Time sparsification of EEG signals in motor-imagery based brain computer interfaces

We propose a method of sparsifying EEG signals in the time domain for common spatial patterns (CSP) which are often used for feature extraction in brain computer interfaces (BCI). For accurate classification, it is important to analyze the period of time when a BCI user performs a mental task. We address this problem by optimizing the CSP cost with a time sparsification that removes unnecessary samples from the classification. We design a cost function that has CSP spatial weights and time window as optimization parameters. To find these parameters, we use alternating optimization. In an experiment on classification of motor-imagery EEG signals, the proposed method increased classification accuracy by 6% averaged over five subjects.