Validating online recursive independent component analysis on EEG data

The needs for online Independent Component Analysis (ICA) algorithms arise in a range of fields such as continuous clinical assessment and brain-computer interface (BCI). Among the online ICA methods, online recursive ICA algorithm (ORICA) has attractive properties of fast convergence and low computational complexity. However, there hasn´t been a systematic comparison between an online ICA method such as ORICA and other offline (batch-mode) ICA algorithms on real EEG data. This study compared ORICA with ten ICA algorithms in terms of their decomposition quality, validity of source characteristics, and computational complexity on the thirteen experimental 71-ch EEG datasets. Empirical results showed that ORICA achieved higher mutual information reduction (MIR) and extracted more near-dipolar sources than algorithms such as FastICA, JADE, and SOBI did while the performance of ORICA approached that of the best-performed Infomax-based algorithms. Furthermore, ORICA outperforms most of ICA methods in terms of the computational complexity. The properties of fast convergence and low computational complexity of ORICA enable the realization of real-time online ICA process, which has further applications such as real-time functional neuroimaging, artifact reduction, and adaptive BCI.