Utility of Independent Component Analysis for Interpretation of Intracranial EEG

Abstract

Electrode arrays are sometimes implanted in the brains of patients with intractable epilepsy to better localize seizure foci before epilepsy surgery. Analysis of intracranial EEG (iEEG) recordings is typically performed in the electrode channel domain without explicit separation of the sources that generate the signals. However, intracranial EEG signals, like scalp EEG signals, could be linear mixtures of local activity and volume-conducted activity arising in multiple source areas. Independent component analysis (ICA) has recently been applied to scalp EEG data, and shown to separate the signal mixtures into independently generated brain and non-brain source signals. Here, we applied ICA to unmix source signals from intracranial EEG recordings from four epilepsy patients during a visually cued finger movement task in the presence of background pathological brain activity. This ICA decomposition demonstrated that the iEEG recordings were not maximally independent, but rather are linear mixtures of activity from multiple sources. Many of the independent component (IC) projections to the iEEG recording grid were consistent with sources from single brain regions, including components exhibiting classic movement-related dynamics. Notably, the largest IC projection to each channel accounted for no more than 20–80% of the channel signal variance, implying that in general intracranial recordings cannot be accurately interpreted as recordings of independent brain sources. These results suggest that ICA can be used to identify and monitor major field sources of local and distributed functional networks generating iEEG data. ICA decomposition methods are useful for improving the fidelity of source signals of interest, likely including distinguishing the sources of pathological brain activity.

Four-dimensional maps of the human somatosensory system

“…Nineteen clustering procedures were performed imposing increasing numbers of clusters (3)(4)(5)(6)(7)(8)(9)(10)(11)(12)(13)(14)(15)(16)(17)(18)(19)(20) and computing silhouette values (75) to evaluate clustering validity. The optimal clustering was defined by the maximal average silhouette value, and leads with negative silhouette values were not considered in the evaluation of the optimal clustering using a repeated-measurement ANOVA with time and cluster as factors.…”

“…Depending on the electrodes used, two kinds of recordings can be made: (i) intraparenchymal recordings, also called stereo-EEG (sEEG) (6), obtained using stereotactically inserted needle-like electrodes with multiple recording leads; and (ii) the electrocorticogram (ECoG), obtained from subdural electrode grids, covering regions of cortex. The latter technique suffers from three disadvantages: (i) the technique only samples from cortical gyri, missing cortical regions buried in sulci (7); (ii) the technique is affected by volume conduction (8); and (iii) the technique does not record directly from gray matter, because pia and arachnoid mater lie in between the electrodes and the cortex, reducing the amplitude and making it more difficult to extract highfrequency activity from the recorded signal. Both approaches suffer from the so-called sparse-sampling problem (i.e., the limited and uneven coverage of the patients' brain because the positioning of the electrodes in any given patient is dictated by clinical criteria), leading to difficulties in carrying out analyses at the population level.…”

Pathological and physiological high-frequency oscillations in focal human epilepsy

“…Patients watched a bedside computer for commands to move a digit. Details of the task have been described in detail previously (Whitmer et al 2010). Briefly, the computer screen displayed an image of a digit or the name of a digit.…”

Recording and analysis techniques for high-frequency oscillations

“…Clinical macro-electrodes placed within or on the surface of the brain record a spatial average of locally generated LFP and volume conducted activity (Kajikawa and Schroeder, 2011; Whitmer et al, 2010). Because of their relatively large surface area macro-electrodes, typically ~1 to 10 mm 2 , do not record single neuron action potentials or multi-unit activity.…”