The information content of physiological and epileptic brain activity


  • Andrew J. Trevelyan,

    1. Institute of Neuroscience, Newcastle University, Newcastle upon Tyne NE2 4HH, UK
    Search for more papers by this author
  • Willy Bruns,

    1. Howard Hughes Medical Institute and Center for Neural Circuits and Behavior and Neurobiology Section, Division of Biology, University of California San Diego, La Jolla, CA 92093-0634, USA
    Search for more papers by this author
  • Edward O. Mann,

    1. Department of Physiology, Anatomy and Genetics, Oxford University, Oxford OX1 3PT, UK
    Search for more papers by this author
  • Valerie Crepel,

    1. Inserm Unite 901, Université de la Mediterranee, UMR S901 Aix-Marseille, and Institut de Neurobiologie de la Mediterranee, Marseille 13009, France
    Search for more papers by this author
  • Massimo Scanziani

    1. Howard Hughes Medical Institute and Center for Neural Circuits and Behavior and Neurobiology Section, Division of Biology, University of California San Diego, La Jolla, CA 92093-0634, USA
    Search for more papers by this author

  • The report was presented at the symposium Why do some brains seize? Molecular, cellular and network mechanisms, which took place at the Epilepsy Research UK Expert International Workshop, Oxford, UK on 15–16 March 2012.

A. J. Trevelyan: Institute of Neuroscience, Newcastle University, Newcastle upon Tyne NE2 4HH, UK. Email:


Abstract  Cerebral cortex is a highly sophisticated computing machine, feeding on information provided by the senses, which is integrated with other, internally generated patterns of neural activity, to trigger behavioural outputs. Bit by bit, we are coming to understand how this may occur, but still, the nature of the ‘cortical code’ remains one of the greatest challenges in science. As with other great scientific challenges of the past, fresh insights have come from a coalescence of different experimental and theoretical approaches. These theoretical considerations are typically reserved for cortical function rather than cortical pathology. This approach, though, may also shed light on cortical dysfunction. The particular focus of this review is epilepsy; we will argue that the information capacity of different brain states provides a means of understanding, and even assessing, the impact and locality of the epileptic pathology. Epileptic discharges, on account of their all-consuming and stereotyped nature, represent instances where the information capacity of the network is massively compromised. These intense discharges also prevent normal processing in surrounding territories, but in a different way, through enhanced inhibition in these territories. Information processing is further compromised during the period of post-ictal suppression, during interictal bursts, and also at other times, through more subtle changes in synaptic function. We also comment on information processing in other more physiological brain states.

[ Andrew Trevelyan (Newcastle University) has recently taken up a Senior Lectureship position, having previously held an Epilepsy Research UK Fellowship. He did his medical training in Oxford and Edinburgh, before doing postdoctoral positions with Julian Jack (Oxford University) and Rafa Yuste (Columbia University). His research addresses how neuronal activity flows through cortical networks. His doctoral thesis concerned the nature of cortical topographic maps, in which information is represented as focal activity. This remains a key interest, because his research now focuses on the different trajectories of epileptic waves, flowing across these topographic maps.]


Seizure is a clinical term, referring to a period of altered consciousness arising from paroxysmal neuronal discharges; epilepsy is the tendency to suffer such events. These events are readily recorded by EEG (both scalp and subdural electrodes), yet despite the cataclysmic nature of these events, there remains much debate about what exactly constitutes a seizure in terms of neuronal firing. This review explores a novel approach using information theory to define what exactly the pathological activity is, and to distinguish it from what may be regarded as a physiological response to the pathological focal discharge, in the surrounding cortex.

Levels of neuronal participation and information content

New molecular biological techniques have allowed a powerful examination of how information can be encoded in the cortex. Intense neuronal firing induces expression of various genes, including a class referred to collectively as immediate early genes. Two research teams have recently utilized this technique, first to define, and then to reactivate a distributed population of hippocampal neurons which encode a particular memory (Garner et al. 2012; Liu et al. 2012). One team coupled expression of the early gene, c-fos, to channel-rhodopsin (Liu et al. 2012); the other took a very similar approach, but coupled c-fos instead to DREADD (designer receptor exclusively activated by designer drug), an exogenous ligand-gated, depolarizing, membrane channel which can be activated by an artificial ligand (Garner et al. 2012). Mice were ‘trained’ with an auditory-cued foot shock in a particular context (size and colour of cage, and the presence or absence of scents), creating a context specific memory which could be tested by their freezing response to repeats of the auditory cue. This well characterized behavioural assay only elicits freezing reliably when the animals are in the same context, but if the scent is changed, or the shape or colour of the cage, then the likelihood of freezing behaviour is greatly reduced. Remarkably, this ‘context’ could be reproduced artificially by light activation of the channel-rhodopsin driven by c-fos (Liu et al. 2012). It was also possible to train in a new context, and then ‘distort’ that context by activation of a different population of neurons marked from a prior training (Garner et al. 2012). The second training was found to elicit freezing behaviour, but freezing was not induced by the distorted context (same environment, but with extra activation of another subset of hippocampal neurons).

These experiments also provided our first mappings of specific memories in a mammalian brain. Because c-fos expression was also coupled to green fluorescent protein (GFP), these mappings could be visualized; what was seen was a sparse distribution of labelled neurons, with no obvious anatomical relationship (there may be a non-obvious pattern of course). If we consider the possible permutations, given the numbers of neurons in mouse dentate gyrus, we see the true information potential of this encoding (and that is just the spatial patterns, not taking into account temporal information within the firing patterns). Previous anatomical studies suggest there to be 0.5–2 million granule cells in the nucleus (Amrein et al. 2004), of which 5% are activated for any one memory (Liu et al. 2012). The number of possible permutations is given by the binomial theorem as inline image. The general formula is referred to as ‘n choose k’, which is the number of different ways of choosing k items (in this case, the number of cells involved in the memory) from a total population of size n (total number of dentate granule cells).

display math

If each permutation provides the possibility for a specific memory, one can see the power of this system; these are astronomically large numbers. The peak number of permutations would occur with activation of 50% of neurons, but there are other considerations to bear on the optimal number of cells active. For instance, Levy & Baxter (1996) considered the trade-off between energy and information content (Barlow, 1959; Niven et al. 2007), and estimated that optimal encoding may also be about 5%, very similar to that recorded with these new experimental techniques (Liu et al. 2012). It is noteworthy that the memory of the home cage appears to be represented more economically, with fewer neurons active, whereas brand new (and thus by definition ‘rare’) events, with an important message (related to pain), are encoded at a higher energy cost.

Interestingly, to examine the efficiency of c-fos expression, Liu et al. (2012) also studied the activation profile following a period of status epilepticus. As a control, this worked beautifully, because it induced expression in 99.8% of neurons, and so provided confidence that c-fos labelling does reflect hippocampal activation (although status may achieve far more intense firing than for the more subtle memory protocol) (Matsumoto & Marsan, 1964). Compared to the memory protocol, though, the information potential of the status epilepticus activation pattern is catastrophically reduced; the binomial theorem tells us that for 100% activation, the number of permutations is just 1!

So far, we have talked about information potential in terms of populations of neurons, but information is also encoded in the pattern of firing. But in this respect too, seizures represent a massive reduction in information potential, because they are such stereotyped events (both clinically and electrographically). The stereotypy of seizure activity suggests that this may also be a kind of attractor state, and the association of some kinds of partial seizures with particular memories (‘auras’) may even indicate its trajectory through the cortex. The tight association of these auras with seizures suggest that the attractor state has already been entered, but the full seizure activity may be averted. Of course, these ideas were being considered by Victorian neurologists; are we merely dressing these old ideas in the ‘new cloth’ of modern terminology? Perhaps, although this may yet provide helpful insights: for instance, it cautions us against trying to work back from the attractor state to identify how the event started, because attractors can be entered from many sites, and if the system is chaotic, then information about that path is rapidly lost once the attractor is entered (Lorenz, 1995).

Notably, the compromised informational state persists beyond the seizure itself. Immediately post-seizure, there is an almost complete suppression of all neuronal activity, leading to a state that is as badly compromised for information processing as the preceding full ictal activity (n choose zero =n choose n= 1). The network recovers from this post-ictal suppression relatively quickly, but there is also evidence from animal studies for a more subtle, persistent compromised state. This evidence comes from brain slice preparations from chronically epileptic rodents (following an initial pharmacological kindling), in which there is raised expression of kainate receptors, with slower opening kinetics than AMPA receptors, with a consequent extension of the integration time of individual neurons (Artinian et al. 2011). This in itself may increase the likelihood of a seizure, but it also changes the output of these neurons in an interesting way. Artinian et al. (2011) compared the firing output arising from temporal summation of repeated, normalized, subthreshold inputs in control versus epileptic animals (Fig. 1). The summation of such inputs lies very close to action potential threshold in control animals, and so creates a firing output that is influenced heavily by noise in the system, with great variation in the interspike intervals. In contrast, this pattern of stimulation in granule cells in epileptic brain slices summates far above action potential threshold, and so spiking is reliably driven in an almost metronomic fashion; the output is thus very predictable, and represents a huge reduction in temporal information compared to the control activity patterns. Given this evidence for changes in information processing in the dentate gyrus following seizures, it would be interesting to chart the level of activation in the footshock–c-fos expression paradigm in these animals.

Figure 1.

Aberrant kainate receptor-operated synapses change the firing tempo in dentate granule cells in temporal lobe epilepsy (adapted from Artinian et al. 2011) 
A and B, subthreshold EPSPAMPA in naïve conditions (A) and subthreshold EPSPKA in epileptic conditions (B); EPSPs are evoked by electrical stimulations performed in the inner one-third of the molecular layer of the dentate gyrus. Note the long lasting kinetics of kainate excitatory postsynaptic potentials (EPSPKA) compared with AMPA mediated EPSPs (EPSPAMPA). C and D, superimposed Nissl staining (green) of dentate granule cell layer and post hoc reconstruction (red) of control (C) and epileptic (D) granules cells filled with biocytin. Note the recurrent mossy fibre (rMF, arrow; oml, outer molecular layer; iml, inner molecular layer; g, granule cell layer; h, hilus). E and F, EPSPKA but not EPSPAMPA trigger a sustained and rhythmic firing pattern in epileptic dentate granule cells; spike discharge for control EPSPAMPA (E), and epileptic EPSPKA (F) evoked at 30 Hz in dentate granule cells. G and H, schematic diagrams, showing how a more powerful excitatory drive can reduce the jitter in spike timing. The noisy trace represents a depolarization from rest, and the boundaries represent the range of the noisy signal. An increase in temporal summation in the epileptic animals means that the summed events drive the cell strongly through the threshold for firing. In contrast, in control animals, summation is weaker and therefore the timing of spikes is affected more by noise. The reduction in spiking jitter in epileptic animals makes spike timing more predictable, and thus represents a drop in information potential.

Role of GABAergic interneurons in cortical information processing

We may make similar argument about the relative information latent in interneuronal and pyramidal cell firing patterns during physiological brain rhythms. For instance fast-spiking parvalbumin-expressing (PV) interneurons fire with high probability during cycles of gamma rhythms (Fig. 2A and B), and so their firing is in this sense highly predictable and thus low in information. In contrast, pyramidal cells fire only intermittently, and so their firing patterns are less predictable, which enhances their capacity for information encoding (Wolfe et al. 2010). Inhibitory GABAergic interneurons, though, do play a key role in enabling such sparse coding strategies, by simultaneously restraining spike rates and coordinating spike times across neuronal networks (Cobb et al. 1995). Interneurons can also inherit the stimulus selectivity of presynaptic excitatory neurons (Maurer et al. 2006), but they are generally seen as ‘low-information’ bearers, which rather shape the flow of cortical activity.

Figure 2.

Tuning fast cortical network oscillations (unpublished data of E. O. Mann) 
A, schematic diagram showing reciprocal connections between fast-spiking interneurons (black) and pyramidal cells (grey). B, simultaneous recordings of local field potential (LFP; top), unit activity of perisomatic-targeting fast-spiking interneuron (middle), and membrane potential of a reciprocally connected layer 3 pyramidal neuron (bottom), during UP state-driven fast oscillations in the superficial layers of the medial entorhinal cortex in vitro. C, cycle-to-cycle variations in phasic inhibitory and excitatory postsynaptic currents (IPSC and EPSC, respectively, see also Atallah & Scanziani, 2009), simulated from parameter distributions recorded from a CA3 pyramidal neuron during network oscillations induced by 100 nm kainate and 3 μm NMDA, shown in DF. D, illustration of the measurement of cycle amplitude and period in the LFP (black; top), with shaded areas showing the corresponding measurement of phasic charge of synaptic events. E and F, histograms showing the relationship between LFP cycle amplitude and period (E), and phasic charge and cycle period (F). The 0.5% contour for the phasic charge of excitatory events (red) is included in the histogram for inhibitory events (blue) for comparison. Spearman's rank correlation coefficients are inset.

The regulation of spike timing by interneurons may itself play an additional role in cortical information processing, with one possibility being that each gamma cycle represents a discrete temporal window for synaptic integration and spike timing-dependent plasticity (Buzsaki & Wang, 2012). In this case, the circuit function of cortical gamma-frequency oscillations would be exquisitely sensitive to the precise cycle period, which displays large variability, and is intimately linked to the level of activity in the network (Atallah & Scanziani, 2009; Ray & Maunsell, 2010; Ahmed & Mehta, 2012).

Deficits in PV interneuron function, and abnormal gamma-frequency synchronization, have been increasingly highlighted as a possible pathophysiological mechanism in various diseases of cortical function, including epilepsy (Li et al. 2012; Tan et al. 2012), schizophrenia (Cunningham et al. 2006) and Alzheimer's disease. Verret et al. (2012) recently found reduced expression of Nav1.1 in cortical PV interneurons in a mouse model of Alzheimer's disease, which correlated with reduced gamma-frequency oscillations and the appearance of spontaneous epileptiform discharges. There are clearly multiple cellular and synaptic deficits within the brains of these mice, but restoring Nav1.1 levels by Nav1.1-BAC expression not only normalized network synchronization and reduced epileptiform discharges, but also helped alleviate the core memory deficits. This is consistent with sparse, heterogeneous and weakly synchronous activity being important for information encoding in cortical circuits.

Optogenetic approaches provide an alternative means of examining pathology arising from dysfunction in subclasses of cortical neurons, and likewise also indicate that deficits in PV interneuron performance may be associated with epileptiform discharges (Fig. 3). Single unit responses to visual stimuli were recorded in cortical layers 2/3 in visual cortex, in awake mice, whilst optogenetically modulating PV interneuron firing (Atallah et al. 2012). Control visual responses were alternated with trials during which PV firing was suppressed by light-emitting diode activation of archaerhodopsin (expressed specifically in these cells). There was a marked and consistent difference between the normal, ‘control’, visual responses, and those elicited when PV firing was suppressed, which commonly resulted in intense, rhythmic discharges, highly characteristic of seizure activity.

Figure 3.

Epileptic discharges induced by physiological stimuli when PV interneurons are suppressed (unpublished data of W. Bruns and M. Scanziani) 
A and B, example extracellular recordings (A) and raster plots from multiple trials (B) recorded in layer 2/3 of mouse visual cortex following presentation of a drifting grating to the contralateral visual field. During alternate trials, PV interneuron firing was suppressed optogenetically. The optogenetic control was provided by expressing archaerhodopsin specifically in PV cells (V-CRE mice infected with AAV-2/9-CBA-Flexed-Arch-GFP (UPenn Vector Core)). Two patterns of activity were recorded: the first was an essentially normal physiological response, with units showing orientation tuning (Atallah et al. 2012); the second was a very large amplitude, ictal-like, rhythmic bursting. C, these two responses are readily distinguished by a semi-automated clustering routine based on the spectral power between 5–50 Hz (grey dots were ambiguous trials). D, the majority of trials show ictal-like activity when PV interneurons are suppressed; 74% of ictal trials (26/35 trials) were during PV suppression, while only 44% of all non-ictal trials (53/121 tials) occurred during PV suppression.

The clinical relevance of ‘seizure-related information deficits’

Information considerations provide a useful framework for considering two very different patterns of activity that may be recorded at different locations during naturally occurring human epileptic seizures. There appears to be a core territory, in which neurons fire intensely and time-locked to the dominant, large amplitude, low frequency rhythms evident in the EEG (Schevon et al. 2012). The apparent intensity of this firing suggests that this is equivalent to the 100% participation visualized by c-fos activation (Liu et al. 2012). Beyond this core territory, though, there are regions which also show an increase in activity, but only to a fractional level of that seen during full ictal recruitment. It seems likely that the information potential in these ‘penumbral territories’ is also compromised but to a far lesser extent than in the fully recruited territories, and for very different reasons, due to inhibition rather than excess participation. Thus, the penumbral surround inhibition, which appears to be a response to the focal discharges, will presumably also block any ongoing physiological inputs into this territory too. Furthermore, the penumbral activity represents a massively reduced energy cost relative to full ictal recruitment, and with the added benefit of preventing further escalation of the activity.

These two patterns of activity are seen clearly also in simple, in vitro models of ictal propagation induced by removing Mg2+ ions from the bathing medium (Trevelyan et al. 2006, 2007; Trevelyan, 2008). Importantly, inhibition is preserved in this model of epilepsy. Studies of this, more tractable, in vitro preparation suggest that an enormous excitatory drive projects forward from the ictal wavefront (Trevelyan et al. 2006), but that it is opposed by a powerful feedforward inhibition. Notably, one study of human multielectrode array recordings only showed this lower level of neuronal firing during the seizure (Truccolo et al. 2011). This study further performed spike sorting (an analytic technique which would be precluded if cells were entering the paroxysmal depolarized state thought to be characteristic of full ictal recruitment; Matsumoto & Marsan, 1964), to show that the most active cells in this territory were likely to be interneurons and not pyramidal cells. Our earlier discussion of the information content of gamma rhythm is pertinent here. There are two possible explanations for this pattern of low level increases in firing, when recorded without a simultaneous recording of fully recruited neurons. The simplest explanation is that there is indeed a core recruited territory elsewhere in the brain that was not recorded in this study. A second possibility is that this lower level of firing distributed over a wide area may constitute the seizure activity, which is manifest at the clinical level by the summed activity over a wide area. Stereotypy may help distinguish these two possibilities: if the clinical picture and the (widely distributed) EEG electrodes indicate highly stereotyped events, and yet the unit activity recorded at a particular locus is not stereotyped, then the parsimonious explanation of this unit activity is that it does not represent the seizure proper.


We believe that these considerations about information potential of cortical networks are a useful way of considering seizure activity, and perhaps other forms of cortical pathology, that will lead to a more enlightened approach to mapping these pathologies in the clinical setting. The recent development of microelectrode arrays for use in humans has provided a more nuanced view of spatial patterns of activity during seizures. The picture that is emerging may redefine our view of cortical seizure activity.