Ph. D. Project
Multiscale brain functioning analysis: forward and inverse problems in micro and macro-electrodes
2015/10/01 - 2018/09/30
For drug-resistant epilepsy patients, brain electrical activities are explored at different spatial scales, from scalp EEG (electrodes on the surface of the head for a global view), SEEG (electrodes implanted in the brain for recording the implanted structures and the neighboring ones) or microelectrodes, which record the activity of individual neurons and their neighbors. In some cases, patients are explored by simultaneous recordings at multiple spatial scales (microelectrodes - SEEG - EEG).

Open issues
According to [1] the microelectrode signal may be approximately separated into a low-frequency component (≤100 Hz, known as Local Field Potential - LFP) due to several more or less distant neurons, and a high frequency component, due to the neighboring neuron (Single Unit Activity - SUA) or neurons (Multiple Unit Activity - MUA). At the same time, these neurons are often parallel, which leads to the superposition of their activities and generates equivalent dipoles that can be recorded by macroelectrodes (SEEG or even surface EEG).
The links between the signals measured at different spatial scales are far from being fully understood, especially the relations between microelectrode recordings and SEEG / EEG signals. This is partly due to the same causes which prevent the establishment of a strict relationship between (depth) SEEG and (surface) EEG, namely the fact that the spatial sampling becomes increasingly sparse and irregular as the scale decreases: the SEEG electrodes are not implanted in the whole brain volume and microelectrodes explore only neural activities in an even smaller volume.
A second difficulty comes from the intrinsically different electrophysiological dynamics of the explored structures at these different scales - the frequencies of the recorded signals are very different between microelectrodes, SEEG and EEG, for several reasons: geometric (spatial structures of groups of neurons), temporal (phase differences and activation times) and biochemical [2].
Finally, feedback effects also appear to be involved: if it is clear that the activity at the microscopic scale is a cause of macroscopic activity (or, in other words, it is the individual neurons and networks that determine the behavior of a brain structure recorded by SEEG or EEG), recent research suggests that the low-frequency remote activity (thus recorded by the macro-electrodes) influences the activation of neurons explored by microelectrodes [3].

From a fundamental point of view, the LFP signals can be seen as the missing link between the individual activities of neurons and macroscopic SEEG / EEG signals. However, the relationship between the signals from the microscopic spatial scale (MUA SUA) and LFP is not yet clearly established [4], [5].
One possible model for a micro-electrode signal x is:
x = SUA / MUA + LFP_local + LFP_propagated
It is clear that these components are both linked together (in terms of cause and effect) and mixed and, before studying the links, they must be separated. This separation is of several types - between the LFP and SUA / MUA spikes (low and high frequency, to study the inter-scale link) between the local and propagated LFPs (mixtures at the mesoscopic scale, to separate the local activity related to close neurons close to the propagated one) and among the spikes (to see which neurons are active at a given time and thus contribute to local LFP).
A first direction to explore is the separation of the two types of activity picked up by the microelectrodes (fast and slow). A possible solution may come from the Bayesian approach proposed in [6], which however uses priors without direct physiological basis (without SEEG validation).
The separation between the local and propagated LFPs is essentially spatial. One approach to estimate the local activity at the site explored by microelectrodes can be inspired by the SEEG localization techniques developed in our team [7], but it can also stem from more generic methods (under development in the team) based on common/specific subspace decompositions.
Finally, the techniques for detecting and separating activities from different neurons (spike detection and spike-sorting) are quite abundant in the recent literature. In general, they are based on feature extraction methods (mainly morphological features of the spikes), followed by supervised or unsupervised classification algorithms [8]. An important class of methods is based on wavelets, which are also used during the preprocessing steps [8].
neurol activations, single unit, local field potential, microelectrodes, SEEG, source separation
Biology, Signals and Systems in Cancer and Neuroscience