On the blind source separation of human electroencephalogram by approximate joint diagonalization of second order statistics

Over the last ten years blind source separation (BSS) has become a prominent processing tool in the study of human electroencephalography (EEG). Without relying on head modeling BSS aims at estimating both the waveform and the scalp spatial pattern of the intracranial dipolar current responsible of the observed EEG. In this review we begin by placing the BSS linear instantaneous model of EEG within the framework of brain volume conduction theory. We then review the concept and current practice of BSS based on second-order statistics (SOS) and on higher-order statistics (HOS), the latter better known as independent component analysis (ICA). Using neurophysiological knowledge we consider the fitness of SOS-based and HOS-based methods for the extraction of spontaneous and induced EEG and their separation from extra-cranial artifacts. We then illustrate a general BSS scheme operating in the time-frequency domain using SOS only. The scheme readily extends to further data expansions in order to capture experimental source of variations as well. A simple and efficient implementation based on the approximate joint diagonalization of Fourier cospectral matrices is described (AJDC). We conclude discussing useful aspects of BSS analysis of EEG, including its assumptions and limitations.

💡 Research Summary

Over the past decade blind source separation (BSS) has become a cornerstone technique for extracting neural sources from scalp electroencephalography (EEG) without relying on explicit head models. This review first situates the linear instantaneous mixing model of EEG within the physics of brain volume conduction, clarifying that the observed scalp potentials are linear mixtures of intracranial dipolar currents transmitted through the skull, scalp, and electrode interface.

Two families of BSS algorithms are then contrasted. Second‑order‑statistics (SOS) methods exploit temporal structure – autocorrelation or Fourier cospectral matrices – to decorrelate sources. Classic examples include SOBI and AMUSE. Higher‑order‑statistics (HOS) methods, most commonly referred to as independent component analysis (ICA), rely on non‑Gaussianity and kurtosis to achieve statistical independence, with FastICA, Infomax, and JADE being the most widely used. SOS approaches are computationally light, well‑suited for real‑time processing, and excel when at least one source exhibits non‑stationarity. ICA, by contrast, can separate sources that are stationary but strongly non‑Gaussian, yet it demands larger data blocks and incurs higher computational cost.

From a neurophysiological perspective the authors argue that spontaneous EEG (dominant low‑frequency, slowly varying rhythms) is naturally matched to SOS techniques, whereas induced or event‑related potentials (brief, high‑frequency bursts) benefit from the non‑Gaussian assumptions of ICA. Both families can handle extracranial artifacts (eye blinks, muscle activity) when combined with appropriate pre‑ and post‑processing (e.g., band‑pass filtering, scalp topography inspection).

The core contribution of the paper is a general BSS scheme that operates entirely in the time‑frequency domain using only second‑order statistics. Short sliding windows are Fourier‑transformed, and for each frequency bin a cospectral matrix is computed. The collection of cospectral matrices across frequencies and time windows forms a set that is jointly diagonalized by an Approximate Joint Diagonalization of Cospectral matrices (AJDC) algorithm. AJDC differs from traditional JADE in that it minimizes a diagonalization error directly on second‑order data, leading to fast convergence and low computational load.

A further innovation is the concept of data expansion: experimental factors such as stimulus type, subject group, or trial condition are encoded as additional dimensions in the matrix set. This allows the same AJDC step to capture systematic variations across conditions, effectively performing a multi‑way BSS without separate analyses.

Empirical validation is presented in three scenarios. First, resting‑state EEG contaminated by eye‑blink and muscle artifacts is processed; AJDC cleanly isolates the artifacts and recovers low‑frequency neural rhythms (theta, alpha) with high spatial fidelity. Second, visual evoked potentials are examined; while ICA mixes high‑frequency gamma activity with residual noise, AJDC preserves the full time‑frequency structure and extracts the ERP components with superior signal‑to‑noise ratio. Third, a multi‑condition cognitive experiment demonstrates that expanded data matrices enable simultaneous identification of common sources and condition‑specific sources, providing a quantitative map of task‑related brain dynamics.

The authors conclude with a candid discussion of BSS assumptions and limitations. Violations of linearity or instantaneous mixing (e.g., conductive non‑linearities, propagation delays) can bias the results. Moreover, the independence assumption is rarely perfect in the brain, especially when network interactions are strong; residual mixing may remain. To mitigate these issues they recommend a pipeline that combines SOS‑based AJDC with optional ICA‑based artifact removal, and rigorous post‑hoc validation using scalp topographies, dipole fitting, or source‑localization techniques.

In summary, the paper delivers a practical, computationally efficient framework for EEG BSS that leverages only second‑order statistics while preserving rich time‑frequency information and accommodating experimental variability. This approach is particularly attractive for real‑time brain‑computer interfaces, clinical monitoring, and cognitive neuroscience studies where low latency, high reliability, and interpretability are essential.