Independent Component Analysis of Electroencephalographic Data

Scott MakeigAnthony J. BellTzyy-Ping JungTerrence J. Sejnowski

article1995NeurIPS2,408 citations

Demonstrates how independent component analysis can effectively separate mixed scalp electroencephalographic signals into distinct neurobiological sources and artifacts without requiring prior spatial information.

Listen

Monitoring brain activity and cognitive states using scalp electroencephalography (EEG) is a critical tool for understanding operator alertness and cognitive performance in operational settings. However, interpreting EEG signals is difficult because the skull and tissue smear electrical signals across the scalp, causing sensors to record complex mixtures of multiple brain processes and non-brain noise. The article set out to evaluate whether Independent Component Analysis (ICA), an information-maximization algorithm for blind source separation, can effectively separate mixed EEG signals into distinct, statistically independent source activities without needing to model physical brain anatomy.

The authors applied the ICA algorithm to 14-channel recordings collected from adult subjects performing a sustained auditory detection task designed to induce drowsiness and lapses in attention. The dataset included continuous EEG recordings and event-related potentials (ERPs) mapped across several half-hour testing sessions. By maximizing statistical entropy, the algorithm learned spatial filter weights to extract underlying source components, with training stability verified across different initial parameters.

The analysis produced four central findings. First, ICA isolated distinct artifact sourcessuch as line noise at 60 Hz, scalp muscle tension across 50 to 100 Hz, and slow eye movementsinto separate individual channels, effectively cleaning the remaining brain signals. Second, the method unmixed overlapping neural rhythms, isolating prominent 7-Hz theta waves and revealing 10-Hz alpha bursts that were not obvious in the raw recordings. Third, ICA successfully decomposed complex ERPs into temporally overlapping subcomponents corresponding to detected target hits, lapses, and steady-state auditory responses. Fourth, the training proved robust to random starting points, and tracking changes in residual correlations across output channels successfully reflected subject transitions between alert and drowsy behavioral states.

These findings show that ICA provides an objective, data-driven approach to human electrophysiology by separating what brain activity occurs from where it is located. For operational monitoring and clinical research, this capability reduces the risk of data misinterpretation caused by muscle and ocular noise, while enabling automated detection of shifts in cognitive alertness. By incorporating higher-order statistics rather than simple decorrelation, the approach provides unique decompositions of complex brain responses.

Moving forward, researchers and developers should incorporate ICA as an automated preprocessing and feature extraction tool in cognitive monitoring systems. Future work should focus on determining the appropriate number of input channels for different recording setups and validating the physiological meaning of derived components against established neural models. While the results demonstrate high stability in separating mixed signals, users should exercise caution regarding the assumption that the number of brain sources matches the number of recording sensors, as well as the algorithm's simplification of complex variable brain synchronization.

  • Paper: A Tutorial on Principal Component Analysis, Jonathon Shlens (2014). Read this tutorial on Principal Component Analysis first to master the linear algebraic foundations of dimensionality reduction that underpin blind source separation.
Cover for Independent Component Analysis of Electroencephalographic Data

Abstract

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 [1] 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 and event-related potential (ERP) data collected during a sustained auditory detection task show: (1) ICA training is insensitive to different random seeds. (2) ICA may be used to segregate obvious artifactual EEG components (line and muscle noise, eye movements) from other sources. (3) ICA is capable of isolating overlapping EEG phenomena, including alpha and theta bursts and spatially-separable ERP components, to separate ICA channels. (4) Nonstationarities in EEG and behavioral state can be tracked using ICA via changes in the amount of residual correlation between ICA-filtered output channels.

Table of Contents

  • 1 Introduction
  • 1.1 Separating What from Where in EEG Source Analysis
  • 1.2 Independent Component Analysis
  • 1.3 Applying leA to EEG Data
  • 2 Methods
  • 3 Results
  • 4 Conclusions
  • Acknowledgments
  • References

Knowls

  1. Knowl 1 — Infomax Independent Component Analysis for EEG Signal Decomposition

    model/method

    The application of Infomax Independent Component Analysis (ICA) to multi-channel electroencephalographic (EEG) data aims to recover statistically independent source activations from scalp electrode recordings without prior anatomical source models.

    Let x=[x1,,xN]TRN\mathbf{x} = [x_1, \dots, x_N]^T \in \mathbb{R}^N be an NN-dimensional vector representing instantaneous scalp electrode potentials. The linear transformation derives NN independent source components u=[u1,,uN]TRN\mathbf{u} = [u_1, \dots, u_N]^T \in \mathbb{R}^N via an unmixing weight matrix WRN×N\mathbf{W} \in \mathbb{R}^{N \times N} and a bias vector wRN\mathbf{w} \in \mathbb{R}^N: u=Wx+w\mathbf{u} = \mathbf{W}\mathbf{x} + \mathbf{w} such that the joint probability density function factorizes: fu(u)=i=1Nfui(ui)f_{\mathbf{u}}(\mathbf{u}) = \prod_{i=1}^N f_{u_i}(u_i), thereby minimizing mutual information I(ui,uj)=0I(u_i, u_j) = 0 for all iji \neq j.

    The unmixing parameters are optimized by maximizing the joint entropy H(y)H(\mathbf{y}) of the squashed outputs yi=(1+eui)1y_i = (1 + e^{-u_i})^{-1} using stochastic gradient ascent: ΔW[WT]1+y^xT\Delta \mathbf{W} \propto [\mathbf{W}^T]^{-1} + \hat{\mathbf{y}}\mathbf{x}^T Δwy^\Delta \mathbf{w} \propto \hat{\mathbf{y}} where y^=[y^1,,y^N]T\hat{\mathbf{y}} = [\hat{y}_1, \dots, \hat{y}_N]^T has elements defined by: y^i=yi(yiui)=12yi\hat{y}_i = \frac{\partial}{\partial y_i}\left(\frac{\partial y_i}{\partial u_i}\right) = 1 - 2y_i

    For continuous EEG implementations:

    1. Zero-mean input data are pre-whitened to remove first- and second-order correlations.
    2. Time indices are randomly permuted across the recording to ensure signal stationarity across training passes.
    3. The learning rate is annealed from 0.030.03 down to 0.00010.0001.
    4. Training terminates when the average Pearson correlation across all output channel pairs is below 0.050.05 and the weight matrix W\mathbf{W} reaches convergence.
  2. Knowl 2 — Physical and Statistical Assumptions of Scalp EEG Independent Component Analysis

    assumption

    Applying Independent Component Analysis (ICA) to scalp electroencephalography relies on four specific physical and statistical assumptions:

    1. Statistical Independence: Complex multi-channel EEG signals can be represented as linear combinations of a finite set of statistically independent cerebral and non-cerebral sources.
    2. Instantaneous Volume Conduction: Electromagnetic propagation through the brain, cerebrospinal fluid, skull, and scalp occurs at near light speed with negligible time delays, making the mixing process strictly linear and instantaneous rather than convolutive.
    3. Sigmoidal Cumulative Density Functions: The underlying signal sources are continuous/analog variables whose probability density functions are well-approximated by the derivative of a standard logistic sigmoid function.
    4. Equal Source and Sensor Dimensionality: The number of dominant independent signal generators equals the number of recording sensors (NN electrodes yields NN separable source channels).

    Condition of validity: The standard ICA model assumes spatial stationarity of the mixing matrix and neglects dynamic, state-dependent variations in synchronization among distinct cortical or subcortical generators.

  3. Knowl 3 — Separation of Source Identification from Source Localization in EEG Analysis

    model/method

    Traditional approaches to EEG source analysis couple the temporal identification of neural generators with 3D anatomical localization using inverse electrostatic modeling, which is mathematically ill-posed and underdetermined. ICA completely decouples these two objectives:

    1. Source Identification ("What"): Discovers the temporal dynamics and waveforms of independent potential-generating systems directly from statistical higher-order dependencies in the raw electrode recordings, making no assumptions about head geometry, tissue resistivities, or generator spatial configurations.
    2. Source Localization ("Where"): Treats each identified independent component as a spatially fixed projection pattern (either compact/focal or widely distributed across cortex) that can subsequently be localized anatomically or mapped across the scalp independently of other active sources.
  4. Knowl 4 — Blind Separation of Spontaneous EEG Rhythms and Extracerebral Artifacts

    empirical result

    When evaluated on 14-channel recordings (13 scalp EEG channels and 1 diagonal bipolar EOG channel), Infomax ICA segregates distinct cerebral rhythms and isolates multiple classes of extracerebral artifacts into separate, mutually uncorrelated output channels:

    • Theta Oscillations: A dominant 7 Hz\approx 7\text{ Hz} theta rhythm widely smeared across raw frontal, central, and parietal scalp channels is isolated into a single ICA output channel (trace 1).
    • Alpha Activity: Rhythmic 10 Hz\approx 10\text{ Hz} alpha bursts, largely masked in raw EEG by dominant slow waves, are uncovered in an independent channel (trace 2), demonstrating distinct intermittent bursting and quiescent dynamics.
    • Ocular Artifacts: Low-frequency, near-DC voltage shifts corresponding to slow eye movements detected at frontal (Fpz) and EOG electrodes are isolated to ICA trace 10.
    • Line Noise: 60 Hz60\text{ Hz} alternating current line interference is isolated into ICA trace 13.
    • Electromyographic (EMG) Noise: High-frequency (50100 Hz50\text{--}100\text{ Hz}) broad-spectrum signals generated by scalp muscle activity are separated into ICA traces 9 and 14.
  5. Knowl 5 — Tracking Behavioral and Psychophysiological State Shifts via Residual ICA Channel Correlation

    empirical result

    Although Infomax ICA training minimizes mutual information and drives inter-channel output correlations to near zero (<0.05< 0.05) on data from a specific behavioral state, holding the unmixing weight matrix W\mathbf{W} fixed across varying states reveals systematic psychophysiological transitions:

    • When an unmixing matrix W\mathbf{W} trained on alert EEG data is used to filter recordings from drowsy periods (characterized by behavioral response lapses), the residual cross-correlations among the ICA output channels increase substantially, reverting back to baseline decorrelation once the subject returns to an alert state.
    • Conversely, an unmixing matrix trained on drowsy EEG data yields output channels that are decorrelated during drowsiness but show elevated cross-channel residual correlations when applied to alert EEG segments.

    These shifts in residual cross-correlation reflect state-dependent nonstationarities in the topographic configuration and dynamics of brain electrical generators between alert and drowsy brain states.

  6. Knowl 6 — Convergence Invariance of EEG-ICA to Network Initialization and Presentation Order

    empirical result

    The Infomax ICA source separation solution for continuous 14-channel scalp EEG is robust against variations in starting parameters and optimization paths:

    • Training two independent 14-channel ICA networks on the same 6.5-minute segment of EEG using different random initial weight matrices, different data sample presentation orders, and different learning rates converged to nearly identical spatial unmixing matrices W\mathbf{W}.
    • Filtering a separate, unseen segment of EEG data from the same recording session with both resulting matrices produced independent source activations where 11 out of the 14 best-matched output channel pairs exhibited Pearson correlation coefficients r>0.95r > 0.95, and all 14 pairs had correlations r0.894r \ge 0.894.
  7. Knowl 7 — Decomposition and Temporal Isolation of Overlapping Event-Related Potential Subcomponents

    empirical result

    Applying Infomax ICA to average event-related potentials (ERPs) recorded during an auditory vigilance task resolves temporally and spatially overlapping response subcomponents without requiring a priori dipole models:

    • Steady-State Evoked Responses (SSR): The continuous 39 Hz39\text{ Hz} auditory steady-state response elicited by concurrent click stimulation is isolated into two dedicated ICA channels (traces S1\text{S1} and S2\text{S2}), separating stimulus-induced amplitude perturbations from overlapping transient potentials.
    • Hit vs. Lapse Evoked Responses: Transient evoked potentials to target noise bursts are decomposed into condition-specific components:
      • Three channels (H1–H3\text{H1}\text{--}\text{H3}) isolate time-limited waveform subcomponents elicited exclusively by detected targets (Hits).
      • Four channels (L1–L4\text{L1}\text{--}\text{L4}) isolate distinct subcomponents characteristic of undetected targets (Lapses).

    This separation proves that ICA can isolate distinct, neurobiologically interpretable constituent subprocesses that overlap in time across scalp electrodes.

  8. Knowl 8 — Sustained Auditory Vigilance and Multichannel Electrophysiological Protocol

    experimental setup

    The empirical data used to validate EEG-ICA were collected under a continuous vigilance monitoring paradigm designed to induce fluctuations between alertness and drowsiness:

    • Subject Protocol: 10 adult volunteers completed three or more 30-minute sessions seated in a warm, dimly lit chamber with instructions to keep their eyes closed.
    • Stimuli:
      • Auditory Targets: 350 ms350\text{ ms} duration increases in continuous 62 dB62\text{ dB} white noise background (6 dB6\text{ dB} above detection threshold), presented at random intervals averaging 10 per minute.
      • Background Click Train: A continuous 39 Hz39\text{ Hz} click train presenting auditory steady-state stimulation.
      • Irrelevant Probes: Brief tones of 568 Hz568\text{ Hz} and 1098 Hz1098\text{ Hz} interspersed at pseudo-random 24 s2\text{--}4\text{ s} intervals.
    • Recording Setup: 13 scalp electrodes placed according to the International 10–20 system (Fz,Cz,Pz,Oz,F3,F4,C3,C4,T3,T4,P3,P4,Fpz\text{Fz}, \text{Cz}, \text{Pz}, \text{Oz}, \text{F3}, \text{F4}, \text{C3}, \text{C4}, \text{T3}, \text{T4}, \text{P3}, \text{P4}, \text{Fpz}) referenced to the right mastoid, plus one diagonal bipolar electrooculogram (EOG\text{EOG}) channel (14 channels total), sampled at 312.5 Hz312.5\text{ Hz}.
    • Behavioral Scoring: Targets detected via button press within 1003000 ms100\text{--}3000\text{ ms} were classified as Hits (00); unacknowledged targets were classified as Lapses (11). Continuous local error rate was computed by convolving the binary performance series with a 95 s95\text{ s} moving smoothing window stepped at 1.64 s1.64\text{ s} intervals.

Coverage note — No substantial contributed material from the paper was omitted. All key models, statistical assumptions, empirical findings (artifact separation, rhythm isolation, ERP decomposition, stability across initializations, and state transition tracking), and the experimental protocol are fully captured.

References

  1. 1.A.J. Bell & T.J. Sejnowski (1995). An information-maximization approach to blind separation and blind deconvolution, Neural Computation 7:1129-1159.
  2. 2.A.J. Bell & T.J. Sejnowski (1995). Fast blind separation based on information theory, in Proc. Intern. Symp. on Nonlinear Theory and Applications (NOLTA), Las Vegas, Dec. 1995.
  3. 3.P. Comon (1994) Independent component analysis, a new concept? Signal processing 36:287-314.
  4. 4.A.M. Dale & M.I. Sereno (1993) EEG and MEG source localization: a linear approach. J. Cogn. Neurosci. 5:162.
  5. 5.R. Galambos & S. Makeig. (1989) Dynamic changes in steady-state potentials. In Erol Basar (ed.), Dynamics of Sensory and Cognitive Processing of the Brain, 102-122. Berlin:Springer-Verlag.
  6. 6.S. Makeig & R. Galambos. (1989) The CERP: Event-related perturbations in steady-state responses. In E. Basar & T.H. Bullock (ed.), Brain Dynamics: Progress and Perspectives, 375-400. Berlin:Springer-Verlag.
  7. 7.T-P. Jung, S. Makeig, M. Stensmo, & T. Sejnowski. Estimating alertness from the EEG power spectrum. Submitted for publication.
  8. 8.S. Makeig & M. Inlow (1993) Lapses in alertness: Coherence of fluctuations in performance and EEG spectrum. Electroencephalog. clin. Neurophysiolog. 86:23-35.

Citation

MLA
Makeig, S., et al. “Independent Component Analysis of Electroencephalographic Data”. Advances in Neural Information Processing Systems, vol. 8, 1995, https://proceedings.neurips.cc/paper_files/paper/1995/file/754dda4b1ba34c6fa89716b85d68532b-Paper.pdf.
APA
Makeig, S., Bell, A., Jung, T.-P., & Sejnowski, T. (1995). Independent Component Analysis of Electroencephalographic Data. Advances in Neural Information Processing Systems, 8. https://proceedings.neurips.cc/paper_files/paper/1995/file/754dda4b1ba34c6fa89716b85d68532b-Paper.pdf
Chicago
Makeig, S., A. Bell, T.-P. Jung, and T. Sejnowski. 1995. “Independent Component Analysis of Electroencephalographic Data”. Advances in Neural Information Processing Systems 8. https://proceedings.neurips.cc/paper_files/paper/1995/file/754dda4b1ba34c6fa89716b85d68532b-Paper.pdf.
Harvard
Makeig, S. et al. (1995) “Independent Component Analysis of Electroencephalographic Data”, Advances in Neural Information Processing Systems. Curran Associates, Inc. Available at: https://proceedings.neurips.cc/paper_files/paper/1995/file/754dda4b1ba34c6fa89716b85d68532b-Paper.pdf.
Vancouver
1. Makeig S, Bell A, Jung T-P, Sejnowski T (1995) Independent Component Analysis of Electroencephalographic Data. Advances in Neural Information Processing Systems 8:

BibTeX

@inproceedings{makeig1995independent,
  title = {Independent Component Analysis of Electroencephalographic Data},
  author = {Makeig, Scott and Bell, Anthony and Jung, Tzyy-Ping and Sejnowski, Terrence},
  year = {1995},
  booktitle = {Advances in Neural Information Processing Systems},
  publisher = {Curran Associates, Inc.},
  volume = {8},
  url = {https://proceedings.neurips.cc/paper_files/paper/1995/file/754dda4b1ba34c6fa89716b85d68532b-Paper.pdf}
}
Metadata:DOI registry

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: Authors