Quantitative approaches to guide epilepsy surgery from intracranial EEG

Over the past 10 years, the drive to improve outcomes from epilepsy surgery has stimulated widespread interest in methods to quantitatively guide epilepsy surgery from intracranial EEG (iEEG). Many patients fail to achieve seizure freedom, in part due to the challenges in subjective iEEG interpretation. To address this clinical need, quantitative iEEG analytics have been developed using a variety of approaches, spanning studies of seizures, interictal periods, and their transitions, and encompass a range of techniques including electrographic signal analysis, dynamical systems modeling, machine learning and graph theory. Unfortunately, many methods fail to generalize to new data and are sensitive to differences in pathology and electrode placement. Here, we critically review selected literature on computational methods of identifying the epileptogenic zone from iEEG. We highlight shared methodological challenges common to many studies in this field and propose ways that they can be addressed. One fundamental common pitfall is a lack of open-source, high-quality data, which we specifically address by sharing a centralized high-quality, well-annotated, multicentre dataset consisting of >100 patients to support larger and more rigorous studies. Ultimately, we provide a road map to help these tools reach clinical trials and hope to improve the lives of future patients.

Source-sink connectivity: a novel interictal EEG marker for seizure localization

Over 15 million epilepsy patients worldwide have drug-resistant epilepsy. Successful surgery is a standard of care treatment but can only be achieved through complete resection or disconnection of the epileptogenic zone, the brain region(s) where seizures originate. Surgical success rates vary between 20% and 80%, because no clinically validated biological markers of the epileptogenic zone exist. Localizing the epileptogenic zone is a costly and time-consuming process, which often requires days to weeks of intracranial EEG (iEEG) monitoring. Clinicians visually inspect iEEG data to identify abnormal activity on individual channels occurring immediately before seizures or spikes that occur interictally (i.e. between seizures). In the end, the clinical standard mainly relies on a small proportion of the iEEG data captured to assist in epileptogenic zone localization (minutes of seizure data versus days of recordings), missing opportunities to leverage these largely ignored interictal data to better diagnose and treat patients. IEEG offers a unique opportunity to observe epileptic cortical network dynamics but waiting for seizures increases patient risks associated with invasive monitoring. In this study, we aimed to leverage interictal iEEG data by developing a new network-based interictal iEEG marker of the epileptogenic zone. We hypothesized that when a patient is not clinically seizing, it is because the epileptogenic zone is inhibited by other regions. We developed an algorithm that identifies two groups of nodes from the interictal iEEG network: those that are continuously inhibiting a set of neighbouring nodes ('sources') and the inhibited nodes themselves ('sinks'). Specifically, patient-specific dynamical network models were estimated from minutes of iEEG and their connectivity properties revealed top sources and sinks in the network, with each node being quantified by source-sink metrics. We validated the algorithm in a retrospective analysis of 65 patients. The source-sink metrics identified epileptogenic regions with 73% accuracy and clinicians agreed with the algorithm in 93% of seizure-free patients. The algorithm was further validated by using the metrics of the annotated epileptogenic zone to predict surgical outcomes. The source-sink metrics predicted outcomes with an accuracy of 79% compared to an accuracy of 43% for clinicians' predictions (surgical success rate of this dataset). In failed outcomes, we identified brain regions with high metrics that were untreated. When compared with high frequency oscillations, the most commonly proposed interictal iEEG feature for epileptogenic zone localization, source-sink metrics outperformed in predictive power (by a factor of 1.2), suggesting they may be an interictal iEEG fingerprint of the epileptogenic zone.

Sources and Sinks in Interictal iEEG Networks: An iEEG Marker of the Epileptogenic Zone

Around 30% of epilepsy patients have seizures that cannot be controlled with medication. The most effective treatments for medically resistant epilepsy are interventions that surgically remove the epileptogenic zone (EZ), the regions of the brain that initiate seizure activity. A precise identification of the EZ is essential for surgical success but unfortunately, current success rates range from 20-80%. Localization of the EZ requires visual inspection of intracranial EEG (iEEG) recordings during seizure events. The need for seizure occurrence makes the process both costly and time-consuming and in the end, less than 1% of the data captured is used to assist in EZ localization. In this study, we aim to leverage interictal (between seizures) data to localize the EZ. We develop and test the source-sink index as an interictal iEEG marker by identifying two groups of network nodes from a patient's interictal iEEG network: those that inhibit a set of their neighboring nodes ("sources") and the inhibited nodes themselves ("sinks"). Specifically, we i) estimate patient-specific dynamical network models from interictal iEEG data and ii) compute a source-sink index for every network node (iEEG channel) to identify pathological nodes that correspond to the EZ. Our results suggest that in patients with successful surgical outcomes, the source-sink index clearly separates the clinically identified EZ (CA-EZ) channels from other channels whereas in patients with failed outcomes CA-EZ channels cannot be distinguished from the rest of the network.

Transfer Entropy between Intracranial EEG Nodes Highlights Network Dynamics that Cause and Stop Epileptic Seizures

Transfer entropy (TE) is used to examine the connectivity between nodes and the roles of nodes in epileptic neural networks during rest, moments before seizure, during seizure, and moments after seizure. There is a set of nodes that dominate information flow to epileptogenic zone (EZ) nodes, regions that trigger seizure, and non-EZ nodes during rest. The TE from the dominant to the EZ nodes decreases shortly before a seizure event and reaches a minimum during seizure. During the seizure, the dominant nodes cease or only weakly interact with the EZ nodes. This supports the hypothesis that seizure occurs when some nodes stop inhibiting the EZ nodes. The TE from the dominant to the EZ nodes peaks immediately after seizure, suggesting that seizure may stop when the brain exerts the highest level of information flow/activation/communication to the EZ nodes. The information flow from the dominant to EZ nodes is different from that to non-EZ nodes. This TE dynamics entering and exiting seizures may identify more accurately the EZ nodes, which may improve surgical planning.

Identifying Sleep Biomarkers to Evaluate Cognition in HIV

Sleep disturbance and cognitive impairment represent two of the most common and debilitating conditions facing seropositive (HIV+) individuals who are otherwise well controlled with antiretroviral therapy. Sleep-assessment-based biomarkers represent an important step towards improving our understanding of the unique mechanistic features that may link sleep disruption and cognition in HIV+ individuals, ultimately leading to advancements in treatment and management options. In this study, a risk score was computed via a generalized linear model (GLM), which optimally combines polysomnography (PSG) features extracted from EEG, EMG, and EOG signals, to distinguish 18 HIV+ Black male individuals with and without cognitive impairment. The optimal set of features was identified via the least absolute shrinkage and selection operator (LASSO) approach, and the risk separation between the two groups, i.e., cognitively normal and cognitive impaired, was significant (and has a P-value < .001). The optimal set of predictive features were all EEG derived and sleep stage-specific. These preliminary findings suggest that sleep-based EEG features may be used as both diagnostic and prognostic biomarkers for cognition in HIV+ subjects.

Estimating Intracranial EEG Signals at Missing Electrodes in Epileptic Networks

Epilepsy can be controlled by targeted treatment of the epileptogenic zone (EZ), the region in the brain where seizures originate. Identification of the EZ often requires visual inspection of invasive EEG recordings and thus relies heavily on placement of electrodes, such that they cover the EZ. A dense brain coverage would be ideal to obtain accurate boundaries of the EZ but is not possible due to surgical limitations. This gives rise to the "missing electrode problem", where clinicians desire to know what neural activity looks like between implanted electrodes. In this paper, we compare two methods for time series estimation of missing stereotactic EEG (SEEG) recordings. Specifically, we represent SEEG data as a sequence of Linear Time-Invariant (LTI) models. We then remove one signal from the data set and apply two different algorithms to simultaneously estimate the LTI models and the "missing" signal: (i) a Reduced-Order Observer in combination with Least Squares Estimation and (ii) an Expectation Maximization based Kalman Filter. The performance of each approach is evaluated in terms of (i) estimation error, (ii) sensitivity to initial conditions, and (iii) algorithm run-time. We found that the EM approach has smaller estimation errors and is less sensitive to initial conditions. However, the reduced-order observer has a run-time that is orders of magnitude faster.

A novel sleep stage scoring system: Combining expert-based features with the generalized linear model

In this study, we aim to automate the sleep stage scoring process of overnight polysomnography (PSG) data while adhering to expert-based rules. We developed a sleep stage scoring algorithm utilizing the generalized linear modelling (GLM) framework and extracted features from electroencephalogram (EEG), electromyography (EMG) and electrooculogram (EOG) signals based on predefined rules of the American Academy of Sleep Medicine (AASM) Manual for Scoring Sleep. Specifically, features were computed in 30-s epochs in the time and frequency domains of the signals and were then used to model the probability of an epoch being in each of five sleep stages: N3, N2, N1, REM or Wake. Finally, each epoch was assigned to a sleep stage based on model predictions. The algorithm was trained and tested on PSG data from 38 healthy individuals with no reported sleep disturbances. The overall scoring accuracy reached on the test set was 81.50 ± 1.14% (Cohen's kappa, κ = 0.73 ± 0.02 ). The test set results were highly comparable to the training set, indicating robustness of the algorithm. Furthermore, our algorithm was compared to three well-known commercialized sleep-staging tools and achieved higher accuracies than all of them. Our results suggest that automatic classification is highly consistent with visual scoring. We conclude that our algorithm can reproduce the judgement of a scoring expert and is also highly interpretable. This tool can assist visual scorers to speed up their process (from hours to minutes) and provides a method for a more robust, quantitative, reproducible and cost-effective PSG evaluation, supporting assessment of sleep and sleep disorders.

A Novel Sleep Stage Scoring System: Combining Expert-Based Rules with a Decision Tree Classifier

Overnight polysomnography (PSG) is the gold standard tool used to characterize sleep and for diagnosing sleep disorders. PSG is a non-invasive procedure that collects various physiological data which is then scored by sleep specialists who assign a sleep stage to every 30-second window of the data according to predefined scoring rules. In this study, we aimed to automate the process of sleep stage scoring of overnight PSG data while adhering to expert-based rules. We developed an algorithm utilizing a likelihood ratio decision tree classifier and extracted features from EEG, EMG and EOG signals based on predefined rules of the American Academy of Sleep Medicine Manual. Specifically, features were computed in 30-second epochs in the time and the frequency domains of the signals and used as inputs to the classifier which assigned each epoch to one of five possible stages: N3, N2, N1, REM or Wake. The algorithm was trained and tested on PSG data from 38 healthy individuals with no reported sleep disturbances. The overall scoring accuracy was 80.70% on the test set, which was comparable to the training set. Our results imply that the automatic classification is highly robust, fast, consistent with visual scoring and is highly interpretable.

Estimating unmeasured invasive EEG signals using a reduced-order observer

Epilepsy affects around 50 million people worldwide. Over 30% of patients are drug-resistant where the only treatment may be surgical resection of the epileptogenic zone (EZ), the region of the brain that generates seizures. Identification of the EZ is often based on invasive EEG recordings. As such, surgical outcome relies heavily on precise and dense placement of EEG electrodes into the brain. Despite large brain regions being removed, success rates barely reach 65%. This gives rise to the "missing electrode problem", where clinicians want to know what neural activity looks like between sparsely implanted electrodes. Solving this problem will enable more accurate localization of the EZ. In this paper, we demonstrate the first steps towards developing a computational platform to estimate neural activity at the "missing electrodes" using a reduced-order observer from control theory. Specifically, we constructed a sequence of discrete time Linear Time-Invariant (LTI) models using the available EEG data from two epilepsy patients. Then, we used the models to simulate EEG data and remove selected signals ("missing" states) from the simulated data set. Finally, we used a reduced-order observer to estimate the signals of these "missing" states and evaluated performance by comparing the observer estimates to the simulated EEG time series.

Linear time-varying model characterizes invasive EEG signals generated from complex epileptic networks

Electrocorticography (ECoG) and stereotactic electroencephalography (SEEG) are popular tools for studying neural mechanisms governing behavior and neural disorders, such as epilepsy. In particular, clinicians are interested in identifying brain regions that start seizures, i.e., the epileptogenic zone (EZ) from such invasive recordings. Currently, they visually inspect signals from each electrode to locate abnormal activity, and are not informed by predictive models that can characterize such recordings and potentially increase accuracy in localizing the EZ. In this paper, we test whether a simple linear time varying (LTV) model is sufficient to characterize both ECoG and SEEG activity. Specifically, we construct linear time invariant models in consecutive time windows before, during and after seizure events creating an LTV model from data collected in one ECoG and one SEEG patient. We find that these LTV models accurately reconstruct both ECoG and SEEG time series measured suggesting that these LTV models can be used for EZ localization.

A look at the strength of micro and macro EEG analysis for distinguishing insomnia within an HIV cohort

HIV patients are often plagued by sleep disorders and suffer from sleep deprivation. However, there remains a wide gap in our understanding of the relationship between HIV status, poor sleep, overall function and future outcomes; particularly in the case of HIV patients otherwise well controlled on cART (combined anti-retroviral therapy). In this study, we compared two groups: 16 non-HIV subjects (seronegative controls) and 12 seropositive HIV patients with undetectable viral loads. We looked at sleep behavioral (macro-sleep) features and sleep spectral (micro-sleep) features obtained from human-scored overnight EEG recordings to study whether the scored EEG data can be used to distinguish between controls and HIV subjects. Specifically, the macro-sleep features were defined by sleep stages and included sleep transitions, percentage of time spent in each sleep stage, and duration of time spent in each sleep stage. The micro-sleep features were obtained from the power spectrum of the EEG signals by computing the total power across all channels and frequencies, as well as the average power in each sleep stage and across different frequency bands. While the macro features do not distinguish between the two groups, there is a significant difference and a high classification accuracy for the scoring-independent micro features. This spectral separation is interesting because evidence suggests a relationship between sleep complaints and cognitive dysfunction in HIV patients stable on cART. Furthermore, there are currently no biomarkers that predict the early development of cognitive decline in HIV patients. Thus, a micro-sleep architectural approach could serve as a biomarker to identify HIV patients vulnerable to cognitive decline, providing an avenue to explore the utility of early intervention.

To Score or Not to Score? A look at the distinguishing power of micro EEG analysis on an annotated sample of PSG studies conducted in an HIV cohort

In this study, we used the Pittsburgh Sleep Quality Index to divide the subjects into two groups, good sleepers and bad sleepers. We computed sleep behavioral (macro-sleep architectural) features and sleep spectral (micro-sleep architectural) features in order to observe if the annotated EEG data can be used to distinguish between good and bad sleepers in a more quantitative manner. Specifically, the macro-sleep features were defined by sleep stages and included sleep transitions, percentage of time spent in each sleep stage, and duration of time spent in each sleep stage. The micro-sleep features were obtained from the power spectrum of the EEG signals by computing the total power across all channels and all frequencies, as well as the average power in each sleep stage and across different frequency bands. We found that while the scoring-independent micro features are significantly different between the two groups, the macro features are not able to significantly distinguish the two groups. The fact that the macro features computed from the scoring files cannot pick up the expected difference in the EEG signals raises the question as to whether human scoring of EEG signals is practical in assessing sleep quality.