Analysis of Large-Scale Brain Data for Brain-Computer Interfaces

Abstract

We present a systematic technique for extraction of useful information from large-scale neural data in the context of brain-computer interfaces. The technique is based on a direct linear discriminant analysis, recently developed for face recognition problems. We show that this technique is capable of extracting useful information from brain data in a systematic fashion and can serve as a general analytical tool for other types of biomedical data, such as images and collections of images (movies). The performance of the method is tested on intracranial electroencephalographic data recorded from the human brain.

Mental State Estimation for Brain--Computer Interfaces

“…RLDA and CSP perform best on subjects S006 and S004, respectively, whereas DLDA and especially PLDA perform poorly. This is consistent with previous studies [29], [48], [50], where these and other face recognition techniques were found ill-suited for classification of noisy biomedical data. Also note that the computational cost of RLDA is an order of magnitude higher than those of the CPCA, PLDA, and CSP techniques.…”

“…Since spatiotemporal ECoG data are essentially an image, it may be worthwhile to cast a classification of ECoG data in the framework of image/face recognition. However, when applied to ECoG data, several representative face recognition techniques produced very modest results [29], [48], [50]. The discrepancy in performance may be attributed to relatively high levels of noise in electrophysiological neural data.…”

“…It is interesting to note that the hippocampus and parietal regions did not seem to carry significant state information, as evidenced by the lack of important coefficients over these areas. These areas, however, typically exhibit directionally selective response and are, therefore, more suitable to decoding of the target location [48], [50], [68].…”

A Classwise PCA-based Recognition of Neural Data for Brain-Computer Interfaces

“…To obtain meaningful data statistics, this irrelevant subspace must be discarded as noise, which is typically accomplished by global dimensionality reduction [5], and variants thereof [9], [10], have been successfully used. However, our experience with face recognition techniques applied to brain data were somewhat disappointing [7]. Motivated by these limitations, we proposed a modification of the direct LDA (DLDA) method [9], by introducing a threshold [7], whose role was to regularize the covariance matrix in a manner similar to the standard shrinkage approach [11].…”

“…While the power/frequency representation is physically intuitive, it is unclear why these ad hoc features should have optimal predictive power. Several studies [6], [7], [8] report significantly better decoding results with the use of other (more abstract) features. Another common strategy, often used in conjunction with the above, is to rank individual features (or recording electrodes) according to some criterion.…”