Content-based EEG database retrieval using a multiclass SVM classifier

CBER (content-based-EEG-retrieval) systems present short data segments as query samples for similar segments in EEG databases. These systems have many applications in large-scale data-mining, but require effective and verifiable retrieval strategies. This paper introduces a new feature strategy based on class probabilities calculated by LIBSVM classification using low-order autoregressive (AR) modeling of the original signal. This second-level feature strategy appears to be effective in capturing structural relationships among classes. We demonstrate the effectiveness of the approach for three retrieval problems: retrieval of segments corresponding to the same task, the same task group, and the same subject. The new approach produces features of much lower dimension and better performance than AR features.