by the kernel trick, data can be mapped to high-dimensional space to increase separability, we propose the application of a classifier based on kernel functions. Due to the selection of optimal kernel, one kernel was replaced by multiple kernel who are combined by weight. In this way, the features can be mapped to multiple spaces to explore the latent features of motor imagery-BCI—especially BCI Illiteracy—in multidimensional and nonlinear spaces. Meanwhile, differential evolution (DE) was employed to find the optimal initial parameters for the purpose of obtaining the classifier suitable for current system. We applied the proposed method to the public BMI-dataset for the study of BCI illiteracy. Compared with the given dataset, the grand average classification accuracy is 70.06%, which is 2.8% higher than that achieved via reference method. Especially, for the sample set of BCI illiteracy, the grand average classification accuracy is 57.88%, exceeding that obtained through reference method by 3.86%. In comparison with the state-of-the-art classifiers, the average classification accuracy of our method is 1.23% more than that of the best control classifier. [ABSTRACT FROM AUTHOR]