Blind Separation of Event-Related Brain Responses into Independent Components

Abstract

Functional imaging of brain activity based on changes in blood flow does not supply information about the relative timing of brief bursts of neural activity in different brain areas 1 ' 2 . Multichannel electric or magnetic recordings from the scalp provide high temporal resolution, but are not easily decomposed into the separate activities of multiple brain networks.We report here a method for the blind separation of event-related brain responses into spatially stationary and temporally independent subcomponents using an Independent The 'unmixing' matrix W performs source separation, while the sigmoidal nonlinearity 5Ü provides necessary higher-order statistical information. Initial sphering of the zero-mean input data 12 (x s = Px, where P = 2(xx T )-2) speeds convergence. W is then initialized to the identity matrix (I) and iteratively adjusted using small batches of data vectors drawn randomly from {x,} without substitution, according to:where e is the learning rate and vector y has elements When ICA is trained on EEG data (Fig. 1), the rows of the resultant matrix (WP)are linear spatial filters which, applied to the input data, produce source activity waveforms (WPx). The columns of the inverse weight matrix (WP)" 1 represent the projection weights from the ICA sources to the sensor array. ICA sources may be spatially distributed, so they need not correspond to single peaks in functional imaging measurements. The minimum number of time points needed for the method appears to be several times the number of recording.channels, which in turn must.be larger than the number of sources to be separated.We first applied the ICA algorithm to 300-s EEG epochs to demonstrate that the algorithm can segregate brain from muscle activity and can track psychophysiological state changes 19 . To test the performance of the ICA algorithm in separating multiple strong and weak brain sources, we then performed numerical simulations using a three-shell spherical head model 20 . Results suggested that the algorithm can successfully separate the independent activities of stronger brain sources from those of weaker signal and noise sources, even when they overlap in time and space 21 .Next, we applied the ICA algorithm to 1-s ERPs time locked to the detected and undetected noise-burst targets (Fig. 2a) presented in an experiment in which a subject was asked to detect weak noise-burst targets continuously during a half-hour session 4 . Results of the analysis consisted of 15 ICA sources (Fig. 2b), including one Hz click train delivered continuously during the experiment. The projected scalp distributions of all these sources were distinct (although not orthogonal) and summed to the scalp distributions of the corresponding response components (Fig. 3). We presume the four remaining sources mixed activity from multiple weak sources.Both the activity waveforms and the scalp projections of the ICA sources with-, largest scalp projections proved robust to changes in initial weights, number and placement of electrodes, epoch lengths, and ...

Blind separation of auditory event-related brain responses into independent components

“…1B). Principal components of data generated by temporally sparse and independent, but spatially nonorthogonal, sources will be linear combinations of activity in all the sources, whereas ICA components of the data will individually identify the larger sources (28). The proposed Varimax extension of the PCA method rotates the PCA vectors to maximize the variance of their activation waveforms (4).…”

“…To explore the strengths and limitations of the method, we ran a number of numerical simulations in which 600-point signals recorded from the cortex of a patient during preparation for operation for epilepsy were projected to simulated scalp electrodes through a three-shell spherical model (28,29). We used electrocorticographic data in these simulations as a plausible best approximation to the temporal dynamics of the unknown ERP brain generators.…”

“…However, given simulated ERP activity arising from separate brain generators whose time courses of activation were substantially correlated, the algorithm parsed the resulting continuously varying scalp responses into distributed activity within overlapping subsets of the simulated sources (28). Similarly, SSR components ICA-8 and ICA-9 (Fig.…”

Neural synchrony indexes disordered perception and cognition in schizophrenia

“…Error and no-response trials were excluded from analyses. Independent component analysis (19) was used to correct for eyeblink artifacts. Corrected single-trial epochs were rereferenced to averaged mastoids.…”

Topographic phase maps using iterative independent component analysis

“…The method developed, termed iterative ICA (iICA), is an iterative implementation of the information maximization algorithm originally proposed by Makeig et al [6]. We have shown that iICA can provide estimates of a particular EP component out of an entire waveform, and this EP component is made clearly visible in each single trial [4,15].Thus, the method allows studying the dynamic evolution of the cortical generators that give rise to speci c EP components.…”