Journal Home Online First Current Issue Archive For Authors Journal Information 中文版

Frontiers of Information Technology & Electronic Engineering >> 2015, Volume 16, Issue 6 doi: 10.1631/FITEE.1400299

Fast removal of ocular artifacts from electroencephalogram signals using spatial constraint independent component analysis based recursive least squares in brain-computer interface

Shanghai Key Laboratory of Power Station Automation Technology, Department of Automation, School of Mechatronic Engineering and Automation, Shanghai University, Shanghai 200072, China

Available online: 2016-01-05

Next Previous

Abstract

Ocular artifacts cause the main interfering signals within electroencephalogram (EEG) signal measurements. An adaptive filter based on reference signals from an electrooculogram (EOG) can reduce ocular interference, but collecting EOG signals during a long-term EEG recording is inconvenient and uncomfortable for the subject. To remove ocular artifacts from EEG in brain-computer interfaces (BCIs), a method named spatial constraint independent component analysis based recursive least squares (SCICA-RLS) is proposed. The method consists of two stages. In the first stage, independent component analysis (ICA) is used to decompose multiple EEG channels into an equal number of independent components (ICs). Ocular ICs are identified by an automatic artifact detection method based on kurtosis. Then empirical mode decomposition (EMD) is employed to remove any cerebral activity from the identified ocular ICs to obtain exact artifact ICs. In the second stage, first, SCICA applies exact artifact ICs obtained in the first stage as a constraint to extract artifact ICs from the given EEG signal. These extracted ICs are called spatial constraint ICs (SC-ICs). Then the RLS based adaptive filter uses SC-ICs as reference signals to reduce interference, which avoids the need for parallel EOG recordings. In addition, the proposed method has the ability of fast computation as it is not necessary for SCICA to identify all ICs like ICA. Based on the EEG data recorded from seven subjects, the new approach can lead to average classification accuracies of 3.3% and 12.6% higher than those of the standard ICA and raw EEG, respectively. In addition, the proposed method has 83.5% and 83.8% reduction in time-consumption compared with the standard ICA and ICA-RLS, respectively, which demonstrates a better and faster OA reduction.

Related Research