Global optimal constrained ICA and its application in extraction of movement related cortical potentials from...

The constrained ICA (cICA) is a recent approach which can extract the desired source signal by using prior information. cICA employs gradient-based algorithms to optimize non convex objective functions and therefore global optimum solution is not guaranteed. In this study, we propose the Global optimal constrained ICA (GocICA) algorithm for solving the conventional cICA problems. Due to the importance of movement related cortical potentials (MRCPs) for neurorehabilitation and developing a suitable mechanism for detection of movement intention, single-trial MRCP extraction is presented as an application of GocICA.

In order to evaluate the performance of the proposed technique, two kinds of datasets including simulated and real EEG data have been utilized in this paper. The GocICA method has been implemented based on the most popular meta-heuristic optimization algorithms such as Genetic Algorithm (GA), Particle Swarm Optimization (PSO) and Charged System Search (CSS) where the results have been compared with those of conventional cICA and two ICA-based methods (JADE and Infomax).

It was found that GocICA enhanced the extracted MRCP from multi-channel EEG better than both conventional cICA and ICA-based methods and also outperformed them in single-trial MRCP detection with higher true positive rates (TPRs) and lower false positive rates (FPRs). Moreover, CSS-cICA resulted in the greatest TPR (91.2232 ± 3.4708) and the lowest FPR (8.7465 ± 3.7705) for single-trial MRCP detection from real EEG data and the greatest signal-to-noise ratio (SNR) (39.2818) and the lowest mean square error (MSE) and individual performance index (IPI) (41.8230 and 0.0012, respectively) for single-trial MRCP extraction from simulated EEG data.

These results confirm the superiority of GocICA with respect to conventional cICA that is due to the ability of meta-heuristic optimization algorithms to escape from local optimal point. As such, GocICA is a promising new algorithm for single-trial MRCP detection which can be used for detecting other types of event related cortical potentials (ERPs) such as P300 and also for EEG artifact removal.