Pyramidal graphs resulting from statistical analyses of EEG recordings can improve our understanding of epileptic seizures.
A statistical approach squeezes more detailed information out of a current method of measuring brain signals in epileptic seizures, adding new insight into how these signals originate and spread.
Visual inspection of electroencephalography (EEG) recordings of epilepsy patients before and during a seizure is a fairly effective method for detecting the part of the brain that can benefit from surgical treatment.
But it’s not sufficient for more challenging cases.
Now, an approach developed by KAUST biostatistician Hernando Ombao and colleagues Yuan Wang of the University of Wisconsin–Madison and Moo K. Chung of the University of California, Irvine, digs deeper into the features of an EEG and can detect abnormalities in brain regions even before a seizure takes place.
The method could provide a clinically useful tool for seizure localization, according to Chung and Wang.
The approach stems from a field of mathematics that analyzes large and complex datasets by studying shape representations of the data and its interactions.
Analyzing these shapes provides information on patterns that exist within the data.
The team applied their method, known as a topological data analysis framework, to see what they could learn from an EEG recording conducted before and during an epileptic seizure.
The statistical approach removes noise from the EEG recording, providing cleaner signals.
A series of shapes are then drawn that directly relate to the signals in the recordings.
The final pyramidal shapes (persistence landscapes) that represent the signals coming from each electrode placed on the scalp provide a good picture of where the seizure originates in the brain and how it spreads.
The analysis of the patient’s EEG recording showed that the seizure originated from a region in or around the electrodes measuring signals from the left temporal lobe of the brain. It then spread to the right temporal lobe.
Further simulation studies showed that the test was robust and sensitive, even when the signal was buried under noise.
“Epileptologists should enhance their toolboxes of data analysis by adding methods like this one that capture topological features as part of their assessment of seizure foci in more challenging cases of epilepsy,” says Ombao.
The team next plans to test their framework on large samples of EEG recordings to clinically validate their findings.
Ombao is also developing statistical methods to study the impact of shocks to the brain, such as in epilepsy or stroke, on the communication network between brain regions and nerve cell populations.
Epilepsy is one of the most common neurological dysfunctions, affecting about 1% of the world population [1].
Recurrent epileptic seizures accompanied with their neurobiological and cognitive consequences impact patients with epilepsy negatively and consistently.
As a chronic disease with age-related features, epilepsy has a huge adverse effect on the quality of life (QoL) of patients, including cognitive impairment, decreased ability of daily activities and the possible social stigma.
To date, a fully understanding of the etiology of epilepsy is not yet available.
The onset of an epileptic seizure is usually accompanied by electrophysiological anomalies and/or behavioral manifestations.
The abnormal synchronized discharge of groups of neurons can happen to the whole brain or originates from several foci and propagates to the whole hemisphere and even the contralateral hemisphere.
The clinical manifestations of seizures vary from uncontrolled convulsions of limbs to impaired awareness and pain.
Anti-epilepsy drugs are commonly used in the treatment of epilepsy, but it is reported that there are about 25% of patients with epilepsy have drug-resistant epilepsy [2].
Long term freedom from seizures was observed in patients with drug-resistant epilepsy after having the epileptogenic zone resected by surgery with the help of neuroimaging techniques and intracranial EEG to identify the location of it.
However, for temporal lobe epilepsy (TLE), one-third of the patients still suffer from recurrent seizures and other complications after surgery.
This may be due to poor localization of the epileptogenic zone.
Observations and recordation on epilepsy date back to almost 2000 years B.C [3].
The first international scheme on the classification of epileptic seizures was proposed by the International League Against Epilepsy (ILAE) in 1964 [4].
Epileptic seizures are classified into one of the five main groups by different electroencephalographic expressions during ictal and inter-ictal period, wherein the anatomical and etiological aspects of different seizure types are also discussed.
In 1969, a revised version [5] of the classification of epileptic seizures proposed in 1964 was given to reflect the new knowledge and the contradictions between experts in this field.
The term ‘ictal electroencephalographic expression’ used in the scheme of classification in 1964 was replaced by ‘electroencephalographic seizure type’, emphasizing the equal importance of clinical and electroencephalographic manifestations in characterizing different types of seizure.
Subtypes of seizures are defined in more detail by clinical and electroencephalographic manifestations, anatomical and etiologic factors and onset age. At the same time, an international classification of the epilepsies [6] was published.
Primary generalized epilepsies, secondary generalized epilepsies and undetermined generalized epilepsies are suggested for diagnostic use with their clinical and ictal and inter-ictal electroencephalographic criteria. Since then, the terminologies of epilepsies and epileptic seizures in these documents had been adopted by numerous neurologists and clinicians worldwide.
In 1981, another revision about the classification of epileptic seizures [7] was launched by ILAE. Different from the two previous schemes, anatomical substrate, etiology and age factors were no longer retained, and epileptic seizures are differentiated via clinical and electroencephalographic (ictal and inter-ictal) manifestations in parallel.
To supplement the former classification schemes whereby a strong emphasis is placed on the separation of individual seizure types, and to refine the use of ‘epilepsies’ as an insufficiently rigorous implication of ‘diseases’, (International Classification on Epilepsies and Epileptic Syndromes) was presented in 1985 [8] with a revised version followed in 1989 [9].
The profound contribution of these proposals is the terminology used for the classification of ‘epileptic syndrome’. Epileptic syndrome is ‘an epileptic disorder characterized by a cluster of signs and symptoms customarily occurring together’ [9].
Different from a ‘disease’, a ‘syndrome’ could be determined without any confirmed etiology or prognosis. Since only few ‘diseases’ are established until then, ‘epileptic syndrome’ maybe a better choice for denoting the diagnostic entity in clinical practice.
The epilepsies are divided into ‘generalized epilepsies and syndromes’, ‘localization-related partial or focal epilepsies and syndromes’, ‘underdetermined’ or ‘special syndromes’.
For each group, syndromes are categorized by more detailed descriptions on etiology, seizure type and onset age.
The terms ‘idiopathic’ and ‘cryptogenic’ are used, for almost all intentions and purposes, as an alternative of ‘primary’ while circumvent possible misunderstandings; ‘symptomatic’ is used for epilepsies with confirmed pathogenesis.
Currently, the most commonly used definitions of epilepsy and epileptic seizure are those proposed by ILAE in 2005 [10] where an epileptic seizure is ‘a transient occurrence of signs and symptoms due to abnormal excessive or synchronous neuronal activity in the brain’ and a conceptual definition of epilepsy is characterized by ‘enduring predisposition to generate seizures’ and its far-ranging negative effects.
It was suggested for the first time that the behavioral disturbances together with the influence from social surroundings caused by recurrent seizures should also be recognized as part of epileptic conditions, to constitute a more comprehensive definition.
After that, continuous improvements and modifications [11,12,13] were performed. Figure 1 shows a sketch of the diagnosis of epilepsy or epileptic syndrome recommended by ILAE.
Five etiological groups (structural, genetic, infectious, metabolic and immune) are proposed [13] as a shared language to facilitate the communications between neurologists and clinicians with the unknown etiology group intending to cover the patients for whom the cause of epilepsy cannot be identified hitherto.
In clinical practice, the exclusion of non-epileptic seizures together with the classification of epileptic seizures is the first step in the process of the diagnostics and treatment of epilepsy or epileptic syndrome.
The identification of the onset and termination of the ictal period is often referred to as epileptic seizure detection or seizure detection.
The durations of individual ictal periods, the frequency and intensity of multiple seizures among a specific period and the influence caused by medical interventions can be achieved by retrospective analysis on medical history.
Until now, video- electroencephalograms (EEGs) still serve as the gold standard for epileptic seizure detection.
They provide both electrophysiological and behavioral information.
Epileptologists can inspect the multichannel EEG recordings and the synchronous video recording to identify the indicative electrophysiological and behavioral features from the background activity.
During a seizure, the occurrence of visually discernible changes in the EEG and the first clinical manifestation may not happen simultaneously, and the evolution of a seizure which originates from several foci and propagates to a larger field like the entire hemisphere or even both hemispheres can also be reflected by multichannel recordings.
For subclinical seizures, behavioral anomalies might not exist and the onset of some specific types of seizures does not have associated EEG reflections.
That implies a fully automated seizure detection method robust to all kinds of seizures is not desirable. However, the automatic extraction and labeling of characteristic features of EEG signals is confirmed to be able to accelerate the inspection of EEG recording and identify specific types of seizures.
In addition, EEG has good time resolution but fairly limited spatial resolution. On the other hand, well-developed neuroimaging technologies, such as positron emission tomography (PET), single photon emission computed tomography (SPECT), magnetic resonance imaging (MRI), functional magnetic resonance imaging (fMRI) have been widely adopted in the localization of epileptogenic zone, as for the diagnosis of epilepsy [14].
It should be noted that in recent years, several ultrafast neuroimaging techniques have been prototyped and proposed [15,16].
It could be more practical to picture the brain activities without having to make a trade-off between time resolution and spatial resolution.
Except for the localization of epileptogenic zone which may cause recurrent seizures, the structural and functional evolution of the brain networks of patients with epilepsy which may be caused by chronic epilepsy can also be reflected by neuroimaging techniques.
For patients with epilepsy, they are more likely to drown because of seizures’ onset during bathing, and Sudden Unexpected Death in Epilepsy (SUEPD) is the biggest epilepsy-related risk factor [1].
In particular, for newborns, epilepsy causes damage to the brain function development, especially for preterm infants. However, most of the patients (about 80%) with epilepsy live in developing countries and do not have access to state-of-art diagnosis, and even in the industrialized country, only a few can get the proper treatment they need.
Furthermore, manual inspection the video-EEG of patients places a heavy burden on clinical staff. Long-term continuous video-EEG monitoring prevents the patients from normal daily activities.
For the aforementioned reasons, automatic seizure detection and prediction based on EEG has been approached since the 1970s [17,18].
However, until now long-term EEG monitoring was still limited to highly structured environments, as the potential value of wearable devices capable of continuously and unobtrusively monitoring the related physiological and behavioral signals (ECG, EMG, EDG and motion signal, etc.) was explored [19,20].
For such seizure detection and prediction systems, pattern recognition and machine learning algorithms are adopted to ‘classify’ an epoch of signal into different classes [21].
Some of the classes indicate the onset of seizure or ictal period of a seizure, and an alarm is supposed to be triggered when a specific epoch of signal is classified into such classes, so that the clinician will verify whether a seizure has happened and take necessary medical intervention timely.
Most of such researches belong to so-called supervised learning paradigm. Regardless of the modalities of the input signals, ‘labeled data’ with clinicians’ knowledge encoded in are used to ‘train’ such a classifier. The unsupervised counterpart is not very popular in this field yet.
Machine learning-based automatic seizure monitoring is promising yet with intrinsic limitations.
machine learning algorithm learns rules from data automatically and evolves as more data is fed in, so it is a straightforward choice suitable for automatic seizure monitoring. Such algorithms embeded in medical instruments and mobile devices have the potential to reduce the misdetection of seizures and improve the overall efficiency of patients’ medical care.
Different from other fields like computer vision, high-quality data are limited in this field because only experts with domain knowledge are capable of labeling bio-signals having high inter- and intra-variability.
Inter-observer variability between experts may be considerable. Furthermore, using only information in bio-signals is deemed insufficient for seizure detection and the diagnosis of epilepsy syndrome.
For some specific types of seizures, the associated change in bio-signals could not be perceivable. On the contrary, epileptiform discharges are not always recognized as a seizure.
The ‘rules’ learned by machine learning algorithms are often only statistically significant relations with poor interpretability, rather than confirmed medical knowledge. Automated decision making will also cause ethical issues and there is a regulation bill requiring the right to an explanation and an option of not being subject to such an automatic decision making procedure has come into force recently (EU GDPR) [22].
It should be noted that although it is self-evident that epileptic seizures can be detected. There have been doubts about whether the onset of a seizure can be anticipated [23].
In recent years, many research groups [20,24] have reported their work on seizure predictions.
In 2014, the Mayo Clinic and the University of Pennsylvania launched two competitions looking for robust seizure detection and prediction algorithms and released datasets comprised of electrocorticograms (ECoGs) collected from the cortex of canines and humans.
Using SVM, random forest and other machine learning techniques, the participants achieved high sensitivity and low false alarm rate on these high-quality datasets [24,25].
A robust automatic seizure/prediction detection algorithm may aid efforts to a closed-loop warning/treatment system.
Although there are still no such algorithms that are widely accepted and acknowledged, an implantable closed-loop treatment system aiming for medically intractable refractory partial epilepsy had been approved by the U.S. FDA in 2013.
The system monitors and analyses the intracranial electroencephalographic activities of the patient.
Neuro stimulus therapy is applied when the approaching of an epileptic seizure is forecasted by the algorithm to terminate the onset and propagation of it. Despite the fact that significant reduction of seizures is observed, complications in patients received implant surgery are also reported [26].
Data analytics of the physiological signal and neuro-images acquired by different diagnostic techniques could provide opportunities for a deeper understanding of the underlying mechanism of epilepsy and enrich the state-of-the-art medical infrastructure.
Features extracted from physiological signals which can solve selectivity-invariance dilemma, which means the features can discriminate data segments recorded during ictal period from that recorded during inter-ictal period while the intra- and inter-individual variability does not lead to misclassification, is also crucial for a closed-loop warning/treatment system with high sensitivity and minimum false alarm rate.
More information: Yuan Wang et al. Topological data analysis of single-trial electroencephalographic signals, The Annals of Applied Statistics(2018). DOI: 10.1214/17-AOAS1119
Journal information: Annals of Applied Statistics
Provided by King Abdullah University of Science and Technology
