Epilepsy is a disorder of the brain characterized by an enduring predisposition to generate seizures and by the neurobiological, cognitive, psychological, and social consequences of this condition (1). Although seizures are an important part of the definition, the associated cognitive and behavioral impairments and learning and memory problems are also important determinants of quality of life (2, 3). The quality of life is highly affected in patients with epilepsy especially if they are in adolescence, affecting the patient's self-esteem and sense of coherence (4). Self-esteem is a main contributor to psychosocial wellbeing, personal reflection, and positive attitude (5), and coherence is the ability to recognize stressors as manageable and solvable (4, 6). In a 5-year follow up study, it was shown that sense of coherence decreased in adolescents with epilepsy. Adolescents who still had seizures showed a greater decrease compared to seizure-free teens, with no effect of seizure frequency on sense of coherence (4). Self-esteem was also decreased in adolescents with epilepsy, however, self-esteem was affected by the seizure frequency where higher seizure frequency was associated with lower self-esteem (4). This indicates that psychosocial wellbeing is affected in adolescents with epilepsy seen by the decrease of both self-esteem and sense of coherence.

Gauffin et al. (7) examined the experience of living with epilepsy and cognitive decline. Their study found out that cognitive decline is persistently present in adults with intractable epilepsy and living with epilepsy and cognitive deficits affected education, employment, self-esteem, social life, and future plans (7). The cognitive deficits are seen in both focal and generalized seizures. Focal seizures occur in a lateralized network in contrast to generalized seizures occurring in a widespread network encompassing both hemispheres (8), however similar cognitive impairments are seen in both epilepsies (9). Patients with focal epilepsy experience various cognitive impairments such as language abnormalities, executive dysfunction, attention deficit and long-term episodic and semantic memory deficits (9–11). Patients with generalized epilepsy experience the same cognitive impairments in addition to acquired knowledge deficits and long-term information processing and retrieval impairment (9, 12, 13).

A common pathophysiological argument for the mechanisms of the relationship between epilepsy and cognition is that the seizure and epileptiform discharges in the EEG directly injure neural networks that are the normal substrate for cognitive function. An alternative hypothesis is that the relationship is indirect; both the seizures and the additional morbidities arise from neural networks that have been disrupted by the etiology of the epilepsy e.g., single gene disorders, malformations of cortical development, or traumatic brain injury. Our starting position is that the action potential is the fundamental unit of information processing in the brain, and that sequences of action potential firing in neuronal populations over time are therefore considered to be mechanisms of cognition, as cognitive function is explicitly about information processing (14, 15). Our goal in the current article is to explore the latter hypothesis in detail, starting at the systems level physiology of brain function and relating this exploration to our understanding of epilepsy. We will describe these system level mechanisms in physiology (Figure 1A) and discuss how these mechanisms are changed or altered in brain diseases associated with epilepsy (Figure 1B), especially in the hippocampus and neocortex, which are important structures involved in memory.

The firing rate of neurons in the brain is a mechanism of information transfer (15). For example, in motor neurons, the degree of muscle flexion depends on the number of action potentials per unit time (14). Tactile and texture perception in rodents is also, at least in part, a function of rate coding (16). In hippocampus and entorhinal cortex, rate coding is illustrated by place, time, and grid cells, respectively (17, 18).

Place cells are hippocampal pyramidal cells that fire when an animal visits a specific region of the environment: the cell's “place field” (19, 20). In any given environment, place fields cover the entire space to create a hippocampal cognitive map that is a representation of that space (19, 20). Repeated recordings of place cell populations have shown that the same cells are activated whenever the animal visits the same region in physical space, which suggests that the cells' representation is held in the network once the animal explores that specific field (21). This representation allows recollection of specific spaces and accurate navigation through the environment. This is further supported by the observation that lesioning the hippocampus results in loss of spatial memory (22). Similarly to place cells, time cells in the hippocampus fire at specific times in a task (e.g., the beginning, middle, or end) called time fields (23). These cells can be time locked to an external stimulus, like a tone, or intrinsically by a neural circuit or an oscillation (14, 15).

Grid cells are entorhinal cortical cells that provide activity-based maps of speed and direction in a certain environment (24). Grid cells fire in different locations in an environment forming a triangular grid. Recording different cells at the same location indicates that these cells have the same orientation to the environment (24, 25). It is important to note that the orientation of the grid relative to the environment is dependent on the hippocampal place cell map (20, 25, 26). Interestingly, the map formed by the grid cells is based on external cues, however the formed map persists even in the absence of these cues (24). Together with the place cell map, the grid cell map is believed to be part of the greater hippocampal cognitive map (20). Given what we know from work on place and grid cells, further experiments have shown that firing rate variations in CA3 place cells depended on signals from the lateral entorhinal cortex (24, 27). Lesioning the lateral entorhinal cortex impairs the hippocampal rate remapping upon changing the configuration of the environment (24, 27) suggesting that inputs from the entorhinal cortex are important for hippocampal rate coding in the formation of the spatial memory and cognitive map.

Rate coding as measured in individual neurons is essential for information processing. However, neurons are functionally connected into a network and interactions between the neurons is also critically important (15, 28). This is known as population coding. Population coding increases robustness of network function and minimizes the effects of noise carried by an individual neuron, ensuring that the signal and processing times are not affected (15). For example, damage to one cell will not have a devastating effect on the information being carried since it is carried by many cells (28). Population coding is common in the nervous system and is illustrated in mammalian visual pathways (29), primary motor cortex activity in cats and monkeys (30), owl auditory cortex, cricket nervous system (31), and mice visual cortex (32–34). Neurons in the visual cortex in cats and monkeys and the auditory cortex in owls have shown the ability to synchronize their firing on a few milliseconds time scale through time cells (14, 35). Synchrony seen in the visual cortex of cats and monkeys takes place when the neurons are activated by one stimulus (35), and this synchrony is lost when the neurons are activated by two independent stimuli. Once two neurons synchronize to represent a certain stimulus, these two neurons always synchronize to represent the same stimulus and will desynchronize when representing two independent stimuli (14).

Population coding influences information processing in mice visual systems. Whole cell recordings in layers 2/3 (L2/3) of awake mice have shown that the excitation/inhibition ratio changes based on the visual stimulus (33). Different studies revealed that patterns of excitation and inhibition are generated in response to various visual stimuli (36, 37). For example, as the stimulus contrast or size increases, the excitation/inhibition ratio decreases. This change in ratio is important for tuning and sharpening of the information processing in the visual system (38) and is thought to be controlled by the somatostatin (SOM) neurons and parvalbumin (PV)-expressing cells. SOM neuron suppression was able to enhance the excitation and inhibition for size tuning, indicating that contrast and size in the visual system in mice depends on excitation/inhibition ratio and tuning of total synaptic input (33). PV cell manipulation was shown to modulate L2/3 pyramidal cells spikes in response to visual stimuli without affecting the cells tuning properties suggesting that PV cells create a connection between certain neuron types and specific computations during sensory processing (32, 39). Furthermore, the excitation/inhibition ratio for the same set of stimuli differs between anesthetized and awake mice and even differs between various behavioral states (33). This change in the ratio is influenced by the recruitment of SOM and vasoactive intestinal peptide-expressing (VIP) inhibitory cells in V1 in both wakefulness and alertness for instance (40–42). The Stabilized Supralinear Network (SSN) model proposes that cortical dynamics can change the excitation/inhibition ratio based on single neurons input/output supralinear relationships, strong recurrent excitation, and feedback inhibition (33, 43). This suggests that the dynamic change of excitation/inhibition ratio depends on the dynamic connections between different cortical regions and the feedback loops between the recruited regions. Further support to this suggestion was seen in experiments investigating the rhythmic activity across cortical layers. Population of excitatory neurons in the mouse's primary visual cortex expressed gamma band oscillation following layer-specific optogenetic stimulation (34). Importantly, rhythm-generating circuits in each layer were able to provide layer-specific excitation/inhibition balances hence influencing the information flow between cortical layers (34). These studies provide evidence that neuronal populations can be recruited through precise synchrony contributing to information processing at a system level.

Notably, place cells in the hippocampus are an excellent example of not only rate coding, but also population dynamics. We previously mentioned that the place cells fire whenever the animal is in a specific place field and these cells are distributed within the hippocampus in a way that covers the whole environment the animal occupies. This distribution within the hippocampus will cause a certain degree of overlap between the place fields, thus a population of cells will respond when the animal goes into the field rather than an individual cell since the cells in the hippocampus are receiving multiple sensory inputs to encode a multidimensional map (44).

Another interesting example of population coding is pattern separation. Pattern separation depends on the discrimination between two closely related places, episodes, or spatial configurations based on experience and can influence successful memory encoding. Pattern separation involves different brain regions where experiences are represented by neural populations. Notably, if the same population or neural pattern established during encoding is activated during retrieval, it can lead to a successful memory retrieval (45). For instance, both the posterior occipitotemporal cortex (OTC) and the hippocampus were recorded while participants performed item recognition tasks. Upon retrieval, both regions showed encoding-specific high frequency activity (HFA) where the strength of this activity was associated with enhanced retrieval, however the discrimination between similar items required a hippocampal activity (45) as the pattern separation mechanism is based on orthogonalizing similar input during encoding thus enabling the distinguishing between highly similar memories with minimal interference (46). Animal and human studies have shown that dentate gyrus (DG) and its projection to CA3 underlie the pattern separation process (47–52).

A final relevant example of population coding is working memory in the prefrontal cortex. Working memory is the temporary maintenance of information involving specialized components of cognition that allows retaining immediate past-experience, supporting new knowledge acquisition, solving problems, reasoning, and planning (53). Early models of working memory suggested that persistent firing activity of the neurons in the prefrontal cortex (PFC) throughout the delay phase of the working memory task was required to maintain information in working memory, however, due to the heterogeneity of neurons within the PFC, recent work has shown that the persistent activity of PFC can be weak or absent (54–56). This was seen during the delay phase of an image-sequence matching task in monkeys and humans. In this task, spiking activity in the PFC decreased during the delay phase of the task in monkeys, and BOLD signal on fMRI decreased during the delay phase in humans performing the task, thus challenging the persistent firing working memory model (57, 58). However, the spiking activity and the BOLD signal re-emerge during image presentation and testing period, indicating that, despite a decrease in firing rate throughout the task, the working memory information is maintained in the collective synaptic weights of populations of neurons in the PFC. The heterogeneity of the neurons allows different neurons to signal during different task events as well. For example, parvalbumin-positive neurons will respond to sensory cues and trial outcomes while somatostatin-positive neurons will respond to a motor action like licking (59). These neurons can fire together in working memory maintenance through population coding when a robust mnemonic stimulus is present (60) as seen in monkeys performing oculomotor delayed response and vibrotactile delayed discrimination, which are working memory tasks, while recording single neurons of lateral PFC (60).

Oscillatory activity is divided into frequency bands as described in Table 1: infra-slow oscillations (0.5–1 Hz), delta (1.5–4 Hz), theta (4–8, 10 Hz), alpha (8, 10–12 Hz), beta (15–30 Hz), gamma (30–80 Hz), in addition to fast (80–200 Hz), and ultra-fast (200–600 Hz) ripples (61). Each band is thought to be related to specific aspects of cognition (Table 1). Both in vivo and in vitro experiments suggest that synaptic inhibition plays a role in generating neuronal oscillations through two different mechanisms, either through interneuron network activity or reciprocal excitatory-inhibitory loops (62). Theta, as previously discussed, and gamma are two important readouts of the hippocampal function and function of connected regions.

Theta rhythm indicates a network that is actively involved in spatial navigation, working memory, and temporal coding (63, 64). It is important to note that theta oscillations and phase-locked neuron discharges with respect to theta oscillations are seen in theta non-generating regions like entorhinal cortex, perirhinal cortex, cingulate cortex, subicular complex, and amygdala (63, 65–67). Lesion or inactivation of the medial septum-diagonal band of Broca neurons recognized as the theta rhythm generators eliminates theta oscillation in all connected cortical regions (68) and leads to spatial and working memory deficits (69–71).

Gamma oscillations ride on top of theta in the hippocampus and dynamically couple hippocampal networks to specific behavioral demands (39, 72, 73). The gamma oscillation is thought to be a readout of information transfer from CA3 to CA1, where CA3 is influencing the activity of CA1 for hippocampal memory retrieval to underlie memory encoding, consolidation, and episodic memory retrieval. Movement sequences, memory encoding and formation, sensory processing and planned trajectories underlying spatial navigation are thought to be temporally organized through timing mechanisms seen through gamma oscillations (74, 75).

The rate of populations of neuronal firing is also modulated in time in respect to theta frequency. For example, in the hippocampus, temporal modulation is manifested as burst firing with bursts occurring at theta frequency (4–12 Hz). The fidelity of burst firing within theta, termed theta modulation, is important for phenomena such as phase precession, phase preference and hippocampal replay, which are believed to allow encoding of space with higher resolution than is possible in the absence of modulation. Phase preference refers to the phenomenon that neuronal firing in the hippocampus is often locked in time with respect to ongoing hippocampal inputs in the theta frequency of the local field potential (LFP). Specific cells fire preferentially at specific phases in the ongoing theta oscillation. For example, directly after the peak of the theta oscillation, PV+ basket cells in CA2/3 area fire at the same phase as pyramidal cells in CA3, but later than basket cells in CA1 (76). CA1 pyramidal cells preferentially fire at the trough of theta (77, 78). Phase precession describes the observation that place cells will discharge whenever the rodent is crossing a place field and this firing occurs at an earlier phase in theta with progressive theta cycle (79, 80). This phase precession is believed to be an important component of information processing. Theta-phase precession could be an indication of item-context associations through spike timing-dependent plasticity (80, 81) given that synaptic inputs need to be precisely synchronized within 5 ms so that EPSPs from different locations will be able to induce postsynaptic firing (15). Umbach et al. showed that neurons in the hippocampus and entorhinal cortex not only fire for space, but also for time. Interestingly, time cells also exhibited theta-phase precession during memory encoding, and the activity of these cells correlates with the use of temporal location during the retrieval phase of the task (82). Neuronal firing coordination with the LFP, like phase-locking and phase precession, offers a key glimpse at the relative timing of inputs in the LFP with outputs of the information processing as the neuronal firing, but neural oscillations are also readouts of synchronized behavior of the network and, as such, are on their own important mesoscale mechanisms of cognition, memory, and behavior.

Coordination between oscillations seen in the LFP or EEG can be an indicator of communication between different brain regions (83, 84). Coherence is a measure of this synchronization with values that range between 0 and 1. The higher the coherence, the more synchronized the regions are (84). The coherence value differs between different brain regions depending on the task performed, for example, theta coherence between hippocampus and striatum during periods of decision is high (>0.8) which indicate learning (84, 85). Not only theta coherence changes during a task, but also gamma as well. Attention tasks in monkeys have shown that gamma coherence increases between the parietal and prefrontal areas (86). Furthermore, CA1 can become coherent with the entorhinal cortex or CA3 through fast or slow gamma characteristics of the entorhinal cortex or CA3, respectively (87). Elevation of hippocampal-entorhinal cortex synchrony was shown to be important for declarative memory formation in epileptic patients performing a memorization task (88). This led to the hypothesis that synchronized brain activity in the gamma range might be an important indicator of controlled flow and routing of information (83, 89), because the rules guiding synaptic plasticity dictate that inputs will be most effective whenever they coincide with peaks of oscillatory network excitability (83). Finally, neural coherence alterations were associated with different disorders like schizophrenia, attention-deficit hyperactivity disorder (ADHD), Alzheimer's disease (AD), and temporal lobe epilepsy (TLE) (90–95) (Table 2).

Temporal, population, and rate coding facilitate plasticity shaped through experiences that enable the brain to adapt to new information. These mechanisms underlie the careful coordination of information between synapses and neurons in the brain that is necessary to promote synaptic plasticity and ensure efficient flow of information between different brain regions required for cognition. In 1949, Donald Hebb postulated that synapse strength can change based on previous activity, which led to what we now know as long-term potentiation (LTP) and long-term depression (LTD), fundamental to network communication. LTP strengthens synaptic transmission through high frequency stimulation of synapses. The first stage of LTP depends on the NMDA and AMPA glutamate receptors (96, 97). The second and third stages of LTP depend on protein synthesis to maintain changes in synaptic strength (96, 97). Maintenance of LTP is essential for place cell stability (98–101). Even though neural plasticity is not a determinant of place cell spatial specificity, rats with neural plasticity deficits had unstable place fields upon revisiting the same environment (102, 103). This shows that neural plasticity is playing a role in the organization of place cells and long-term maintenance of this representation.

The frequency of action potential timing matters; low frequency firing induces LTD, which decreases synaptic efficacy. LTD is also essential for memory formation as it counteracts the LTP to allow new memories to form. Recent evidence has shown that LTD may be involved in formation and maintenance of place fields (104), supporting previous experiments showing that a decrease in the expression of LTD impairs spatial memory retention and consolidation (105, 106). It is worth noting that plastic synapses can form positive feedback loops on the rate, temporal, and population coding mechanisms where this positive loop aids the mechanisms in refining and precisely timing the neuronal firing leading to a more efficient information processing. Another form of plasticity known as short-term plasticity (STP) takes place on a millisecond to minutes timescale and depends on presynaptic calcium accumulation and vesicle depletion (107). This form of plasticity is thought to play a role in information transfer across synaptic connections, activity-dependent synaptic efficacy modulation, promoting synchronization and working memory (107–109). Careful coordination of the firing of populations of neurons in time supports appropriate short and long-term forms of plasticity that are critical for information processing, learning and memory.

Cognitive impairments in people with epilepsy are extremely common and have a major negative influence on quality of life. Patients with focal epilepsy have shown to have a significant decrease in their quality of life compared to patients with generalized epilepsy and healthy controls, however, both focal and generalized epilepsy patients have a decreased self-esteem and increased anxiety compared to healthy controls (9). Identification of mechanisms of cognitive impairment is important as this will help to guide development of novel therapeutic strategies to improve outcomes. A major emphasis has been on studying the impact of seizures on cognition and exploring the epileptic encephalopathy hypothesis. Researchers have focused on the time of onset and frequency of seizures, however, there are few clear correlations between seizure characteristics and cognitive outcomes. Taken into consideration that cognitive functions are dependent on complex brain networks and both focal and generalized epilepsy groups share the same cognitive impairments (9), we can say that seizure location is less of a determinant of the cognitive impairments than an altered or dysfunctional network in the brain. This suggests that there must be other factors influencing cognitive outcomes in epilepsy. This assertion is supported by several lines of evidence. Studies assessing the effect of age of onset on cognitive impairment have shown that cognitive impairment already exists at pretreatment baseline in newly diagnosed children (3, 110–112). This led some researchers to question whether cognitive deficits could be present even before seizure onset (113) and this might suggest that the pre-existing impairments could be a result of the same dysregulation that underlies the seizures in the first place. Indeed, the need for special educational services prior to epilepsy onset is more common in children who were later diagnosed than those who weren't (114), suggesting that impairment in cognitive function may be present even before the first seizure. The best predictor of cognitive outcome up to 3 years after the diagnosis of epilepsy in infants is the initial cognitive profile, not any seizure or medication related factor. In adults, these impairments extend to deficits in visual motor tasks, mental flexibility, memory, reaction times, and attention (3, 115).

Others have studied the effects of the disorder duration on cognition, finding a negative correlation between the number of years of the disorder and brain volume (116, 117). This has been interpreted by some to mean that the length time since disease onset is related to the amount or significance of cognitive impairment. However, an alternate interpretation is that early onset of an epilepsy disorder is an indication of a more fundamentally dysfunctional network, leading to early development of recurrent seizures. Similarly, seizure frequency has been noted to be associated with a detrimental effect on cognition, with higher seizure frequency being correlated with lower performance on cognitive tasks and vice versa (9, 118). This was always interpreted to mean that seizures themselves were detrimental to cognition, rather than more frequent seizures being an indicator of a more abnormal neural network that underlies abnormal functioning during the interictal period manifesting as cognitive impairment. Hence, these impairments might not be due to age of onset or frequency of seizures, instead it could be attributed to the fact that children in this situation have a brain disease that presented earlier, and this difference in disease presentation may be an indicator of more severe network dysfunction, and more severe impairments. Further investigation of memory dysfunction in patients with epilepsy showed that people with epilepsy had significant deficits in both semantic and episodic autobiographical memory (9, 119, 120). This deficit was associated with young age at onset, more frequent seizures, and reduced working memory in early-onset epilepsy patients. In contrast, the same deficit was related to depression and lesion (3, 9, 120). The presence of both neurobiological and psychological factors suggests that information processing mechanisms might be altered (121).

Information processing through the mechanisms discussed in this review is shown to be altered in epilepsy and associated disorders. CA1 place cells are unstable in epileptic mice and undergo remapping a few weeks after pilocarpine-induced temporal lobe epilepsy (TLE). The number of place cells decreases, and the spatial tuning curve is less stable over time (122, 123). Prolonged recording over days from populations of neurons in CA1 and dentate gyrus has shown desynchronized interneuron firing between these two areas (124), which suggests that disruption of spatial coding is due to the loss of information processing control by interneurons. The desynchronized interneuron firing can affect the timing of the inputs being sent to the CA1. This was supported by the observation of theta rhythm temporal coordination loss in the dentate gyrus, where these neurons were firing at inconsistent phases of the CA1 theta rhythm (124). Spatial memory alteration was previously shown to be present even during the latent, seizure-free, period after either the pilocarpine-induced status epilepticus (SE) or early life seizures during the 1st weeks of life (39, 125, 126). These deficits were associated with a decrease in the power of theta oscillations (125). It is important to note that spontaneous seizures did not modify or affect any of the spatial deficits that were already present (125). Interestingly, Shuman et al. (124), found that place cell deterioration and place coding alteration occurred several weeks after pilocarpine induction, showing that the development of seizures is not solely responsible for place cell deterioration. Notably, place coding alteration, place cell deterioration, more dispersed place fields, and fewer place field responses were also seen after silencing either CA3, entorhinal cortex or both (124, 127, 128).

In addition, we and others have shown dysregulated population coding in epilepsy models. In-vivo single-unit recording showed that CA1 pyramidal cells are functionally connected to other pyramidal cells and fire in a coordinated fashion during spatial memory tasks; this connectivity is altered in TLE where neuronal reactivation and synchrony predicts the behavioral outcome in a TLE model (129). Population coding functional connectivity is also crucial within the hippocampus and between the hippocampus and PFC to underlie spatial working memory (SWM) (130, 131). During a SWM task, the hippocampal-PFC network shows a distributed dynamic code, seen through temporally regulated firing within and between brain regions, which is needed to combine separate processes together to execute a SWM task (131). The coordinated firing of cells in time is important for several components like attention, decision making and long-term memory, which can predict task performance. The temporal modulation of populations of neurons predicted SWM accuracy in a delayed non-match-to-sample task in control rats and rats with a cortical malformation that, in humans, is an important etiology in epilepsy. Animals with cortical malformations showed deficits in hippocampal firing modulation in addition to decreased functional connectivity between neurons (131).

Furthermore, population coding and neural dynamics are important for pattern separation and this process has been shown to be altered in hippocampal injury and epilepsy. The pattern separation depends on a network spanning different brain regions other than hippocampus, like the dorsal medial prefrontal cortex (dmPFC), however the hippocampus and the parahippocampal cortex serve as a hub for this network (132), thus it is expected that a hippocampal injury will alter the network communication causing pattern separation deficits (47, 133, 134). TLE patients and amnestic mild cognitive impairment (aMCI) patients have pattern separation deficits, and this could be due to hippocampal dysfunction involving DG and CA3 (46, 135). Another reason could be due to the failure of separating similar information during encoding by the hippocampus, hence memories will not be accurately encoded or retrieved. Studies investigating aMCI and TLE patients have shown that aMCI patients have an excess activation of the DG/CA3 area in fMRI compared to control groups and this excess activation is correlated with poor performance on pattern separation tasks. The same poor performance was seen in TLE patients performing the Mnemonic Similarity Task (MST). TLE patients demonstrated poor pattern separation performance compared to controls, however, it is important to note seizure and hippocampal sclerosis did not affect the performance of patients in this task (46, 133). Following studies showing that TLE patients have spatial mnemonic discrimination impairment and that TLE mice have DG-dependent object location memory deficits (50, 136), Madar et al. (52) tested pattern separation in TLE patients and mice with TLE, and then used mouse brain slices to record the spiking patterns of single granule cells (GC) in the dentate gyrus. TLE patients performing object recognition-based MST had a significant deficit in identifying similar but not identical objects suggesting that TLE might be impairing the DG-dependent mnemonic discrimination. Similar deficits were seen in mice with TLE as the mice had a decrease in object-location mnemonic discrimination compared to control mice (52). Slice electrophysiology in the same mice utlized inputs mimicking the same recorded inputs during behavior, and indicated that the output spike-trains of GCs had a higher average correlation compared to input correlation, which signifies a deficit in pattern separation in mice with TLE. Different input ranges demonstrated decreased pattern separation and convergence in DG at multiple timescale levels (52). This shows the importance of population dynamics underlying spatial deficits and signals the importance of assessing functional connectivity.

Imaging and histological experiments showed that structural and functional connectivity were altered in TLE patients as well (3, 137). Histological changes have been observed in the amygdala, entorhinal and parahippocampal cortices in TLE patients (3, 138–142). MRI images investigating hippocampal sclerosis associated with TLE, found that in addition to hippocampus, atrophy is present in the adjacent mesiotemporal, temporopolar structures, and thalamus (3, 143–146), and this atrophy increases over time (147–149). Experiments investigating tissue microstructure and structural covariance indicate that structural connectivity was impacted in TLE. Diffusion tensor MRI showed a disorganization in fiber arrangement in temporolimbic and adjacent regions (3, 150–153). Structural covariance such as cortical thickness or gray matter volume was altered between the mesiotemporal and neocortical regions and within the corticocortical networks (154–156). Resting state functional connectivity revealed a deficit in network connectivity in TLE patients compared to healthy controls. TLE patients had a decrease in ipsilateral mesiotemporal networks connectivity and ipsilateral and contralateral hippocampi connectivity (3, 153, 157–159). This decrease in connectivity extends beyond the temporal lobes into the posterior cingulate, inferior parietal, and medial prefrontal cortices disrupting the default mode network (DMN) (160–164). These changes and deficits suggest that structural connectivity is impacted in TLE patients and that TLE is also associated with functional connectivity deficits and reorganization.

Early stage TLE patients experience functional connectivity deficits mainly in the ipsilateral hemisphere (162, 165) in addition to disturbed interhemispheric connections (3, 158, 166). However, in patients with generalized epilepsy, there is an increase in the interhemispheric connectivity in addition to reduced functional connectivity (167–171). fMRI studies investigating the network connections in epileptic brains showed an increase in functional connectivity within the temporal lobe, alongside a decrease between temporal and other regions. Also, there is a decrease in the connection probability between neighboring brain regions, known as the clustering coefficient, within the DMN (161, 172). This decrease in clustering coefficient as well as increased path length, i.e., distance between one node and another, was revealed to be associated with cognitive decline in patients with cryptogenic epilepsy and only seen in patients with cognitive decline (172–174). The decreased cluster coefficient within the DMN could underlie the language impairment in patients with generalized epilepsy without focal brain damage. Gauffin et al. (175) conducted an experiment where patients with generalized epilepsy without focal damage performed a sentence-reading task while going through fMRI. Patients with generalized epilepsy took longer time to read both congruent (simple) and incongruent (complex) sentences compared to healthy controls with no reading time difference between congruent and incongruent sentences in the patients group which suggests that patients perceived both types as complex (175). BOLD fMRI indicated the activation of a left-lateralized frontotemporal network, anterior cingulate cortex and occipital cortex in both patients and controls upon reading both types of sentences (175). However, patients with generalized epilepsy had reduced DMN suppression compared to healthy controls (175). Further lack of suppression was seen in the left anterior temporal lobe and the posterior cingulate cortex, in addition to irregular activation of the right hippocampus proper and right parahippocampal gyrus (175). The reduced DMN activity suppression can be due to reduced functional segregation of the DMN in generalized epilepsy patients (170) where this can alter the balance between activated and deactivated neural networks hence disturbing the cognitive function (176, 177). Further evidence of network alteration in TLE patients was seen by Bernhardt et al. (178) upon analyzing hub nodes between controls and TLE patients. Hubs are also known as nodes that have multiple connections within a network with one central position and the connections formed by the hub nodes are essential for communication and network synchronization (179). Hub nodes in TLE patients were mainly located in the limbic and temporal association cortices instead of being evenly distributed between different lobes and this was thought to be due to connectivity disturbances between the temporolimbic and extratemporal neocortical structures (178) providing evidence that epileptic brains express decreased integration and enhanced segregation (172). It is also important to note that memory impairments are present in patients who don't show a lesion with MRI (180) which further supports the notion that cognitive impairments depend on the affected network rather than a structural lesion (9). These studies emphasize the necessity to move beyond the classical lesion model into a network approach which can provide several advantages by helping track or predict cognitive decline in epilepsy patients, improving diagnosis, and developing more accurate resection surgeries by targeting the areas where the hub nodes are mostly concentrated.

It is critical to note that experimental designs that induce an underlying disorder associated with epilepsy, but in which there are no overt seizures, and no other subclinical epileptiform activity was noted, show changes in information processing and behavioral deficits. Loss of function of sodium channels Nav1.1 associated with human epilepsy in CA1 can cause disruptions to place cells and spatial cognition without producing seizures (181). Nav1.1 knockdown in the medial septum causes alterations in temporal and rate coding in those neurons, and deficits in working memory that are correlated with the degree of LFP alteration in the hippocampus rather than seizure frequency (39, 182). Similar effects are seen in animals with a malformation of cortical development where no overt or subclinical seizures were noted. These animals have reduced fidelity of place cells, reduction in the magnitude of theta modulation, and disrupted population coding in addition to spatial and working memory deficits. The addition of induced seizures in this model did not make the behavioral deficits worse, indicating that the main contributor to the cognitive impairment was the underlying brain substrate and not seizures (183).

Notably, subclinical epileptiform activity or inter-ictal spikes (IIS) can disrupt cognitive function; however the number of spikes is not a reliable indicator of the associated cognitive impairment. Several studies have previously shown that patients with benign epilepsy with centro-temporal spikes experience IQ and school performance deficits (184, 185). These deficits were correlated with the frequency of IIS but not seizure frequency (184, 185). However, this may be related to timing of the IIS relative to ongoing cognitive processing, as the presence of the IIS may be an indicator that the brain is not in a state where it can be performing cognitive computations. Kleen et al. (186) investigated the effect of focal IIS on hippocampus in TLE. They showed that rats with unilateral intrahippocampal pilocarpine infusion developed hippocampal spikes that caused a response latency deficit in hippocampal-dependent operant behavior task, delayed-match-to-sample (186). However, the hippocampal spikes only altered the cognitive performance when they occur at the same time during memory retrieval; spikes occurring during memory encoding or maintenance did not affect the cognitive performance and overall IIS frequency during a trial was not predictive of accuracy during that trial (186). Similar results were seen in patients with refractory seizures performing Sternberg task, a delayed information task that depends on short-term memory processes, along with EEG recordings (187). Contralateral or bilateral to seizure focus hippocampal interictal epileptiform discharges (IED) during memory retrieval disrupted memory retrieval, and bilateral IED during memory maintenance was able to disrupt that process, however no effect was seen on memory encoding (187). These studies show that focal IIS and hippocampal IED are associated with disruptions in memory maintenance and retrieval only when they occur during the same time window as the memory processes. Taken together, this suggests that IIS/IED are indicators of disrupted network processing underlying cognition.

In addition to deficits in rate, temporal and population coding, plasticity deficits are also present in epilepsy, in accordance with the view that these neural coding mechanisms support plasticity. Kainic-acid induced status epilepticus (SE) model in rats shows a significant decrease in hippocampal LTP in addition to cell loss, and signs of hippocampal sclerosis (188). These rats also have deficits in the hippocampal-dependent novel object recognition spatial memory task that positively correlated with LTP magnitudes (188). These findings were also seen in the pilocarpine model where the mice showed a significant decrease in the hippocampal synaptopodin acting-binding protein in CA1 region which alters the ability of the neurons to express synaptic plasticity leading to a decrease in LTP induction in Schaffer collateral-CA1 synapses (189). STP and working memory are also altered in kainic acid-induced SE. Following kainic acid-induced SE, there was a decrease in STP, reduced LTP capacity, impaired spatial learning, and increased inhibition in the dentate gyrus (190). STP was altered in a model with recurrent hyperexcitability leading to seizures during development (191), as well as a model with aberrant GABA signaling during development leading to frequent interictal discharges. Animals with frequent IDs in the developing PFC showed a decrease in attention, and sociability alongside these changes in STP (192). The growing evidence on neural networks and epilepsy shows that these disrupted neural networks are likely responsible for the cognitive impairments seen with the disease and that the underlying etiology is the cause of both the disease and coding impairments seen in epilepsy animals and patients as well. The corollary is that recovering neural networks toward normal has potential for recovering cognitive impairments (Tables 3, 4).

The main issue with finding the appropriate treatment is whether neural network function can be recovered even in the context of a diseased brain. Here we will discuss potential therapeutic approaches that might influence the neural network function.

In this review, we addressed the role of neuronal dynamics in supporting proper cognition, learning and memory, and discussed how these dynamics are altered in epilepsy. The data suggest that cognitive impairments seen in patients with epilepsy and preclinical models of epilepsy are likely due to plasticity changes, alterations to neuronal coding regimes, desynchronization, and functional connectivity disruptions from the effect of underlying etiology, rather than seizures themselves. Although we accept that the seizures could also have some negative impact on network behaviors, we strongly argue that the seizure effect is very small when compared to the etiology effect. We therefore suggest that these deficits should be approached from a systems neuroscience perspective, while being informed by mechanisms needed for normal cognitive function and development in a dynamic experience-dependent and plastic network. Importantly, this calls us to move beyond seizures into network science that is guiding possible treatments and defining new pathophysiology. This might help advance the epilepsy research forward and open the door potentially to answer unsolved questions in the field.

MK, AH, and RS: draft manuscript writing and editing. All authors reviewed the final draft manuscript. All authors contributed to the article and approved the submitted version.

AH and MK were funded by an NIH NINDS K22NS104230. RS was funded by an NIH NINDS R21NS117112.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.