Biomimetic Deep Learning Networks With Applications to Epileptic Spasms and Seizure Prediction

OBJECTIVE

In this study, we present a novel biomimetic deep learning network for epileptic spasms and seizure prediction and compare its performance with state-of-the-art conventional machine learning models.

METHODS

Our proposed model incorporates modular Volterra kernel convolutional networks and bidirectional recurrent networks in combination with the phase amplitude cross-frequency coupling features derived from scalp EEG. They are applied to the standard CHB-MIT dataset containing focal epilepsy episodes as well as two other datasets from the Montefiore Medical Center and the University of California Los Angeles that provide data of patients experiencing infantile spasm (IS) syndrome.

RESULTS

Overall, in this study, the networks can produce accurate predictions (100%) and significant detection latencies (10 min). Furthermore, the biomimetic network outperforms conventional ones by producing no false positives.

SIGNIFICANCE

Biomimetic neural networks utilize extensive knowledge about processing and learning in the electrical networks of the brain. Predicting seizures in adults can improve their quality of life. Epileptic spasms in infants are part of a particular seizure type that needs identifying when suspicious behaviors are noticed in babies. Predicting epileptic spasms within a given time frame (the prediction horizon) suggests their existence and allows an epileptologist to flag an EEG trace for future review.

Phase-Amplitude Coupling Features Accurately Classify Multiple Sub-States Within a Seizure Episode

Epilepsy is frequently characterized by convulsive seizures, which are often followed by a postictal EEG suppression state (PGES). The ability to automatically detect and monitor seizure progression and postictal state can allow for early warning of seizure onset, timely intervention in seizures themselves, as well as identification of major complications in epilepsy such as status epilepticus and sudden unexpected death in epilepsy (SUDEP). To test whether it is possible to reliably differentiate these ictal and postictal states, we investigated 52 seizure records (both intracranial and scalp EEG) from 19 patients. Phase-amplitude cross-frequency coupling was calculated for each recording and used as an input to a convolutional neural network model, achieving the mean accuracy of 0.890.09 across all classes, with the worst class accuracy of 0.73 for one of the later ictal sub-states. When the trained model was applied to SUDEP patient data, it classified seizure recordings as primarily interictal and PGES-like state (70% and 26%, respectively), highlighting the fact that in SUDEP patients seizures primarily exist in postictal states and don't show the ictal sub-state evolution. These results suggest that using frequency coupling markers with a machine learning algorithm can reliably identify ictal and postictal sub-states, which can open up opportunities for novel monitoring and management approaches in epilepsy.

Delta-gamma phase-amplitude coupling as a biomarker of postictal generalized EEG suppression

Postictal generalized EEG suppression is the state of suppression of electrical activity at the end of a seizure. Prolongation of this state has been associated with increased risk of sudden unexpected death in epilepsy, making characterization of underlying electrical rhythmic activity during postictal suppression an important step in improving epilepsy treatment. Phase-amplitude coupling in EEG reflects cognitive coding within brain networks and some of those codes highlight epileptic activity; therefore, we hypothesized that there are distinct phase-amplitude coupling features in the postictal suppression state that can provide an improved estimate of this state in the context of patient risk for sudden unexpected death in epilepsy. We used both intracranial and scalp EEG data from eleven patients (six male, five female; age range 21–41 years) containing 25 seizures, to identify frequency dynamics, both in the ictal and postictal EEG suppression states. Cross-frequency coupling analysis identified that during seizures there was a gradual decrease of phase frequency in the coupling between delta (0.5–4 Hz) and gamma (30+ Hz), which was followed by an increased coupling between the phase of 0.5–1.5 Hz signal and amplitude of 30–50 Hz signal in the postictal state as compared to the pre-seizure baseline. This marker was consistent across patients. Then, using these postictal-specific features, an unsupervised state classifier—a hidden Markov model—was able to reliably classify four distinct states of seizure episodes, including a postictal suppression state. Furthermore, a connectome analysis of the postictal suppression states showed increased information flow within the network during postictal suppression states as compared to the pre-seizure baseline, suggesting enhanced network communication. When the same tools were applied to the EEG of an epilepsy patient who died unexpectedly, ictal coupling dynamics disappeared and postictal phase-amplitude coupling remained constant throughout. Overall, our findings suggest that there are active postictal networks, as defined through coupling dynamics that can be used to objectively classify the postictal suppression state; furthermore, in a case study of sudden unexpected death in epilepsy, the network does not show ictal-like phase-amplitude coupling features despite the presence of convulsive seizures, and instead demonstrates activity similar to postictal. The postictal suppression state is a period of elevated network activity as compared to the baseline activity which can provide key insights into the epileptic pathology.

Glial Modulation of Electrical Rhythms in a Neuroglial Network Model of Epilepsy

OBJECTIVE

An important EEG-based biomarker for epilepsy is the phase-amplitude cross-frequency coupling (PAC) of electrical rhythms; however, the underlying pathways of these pathologic markers are not always clear. Since glial cells have been shown to play an active role in neuroglial networks, it is likely that some of these PAC markers are modulated via glial effects.

METHODS

We developed a 4-unit hybrid model of a neuroglial network, consisting of 16 sub-units, that combines a mechanistic representation of neurons with an oscillator-based Cognitive Rhythm Generator (CRG) representation of glial cells-astrocytes and microglia. The model output was compared with recorded generalized tonic-clonic patient data, both in terms of PAC features, and state classification using an unsupervised hidden Markov model (HMM).

RESULTS

The neuroglial model output showed PAC features similar to those observed in epileptic seizures. These generated PAC features were able to accurately identify spontaneous epileptiform discharges (SEDs) as seizure-like states, as well as a postictal-like state following the long-duration SED, when applied to the HMM machine learning algorithm trained on patient data. The evolution profile of the maximal PAC during the SED compared well with patient data, showing similar association with the duration of the postictal state.

CONCLUSION

The hybrid neuroglial network model was able to generate PAC features similar to those observed in ictal and postictal epileptic states, which has been used for state classification and postictal state duration prediction.

SIGNIFICANCE

Since PAC biomarkers are important for epilepsy research and postictal state duration has been linked with risk of sudden unexplained death in epilepsy, this model suggests glial synaptic effects as potential targets for further analysis and treatment.

Cross-Frequency Coupling Features of Postictal Generalized EEG Suppression State

In patients with epilepsy, convulsive seizures are often followed by a postictal generalized EEG suppression (PGES) state characterized by reduced background activity. Recent studies found a correlation between seizure termination state and PGES duration, and suggested that PGES is the result of the cessation of neuronal activity. To test that assertion, we investigated ten seizure records obtained from intracranial EEG (iEEG) from six patients, four of which had Engel Class 1 surgical outcome. In each case expert neurologists identified the most likely seizure onset electrode. We found the iEEG equivalent of PGES and an artifact-free preictal quiescent state of the same window size. Using index of cross-frequency coupling (ICFC) we identified the degree of coupling and dominant frequency bands involved in PGES and preictal quiescent states, and quantified the areas of high ICFC. We found that there was an increase in the degree of coupling between the 0.5-1.5Hz with high gamma frequency bands in the PGES states. We found that among all of the patients, as well as in Engel Class 1 patients specifically, the change in the quantified area of high ICFC was significant (p <; 0.05) between PGES and preictal quiescent states. Furthermore, we were able to identify whether a recording was from a depth or subdural electrode, or whether it was from seizure onset zone or not using ICFC markers in PGES. This suggests that there are frequency coupling markers that successfully identify PGES and that there are underlying dynamics that occur in this seemingly quiet postictal state.

Excitation and Inhibition Balance Underlying Epileptiform Activity

OBJECTIVE

The phenomenon of postictal generalized EEG suppression state (PGES) - a period with suppressed activity following seizure termination and has been found to be associated with sudden unexpected death in epilepsy - remains poorly understood. This article aims to examine the how the balance of excitation and inhibition (E/I balance) affect the dynamics of seizure and PGES.

METHODS

A network of 1000 Izhikevich model neurons was developed and only the strengths of synaptic connections were adjusted to recreate the dynamics observed in recordings of seizure and PGES from human patients.

RESULTS

A rapid rise followed by a slow decay of dominant frequency was observed in iEEG recordings of ictal periods and reproduced in the simulated local field potential by changing the E/I balance of the model network. The rate of this dominant frequency evolution was quantified by a single measure, β, which was found to have a significant rank correlation with the duration of PGES in iEEG data and the rate of E/I balance shift in the model. Significance and Conclusion: (i) highlighting the importance of E/I balance in the dynamics of seizure and PGES; (ii) suggesting the measure, β, as a marker for PGES and the shift in E/I balance as a neural correlate for this marker.

Neuro-Glial Network Model Of Postictal Generalized EEG Suppression (PGES)

Over the past couple of decades, glial cells have been highlighted as active agents in hyperexcitability of neuronal networks, specifically playing key roles in seizure onset and termination. In particular, microglia have been suggested to have both neuroprotective and neurotoxic effects on the brain. Investigation into seizure termination is of particular interest, as it is sometimes followed by a postictal generalized EEG suppression (PGES) - a low activity state that is potentially associated with sudden unexpected death in epilepsy. In this study, we attempt to link glial effects - synaptic pruning and astrocytic potassium clearance - to the duration of spontaneous epileptiform discharges (SEDs) as well as interSED intervals (iSEDs). We build upon an earlier model of a neuroglial network by translating it into the cortical paradigm and including microglial units. Preliminary findings of our model demonstrated that the duration of SEDs is largely determined by the astrocytic potassium clearance, whereas iSEDs significantly increased with microglial-driven synaptic pruning. In our model, astrocytic potassium clearance itself did not bring a PGES-like state, whereas microglial effects did, which suggests a potential biomarker for PGES phenomena.

Coupled Oscillators Model of Hyperexcitable Neuroglial Networks

Glial populations within neuronal networks of the brain have recently gained much interest in the context of hyperexcitability and epilepsy. In this paper, we present an oscillator-based neuroglial model capable of generating Spontaneous Electrical Discharges (SEDs) in hyperexcitable conditions. The network is composed of 16 coupled Cognitive Rhythm Generators (CRGs), which are oscillator-based mathematical constructs previously described by our research team. CRGs are well-suited for modeling assemblies of excitable cells, and in this network, each represents one of the following populations: excitatory pyramidal cells, inhibitory interneurons, astrocytes, and microglia. We investigated various pathways leading to hyperexcitability, and our results suggest an important role for astrocytes and microglia in the generation of SEDs of various durations. Analysis of the resultant SEDs revealed two underlying duration distributions with differing properties. Particularly, short and long SEDs are associated with deterministic and random underlying processes, respectively. The mesoscale of this model makes it well-suited for (a) the elucidation of glia-related hypotheses in hyperexcitable conditions, (b) use as a testing platform for neuromodulation purposes, and (c) a hardware implementation for closed-loop neuromodulation.

Low-to-High Cross-Frequency Coupling in the Electrical Rhythms as Biomarker for Hyperexcitable Neuroglial Networks of the Brain

OBJECTIVE

One of the features used in the study of hyperexcitablility is high-frequency oscillations (HFOs, >80 Hz). HFOs have been reported in the electrical rhythms of the brain's neuroglial networks under physiological and pathological conditions. Cross-frequency coupling (CFC) of HFOs with low-frequency rhythms was used to identify pathologic HFOs in the epileptogenic zones of epileptic patients and as a biomarker for the severity of seizure-like events in genetically modified rodent models. We describe a model to replicate reported CFC features extracted from recorded local field potentials (LFPs) representing network properties.

METHODS

This study deals with a four-unit neuroglial cellular network model where each unit incorporates pyramidal cells, interneurons, and astrocytes. Three different pathways of hyperexcitability generation-Na - ATPase pump, glial potassium clearance, and potassium afterhyperpolarization channel-were used to generate LFPs. Changes in excitability, average spontaneous electrical discharge (SED) duration, and CFC were then measured and analyzed.

RESULTS

Each parameter caused an increase in network excitability and the consequent lengthening of the SED duration. Short SEDs showed CFC between HFOs and theta oscillations (4-8 Hz), but in longer SEDs the low frequency changed to the delta range (1-4 Hz).

CONCLUSION

Longer duration SEDs exhibit CFC features similar to those reported by our team.

SIGNIFICANCE

First, Identifying the exponential relationship between network excitability and SED durations; second, highlighting the importance of glia in hyperexcitability (as they relate to extracellular potassium); and third, elucidation of the biophysical basis for CFC coupling features.

Effects of astrocytic mechanisms on neuronal hyperexcitability

While originally astrocytes have been thought to only act as support to neurons, recent studies have implicated them in multiple active roles, including the ability to moderate or alter neuronal firing patterns and to possibly be involved in both the prevention and propagation of epileptic seizures. In this study we propose a new model to incorporate pyramidal cells and interneurons (a common neural circuit in CA3 hippocampal slices) as well as a model of astrocyte. As both potassium and calcium ions have been shown to potentially affect neuronal hyperexcitability, the astrocytic model has both mechanisms--the clearance of potassium through potassium channels (such as KIR, KDR and sodium-potassium pump), and the influence of astrocyte in the synapse (forming the tripartite synapse with calcium-glutamate interactions). Preliminary findings of the model results show that when potassium conductances in the astrocyte are decreased, it results in the accumulation of extracellular potassium, leading to both spontaneous discharges and depolarization block, while the alteration of normal calcium response in the astrocyte can lead to just hyperexcitable conditions without the depolarization block.

Balance of synaptic and electrotonic connections controls the excitability of networks in biophysical model of epilepsy

Recent studies have implicated astrocytes in multiple active roles in neuronal networks. In particular they have been shown to be able to moderate and alter neural firing patterns both in normal and epileptic conditions. In addition, it has been proposed that one of the roles of gap junctions between astrocytes, as well as neurons is in increasing synchronization of neuronal firing and potential epileptogenic effect. In this study we build upon a model of a network that incorporates both pyramidal cells and interneurons as well as astrocytes with potassium clearance mechanisms and basic calcium dynamics. We include electrotonic connections between cells to be able to separate the effects of synaptic connections and gap junctions on neuronal hyperexcitability. Preliminary findings of this model show that under normal conditions, when gap junctions are blocked the network exists in an interictal-like state. When the system is put in a zero calcium environment (i.e. synaptic connections are disabled), the network enters spontaneous rhythmic bursting with very regular spiking. This suggests that electrotonic connections play a crucial role in the epileptogenesis within the neuronal network.