Measuring the time course of brain activity

Once the coding and processing stages have been identified, one of the key questions in cognitive psychology is how they are organized in time. Can a cognitive task be broken down into elementary stages? How long does each step take? How does this time vary according to the parameters of the task? Can the order in which steps are performed vary? Are they executed in parallel or in series? Is the transmission of information a continuous or discrete process? Why does processing time vary from trial to trial, or between individuals?

High temporal resolution brain imaging could provide answers to these questions. The lecture therefore examined the extent to which fMRI, but also other methods such as electro- and magneto-encephalography, are likely to break down brain activations over time.

A common misconception is that fMRI does not have the temporal resolution required to track cognitive processes. This is not true. Insofar as the delay imposed by hemodynamic function is relatively stable, it is possible to detect shifts of the order of a few hundred milliseconds in the onset of activity in different brain regions, or in the same region under different experimental conditions. In particular, Ravi Menon and his colleagues have shown that it is possible to distinguish between regions whose duration of activation varies with reaction time, and others whose onset of activity indexes the timing of motor programming (Menon & Kim, 1999; Menon, Luknowsky, & Gati, 1998).

In the laboratory, we have exploited this temporal resolution through spectral analysis. In a paradigm where a stimulus is presented periodically at long intervals (e.g. a sentence every 15 seconds), the activity evoked in the regions involved is also periodic. Extracting its phase and amplitude provides valuable insights into the latency and organization of language areas. We have been able to show a hierarchical organization of temporal and inferior frontal regions, with response to sentences becoming slower the further one moves away from the primary auditory area, on the one hand towards the posterior part of the left temporal lobe, and on the other towards the bilateral temporal pole and left Broca's area (Dehaene-Lambertz et al., Human Brain Mapping, 2006). This temporal hierarchy, which suggests embedded processes of temporal integration, is observed as early as 2-3 months of age in infants (G. Dehaene-Lambertz et al., PNAS, 2006).

The phase and amplitude of fMRI activations can be exploited in even finer detail. Suppose the experiment manipulates an F factor at several levels. If F affects the moment of entry into activity of a brain region, but not its duration, we can show mathematically that phase should vary linearly with F (with a slope measuring the slowdown caused by F), while signal amplitude remains constant. Conversely, if F affects duration without altering the moment of entry into activity, then phase must vary with F (with a slope equal to half the size of F's effect on the duration of the measured process), and amplitude must also vary linearly with F. Other combinations of effects are possible, and each is associated with a "signature" of phase and frequency variations with F. Using this method, Mariano Sigman, a post-doctoral fellow at the laboratory, showed that it was possible to break down a complex cognitive task, lasting just under two seconds, into five distinct temporal stages (Sigman et al, NeuroImage, 2006).

Despite these advances, fMRI remains limited in its temporal accuracy by the blurring induced by the slow hemodynamic response. More direct access to the temporal course of cognitive operations is provided by electro- and magnetoencephalography (EEG and MEG). These methods measure, on a millisecond scale, the minute electrical currents and magnetic fields generated outside the cranium by the synchronous activity of neuron populations. It was beyond the scope of this lecture to review all the temporal decomposition results obtained by EEG and MEG. We were simply interested in whether the gain in temporal resolution was necessarily accompanied by a loss of spatial resolution. In the case of EEG, while it is possible to follow a series of processing steps over time, the scattering and superposition of their topographies on the scalp surface makes their spatio-temporal reconstruction difficult and ambiguous (although significant progress has been made in the reconstruction of cortical sources). MEG, on the other hand, is much more spatially selective. MEG machines have recently undergone significant advances. Some have over 300 sensors, whose spatial selectivity can be enhanced by a "planar gradiometer" configuration. We have reviewed recent work by Rita Hari's group, which has succeeded in breaking down five successive stages of facial action imitation with great temporal and spatial precision (Nishitani & Hari, 2002).

Another technological advance is that EEG recording can now be combined with fMRI brain activation. Simultaneous EEG+fMRI recording requires suitable amplifiers, special electrodes compatible with the imager's magnetic field, and above all sophisticated signal processing to eliminate artifacts arising from MRI gradients, the cardio-ballistogram, and movements induced in particular by the subject's breathing. Electrode placement also induces a slight loss of fMRI signal. This method should therefore be reserved for very specific cases where the signals of interest cannot be measured in separate experiments (studies of epilepsy, spontaneous endogenous brain activity, etc.).

EEG+fMRI recording also opens up the possibility of exploiting spontaneous EEG fluctuations, from trial to trial, to propose a new strategy for spatio-temporal localization of brain activation. The idea is to isolate a component of the evoked potential that occupies a well-defined temporal window; to measure its trial-to-trial fluctuations; and to use these fluctuations, after convolution with the canonical hemodynamic function, to predict fMRI signals, in order to determine which brain regions are involved. The method is based on the strong relationship between local potential fluctuations and the BOLD signal measured on fMRI (see above). It should be noted that this is the only EEG source localization method that does not require realistic modeling of tissue conductance. In principle, this method can provide a temporal label for fMRI activations, whatever their intrinsic spatial resolution. We have looked at a very recent example of the exploitation of this technique, which identified in time and space a major step in the error detection and correction system (Debener et al., 2005). It remains to be shown, in future work, that several temporal labels, each associated with a different window of the evoked potential, can be affixed to cerebral activations in this way in order to break down their course more finely.