Because of the distance between the skull and brain and their different resistivities, electroencephalographic (EEG) data collected from any point on the human scalp includes activity generated within a large brain area. This spatial smearing of EEG data by volume conduction does not involve significant time delays, however, suggesting that the Independent Component Analysis (ICA) algorithm of Bell and Sejnowski (Bell and Sejnowski, 1995) is suitable for performing blind source separation on EEG data. The ICA algorithm separates the problem of source identification from that of source localization. First results of applying the ICA algorithm to EEG data collected during a sustained auditory detection task show: (1) ICA training is insensitive to different random seeds. (2) ICA analysis may be used to segregate obvious artifactual EEG components (line and muscle noise, eye movements) from other sources. (3) ICA analysis is capable of isolating overlapping alpha and theta wave bursts to separate ICA channels (4) Nonstationarities in EEG and behavioral state can be tracked using ICA analysis via changes in the amount of residual correlation between ICA-filtered output channels.

Because of the spread of electromagnetic signals through CSF and skull through volume conduction, EEG data recorded at different points on the scalp tend to be correlated. Bell and Sejnowski (1995) have recently presented an artificial neural network algorithm that identifies and separates statistically independent signals from a number of channels composed of linear mixtures of an equal number of sources. Here we present a first application of this Independent Component Analysis (ICA) algorithm to human EEG data. Conceptually, ICA filtering separates the problem of source identification in EEG data from the related problem of physical source localization. Three subjects performed a continuous auditory detection task in two half hour sessions. ICA filters trained on 14-channel EEG data collected during these sessions identified 14 statistically independent source channels which could then be further processed using event-related potential (ERP), event-related spectral perturbation (ERSP), and other signal processing techniques. One ICA source channel contained most eye movement activity, and another two collected line noise and muscle activity, while others were free of these artifacts. Changes in spectral power in several ICA channels covaried with changes in performance. If ICA sources can be shown to have distinct and consistent relationships to behavior or other physiological signals, ICA filtering may reveal meaningful aspects of event-related brain dynamics associated with sensory and cognitive processing but hidden within correlated EEG responses at individual scalp sites.

The problem of determining brain electrical sources from potential patterns recorded on the scalp surface is mathematically underdetermined. We have applied the Independent Component Analysis (ICA) algorithm of Bell and Sejnowski to the problem of source identification (What) considered apart from source localization (Where). By maximizing the joint entropy of a set of output channels derived from input signals by linear filtering without time delays, the ICA algorithm attempts to derive independent sources from highly correlated scalp EEG signals without regard to the locations or configurations (focal or diffuse) of the source generators. We report simulation experiments to determine (1) whether the ICA algorithm can successfully isolate independent components in simulated EEG generated by focal and distributed sources, and (2) whether ICA performance is severely affected by sensor noise and additional low-level brain noise sources. We will also show examples of ICA applied to actual EEG and cognitive ERP data.

Recordings of event-related potentials (ERPs) can reveal the time course of brain events associated with visual perception and selective attention. ERP studies of visual-spatial attention indicate that cortical processing of stimuli appearing in the attended location is augmented as early as 80 ms after stimulus onset. However, separation of the multiple brain processes contributing to the surface-recorded components of ERP waveforms has proven difficult. Recently, an `infomax' algorithm for the blind separation of linearly mixed inputs has been devised (Bell and Sejnowski, 1995) and applied to EEG and ERP analysis (Makeig et al., 1996). The neural generators of ICA sources are not specified by the algorithm and may be either physically compact or distributed.

Results of applying this Independent Component Analysis (ICA) algorithm to single-subject and group-mean ERPs recorded during a visual selective attention experiment (Anllo-Vento and Hillyard, 1996) suggest that ERP waveforms represent a sum of overlapping, discrete and time-limited brain processing events whose amplitudes are modulated by selective attention without affecting their time course. These source components identified by ICA appear to index independent stages of visual information processing. Spatial attention operates on early source components in a manner similar to a sensory gain-control mechanism, while later components appear to reflect further processing of stimulus features and feature conjunctions.

The problem of objectively decomposing event-related brain responses into neurophysiologically meaningful components is a major difficulty in the evoked response field. Traditional methods of identifying and measuring response subcomponents based on measuring the amplitudes and latencies of peak excursions in the waveforms at individual scalp sites fail when subcomponents overlap substantially, while current source localization procedures based on fitting single or multiple dipole models give ambiguous results when source geometry is unknown or complex. The Independent Component Analysis (ICA) algorithm of Bell and Sejnowski (1995) is an artificial neural network which maximizes the overall entropy of a set of non-linearly transformed input vectors using stochastic gradient ascent, without regard to the physical locations or configuration of the source generators. Trained on one or more multichannel electric or magnetic evoked responses, the algorithm converges on spatial filters which separate the input data into independent time courses and distinct scalp topographies arising in multiple, spatially-stationary 'effective brain sources.' Response decompositions produced by the ICA algorithm can be used to measure the effects of experimental manipulations on individual response components, even when these are overlapping in time or space. Typically, response components identified by the algorithm are recaptured in repeated analyses, regardless of changes in initial weights, sensor montage, and data length. I will explain the theory and practise of ICA decomposition and its differences from PCA, demonstrate results of EEG simulations, and present applications to EEG and MEG data analysis.

Event-related potentials (ERPs) can reveal the time course of brain events associated with visual perception and selective attention. ERP studies of visual-spatial attention indicate that cortical processing of stimuli appearing in the attended location is augmented as early as 80 ms after stimulus onset. However, separation of the multiple brain processes contributing to the surface-recorded components of ERP waveforms has proven difficult. Recently, an `infomax' algorithm for the blind separation of linearly mixed inputs has been devised (Bell and Sejnowski, 1995) and applied to EEG and ERP analysis (Makeig et al., 1996). The neural generators of ICA sources are not specified by the algorithm and may be either physically compact or distributed. Results of applying this Independent Component Analysis (ICA) algorithm to single-subject and group-mean ERPs recorded during a visual selective attention experiment (Anllo-Vento and Hillyard, 1996) suggest that ERP waveforms represent a sum of overlapping, discrete and time-limited brain processing events whose amplitudes are modulated by selective attention without affecting their time course. These source components identified by ICA appear to index independent stages of visual information processing. Spatial attention operates on early source components in a manner similar to a sensory gain-control mechanism, while later components appear to reflect further processing of stimulus features and feature conjunctions.

The Independent Component Analysis (ICA) algorithm of Bell and Sejnowski (Bell and Sejnowski, 1995) is an information-theoretic unsupervised learning algorithm which can be applied to the problem of separating multichannel electroencephalographic (EEG) data into independent sources (Makeig et al., 1996). We tested the potential usefulness of the ICA algorithm for EEG source decomposition by applying the algorithm to simulated EEG data. This data was constructed by projecting known input signals from single- and multiple-dipole sources in a three-shell spherical model head (Dale and Sereno, 1993) to simulated scalp sensors. In different simulations, we (1) altered the relative source strengths, (2) added multiple low-level sources (weak brain sources and sensor noise) to the simulated EEG, and (3) permuted the simulated dipole source locations and orientations. The algorithm successfully and reliably separated the activities of relatively strong sources from the activities of weaker brain sources and sensor noise, regardless of source locations and dipole orientations. These results suggest that the ICA algorithm should be able to separate temporally independent but spatially overlapping EEG activities arising from relatively strong brain and/or non-brain sources, irrespective of their spatial distributions.

Event-Related Potential (ERP) averages of electrical responses to sensory stimuli recorded at the human scalp capture voltage fluctuations both time locked and phase locked to occurrence of the stimuli. It is widely suspected, though poorly documented, that in single stimulus epochs the response activity may vary widely in both time course and scalp distribution. The major difficulty in comparing single trials is that the spontaneous EEG activity may obscure response-evoked activity, since spontaneous EEG is typically much larger than the evoked response. Independent Component Analysis (ICA) constructs spatial filters that can separate ERPs into spatially-fixed, temporally-sparse components that are temporally independent of one another. By adjusting the amount of ERP and single-trial EEG data used to train the algorithm, the resulting filters can separate larger EEG activity from ERP component activity, allowing a more accurate analysis of changes in the time course and/or the spatial distribution of ERP activity in single trials. Analysis of data from an auditory ERP experiment supports the observation that the relative amplitudes, latencies and scalp distributions of individual ERP components vary independently across single trials from the same subject and session. For example, a component composing N100 may be measurable in some but not all trials, independent of the presence of a component accounting for P300. This suggests that EEG and ERP activity may interact in ways that deserve further study. (Research supported by the Howard Hughes Medical Institute and the Office of Naval Research).

Independent Component Analysis (ICA) is a new signal processing technique for decomposing spontaneous or evoked electrophysiological data into spatially fixed and temporally independent components. ICA allows comparison of component amplitudes and time courses across related conditions. Applied simultaneously to target and nontarget responses in 30 conditions of a visual selective attention experiment (see Westerfield et al., this session), ICA derived at least four components of the early visual evoked response which were differently amplitude-modulated by spatial location and attention without effects on component latency, but were not affected by the target/nontarget distinction. Other components accounted for data artifacts in single conditions. Later components common to several conditions were sensitive to both spatial attention and target feature. ICA allows quantitative comparison of objectively-derived and temporally-sparse ERP components and subcomponents across 30 or more stimulus or task conditions. (Supported by the Office of Naval Research, Howard Hughes Medical Institute, NINDS NS34155 and NIMH MH36840)

Independent Component Analysis (ICA) is a new signal processing technique for decomposing spontaneous or evoked EEG and MEG data into temporally independent and spatially fixed components. The scalp distribution of the auditory steady-state response near 40 Hz appears to sweep from the front to the back of the scalp every cycle. ICA decomposes this apparent movement into the sum of at least two bilateral components with different scalp distributions and phase lags. ICA accounts for the transient perturbations in SSRs produced by experimental events using the same components producing the SSR, supporting the hypothesis that these transient (CERP) perturbations represent modulation of the ongoing response. Application of ICA algorithms capable of both sub-Gaussian and super-Gaussian components will be presented and psychophysiological implications of new blind decomposition techniques discussed. (Research supported by the Howard Hughes Medical Institute and the Office of Naval Research).

A recently-derived 'infomax' algorithm for performing Independent Component Analysis (ICA) is a new information-theoretic approach to the problem of separating multichannel electroencephalographic (EEG) or magnetoencephalographic (MEG) data into temporally independent and spatially stationary sources. In a previous report, we have shown that the algorithm can separate simulated EEG source waveforms (independent simulated brain source activities mixed linearly at the scalp sensors), even in the presence of multiple low-level model brain and sensor noise sources. Here, we demonstrate the ability of the ICA algorithm to decompose brief event-related potential (ERP) data sets into temporally independent components by applying it to simulated ERP-length EEG data synthesized from 3-sec (600-point) electrocorticographic (ECoG) epochs recorded from the cortical surface of a human undergoing pre-surgical evaluation.

Six asynchronous single-channel ECoG data epochs were projected through single- and multiple-dipole model sources in a three-shell spherical head model to six simulated scalp sensors to create simulated EEG data. In two sets of simulation experiments, we altered relative source strengths, added multiple low-level sources (synthesized from ECoG data and uniform- or Gaussian-distributed noise), and permuted the simulated dipole source locations and orientations. The algorithm reliably separated the activities of the relatively strong sources, regardless of source location, dipole orientation, and low-level source distributions. Thus, the ICA algorithm should identify relatively strong, temporally independent and spatially overlapping ERP components arising from multiple brain and/or non-brain sources, regardless of their spatial distributions. This shows that the ICA algorithm can decompose ERPs generated by uncorrelated sources.

A third ERP simulation tests how the algorithm treats a simulated ERP epoch constructed using model ERP generators whose activations are partially correlated. In this case, the algorithm parsed the simulated ERP waveforms into a sum of temporally independent and spatially stationary components reflecting the changing topography of correlated source activity in the simulated ERP data. Each of the affected components sums activity from one or more concurrently-active brain generators. This suggests the ICA algorithm may also be useful for identifying event-related changes in the correlation structure of either spontaneous or event-related EEG data. Paradoxically, adding four simulated ``no response'' epochs to the training data minimized the relative importance of partial correlations in the original data epoch and allowed the algorithm to separate the concurrently active sources. Likewise, submitting ERPs from more than one stimulus or experimental condition to concurrent ICA analysis may allow the algorithm to separate sources from brain generators whose activations are partially correlated in some but not all response conditions.