2021 West - 15500594211008285

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Neurology/Medicine

Clinical EEG and Neuroscience


2022, Vol. 53(1) 74–78
A Review of Network and Computer Analysis © EEG and Clinical Neuroscience
Society (ECNS) 2021
of Epileptiform Discharge Free EEG to Article reuse guidelines:
sagepub.com/journals-permissions

Characterize and Detect Epilepsy DOI: 10.1177/15500594211008285


journals.sagepub.com/home/eeg

Caitlin West1, Wessel Woldman2, Katy Oak3, Brendan McLean3 ,


and Rohit Shankar4,5

Abstract
Objectives. There is emerging evidence that network/computer analysis of epileptiform discharge free electroencephalograms
(EEGs) can be used to detect epilepsy, improve diagnosis and resource use. Such methods are automated and can be performed
on shorter recordings of EEG. We assess the evidence and its strength in the area of seizure detection from network/computer
analysis of epileptiform discharge free EEG. Methods. A scoping review using Preferred Reporting Items for Systematic Reviews
and Meta-Analyses (PRISMA) guidance was conducted with a literature search of Embase, Medline and PsychINFO. Predesigned
inclusion/exclusion criteria were applied to selected articles. Results. The initial search found 3398 articles. After duplicate
removal and screening, 591 abstracts were reviewed, 64 articles were selected and read leading to 20 articles meeting the requi-
site inclusion/exclusion criteria. These were 9 reports and 2 cross-sectional studies using network analysis to compare and/or
classify EEG. One review of 17 reports and 10 cross-sectional studies only aimed to classify the EEGs. One cross-sectional study
discussed EEG abnormalities associated with autism. Conclusions. Epileptiform discharge free EEG features derived from network/
computer analysis differ significantly between people with and without epilepsy. Diagnostic algorithms report high accuracies and
could be clinically useful. There is a lack of such research within the intellectual disability (ID) and/or autism populations, where
epilepsy is more prevalent and there are additional diagnostic challenges.

Keywords
network analysis, computer analysis, diagnostic algorithms, artificial intelligence

Received December 7, 2019; revised August 18, 2020; accepted March 7, 2021.

Introduction a relatively high specificity (78%-98%) but low sensitivity


(25%-56%) for the diagnosis of epilepsy.1
Epilepsy and Electroencephalography
Electroencephalography (EEG) records brain activity and is
used to support the diagnosis of epilepsy, classify epileptic syn-
Network and Computer EEG Analysis
dromes, help guide pharmacological management and in pre- Automated computerized methods, which include network
surgical evaluation.1 The National Institute for Health and analysis and machine learning, are being used to classify
Care Excellence (NICE) recommends that a standard EEG EEG from shorter recordings and do not require the pres-
recording is performed after 1 or 2 seizures if the clinical ence of visual epileptiform discharges.3 There are machine-
history suggests that the events could have an epileptic learning approaches, which aim to automatically identify
basis.2 A standard EEG recording usually takes 20–40 min.
The use of electroencephalograms (EEGs) in epilepsy was 1
Exeter Medical School, Knowledge Spa, Truro, UK
reviewed by Smith.1 The likelihood of capturing a seizure on 2
Centre for Systems Modelling and Quantitative Biomedicine, University of
a standard EEG is low, and enhancing stimulus such as hyper- Birmingham, Birmingham, UK
3
ventilation and photostimulation may be used to trigger epilep- Royal Cornwall Hospitals Trust Truro, Truro, UK
4
Cornwall Partnership NHS Foundation Trust, Truro, UK
tiform activity.1 Recordings are reviewed for the presence of 5
Cornwall Intellectual Disability Equitable Research (CIDER), University of
epileptiform activity; most commonly they contain interictal Plymouth Medical School, Truro, UK
epileptiform discharges (IEDs), which are present on the first
Corresponding Author:
EEG recording in up to 65% of people subsequently diagnosed Rohit Shankar, Chygovenck, Threemilestone Industrial Estate, Truro Cornwall
with epilepsy. Repeat, sleep-deprived and prolonged EEG TR4 9LD, UK.
recordings increase sensitivity to 80% to 90%.1 An EEG has Email: [email protected]
West et al 75

and classify spikes, discharges, or other IEDs from long- methods, and the differences in results, particularly between
term EEG.4-6 papers analyzing the same network feature, are notable findings.
Of these studies, a systematic review and meta-analysis by van
Diessen et al,25 which includes results from 13 articles, provides
Aim particularly strong evidence regarding 2 of the network features
The main aim of this study was to assess the evidence and its assessed. However, these results are limited by differences in
strength in the area of epilepsy detection from network analysis, methodology across the included studies.
and other computer analysis methods, of interictal EEG.

Summary of Review Findings


Methods
This was a scoping review conducted using PRISMA guidance.
Network Analysis
A search of Embase, Medline and PsychInfo was performed In network analysis, the electrodes of the EEG are modeled as
using Healthcare Databases Advanced Search with terms relating nodes of a network.27 A connection between 2 nodes is typically
to computer/network analysis, EEG, and epilepsy. The full list of quantified by comparing the relationship between the signals
terms used is provided in Supplemental Appendix A. Results from 2 electrodes. After a network is derived from the electro-
were limited to full reports published in English from January physiological recording, graph measures can be used to capture
2008 to December 2018 so only recent material was included. properties of the network structures, which can then, in turn, be
No other sources were used. used as potential markers in comparing patients and controls.28
Screening was separately performed by 2 coauthors. After Connectivity, mean degree, path length, and clustering coefficient
the initial screening of titles for relevant articles, abstracts are the network features used most commonly across the included
were screened using predefined eligibility criteria by both and studies. Nine individual studies found significant differences
differences in selection (n = 1) was reconciled by consensus. between the interictal EEGs of people with epilepsy in the fea-
tures included in Supplemental Appendix D.
A study found no significant differences in any network
Criteria
feature between patients and controls, but created a successful
Studies were included if they used resting-state/interictal EEG classifier through a combination of different network features.23
and excluded if they studied exclusively preictal/postictal A later report by the same group also found no significant dif-
and/or ictal EEG. Studies were excluded if they used functional ferences in path length and clustering coefficient, but a different
magnetic resonance imaging (fMRI) or an investigation other network measure (minimum spanning tree) was significantly
than EEG for analysis and if they used the presence of IEDs different between patients and controls.24
to classify the EEG. Papers were included if they reviewed or The results for individual network features are commonly
included primary data. presented as a pattern of increased/decreased feature values,
dependent on anatomical location and frequency of the EEG,
rather than as an overall value. An example of this is connectiv-
Results ity, the feature that is most commonly investigated and this was
The initial search identified 3398 articles which when examined the case in 7 of the included studies. As shown in Supplemental
by title for relevance, and duplication resulted in 591 abstracts Table 2, results vary widely. Connectivity is calculated and rep-
being screened. Sixty-four papers were read in full and the eli- resented in a variety of ways, which include synchronization,
gibility criteria applied. A flowchart of the search process is coherence, and outflow. One study investigated the EEGs of
provided in Supplemental Appendix B. people with temporal lobe epilepsy (TLE) and found that
Twenty articles have been included and are detailed in they had decreased functional connectivity in the default-mode
Supplemental Appendix C.7-26 Thematic analysis was used to network regions, which are reported to be involved in TLE.14
analyze these papers looking into the network analysis, classi- Whereas 1 study reported increased synchronization likelihood
fication, and influence on special populations. Nine reports in people with epilepsy.11
and 2 cross-sectional studies use network analysis to compare Generally, individual papers report significant differences in
and/or classify EEG. The results from these papers are com- the network features of patients and controls, but results across
pared by network feature in Supplemental Appendix D. One the 11 reports are varied. This may well be due to significant meth-
review of 17 controlled trials and a further 10 cross-sectional odological heterogeneity, discussed under ‘General limitations.’
reports, including 2 which use network analysis, aim to classify The included studies suggest that network analysis can esti-
the EEG. One cross-sectional study also discusses EEG abnor- mate differences between interictal EEG from people with epi-
malities associated with a special population of autism. lepsy and from healthy controls. There are, however, no studies
A review by Acharya et al7 includes promising results from 17 that classify EEG using network analysis, and there is little con-
articles and is particularly significant to this review. Across the sensus on the ‘best’ way to apply network analysis with regard
studies using network analysis, the differences between analysis to feature calculation and selection.
76 Clinical EEG and Neuroscience 53(1)

Classification The method used to record the EEG also differed between
studies, and was performed with varying numbers of electrodes.
Twelve reports involve the classification of EEG. This The University of Bonn data set was produced from intracranial
includes 1 review of 17 articles and reports of 10 cross-sectional EEG, which limits the applicability of the results to standard
studies. The review7 and a further 6 cross-sectional EEG.
studies10,12,14,16,20,23 highlighted that interictal and healthy EEG
could be classified with over 90% accuracy. This level of accu-
racy suggests that computer analysis of interictal EEG could be Limitations of This Study
clinically useful. The classification studies use a wide variety This scoping review had a focus on network analysis, and it is
of methods to differentiate the healthy and interictal EEGs, beyond the scope of this review to provide a comparison of
which include nonlinear feature extraction and network analysis. other methods. Articles using fMRI for analysis were excluded
from this study, as EEG alone is more suitable for a standard
Studies in Special Populations diagnostic test (due to lower cost, etc). These studies may
provide further insight into measurable differences between
There is an increased prevalence of epilepsy in people with ID the brains of people with epilepsy and those without.
(15%-30%).29 Prevalence is also increased in people with
autism.30 There are diagnostic challenges within these popula-
tions as an EEG recording can be distressing for some people Conclusion
and there may be no cooperation. Nonepileptic events may be There are suggestions that network analysis can be used to dif-
misinterpreted which may lead to misdiagnosis.29 There are ferentiate interictal EEG in people with epilepsy from findings
also suggestions that people with ID and/or autism may have in healthy controls. Classification methods using network anal-
existing EEG abnormalities that are independent of epilepsy.31 ysis and other methods are highly accurate in trials. Generally,
Machine learning has been applied in the area of EEG analysis the results are limited by significant methodological variability,
for the diagnosis of autism. A recent study used a machine- and further research using standardized methodology and wider
learning method to distinguish EEG from children with EEG data sets is required to assess the reproducibility of results
autism and those with other neuropsychiatric diagnoses, such from individual studies. There is a lack of research within the
as attention deficit hyperactivity disorder, as well as children ID/autism populations, where such methods would arguably
developing typically, with an accuracy of 95%.32 be most useful, as the prevalence of epilepsy is increased
A search was performed to look for reports that compared within these populations and there are additional diagnostic
and/or classified interictal EEG from people with epilepsy challenges.
and EEG from people without epilepsy, in the ID and/or
autism populations. No reports were identified. Acknowledgments
The authors acknowledge Professor Sander JWL, University College
Limitations of the Studies Reviewed Hospitals, London, UK, for providing technical support in putting
together this paper.
Methodological differences between studies make comparisons
difficult. Epilepsy syndrome and/or seizure type of the popula- Author Contributions
tion considered vary across the items. Visually, epileptiform All authors were involved in developing, editing, writing and revising
activity differs between epilepsy types.1 Classification the paper. All authors satisfy the ICJME criteria for authorship.
methods are likely to need significant adaptation to different
epilepsy types. Differences may arise due to choice of specific Declaration of Conflicting Interests
network derivation method (linear, nonlinear, phase based or
The authors declared the following potential conflicts of interest with
amplitude based), choice of epochs and epoch length, referenc-
respect to the research, authorship, and/or publication of this article:
ing method, sampling rate, and frequency bands. CW, OK and BMcL have no conflicts of interest to report. WW has
This is particularly significant, as many of the classifiers use a commercial interest and is the scientific director in the company
machine learning, in which a classifier is ‘trained’ on a data set. Neuronostics (https://www.neuronostics.com/) focused on the devel-
Such classifiers are vulnerable to sensitization to an individual opment and commercialization of a novel and potentially revolutionary
data set, which would reduce the probability of successful clas- seizure susceptibility assessment technology called #BioEP™. RS is
sification when applied to new out-of-sample EEG recordings. the clinical chief investigator for the NIHR adopted study for investi-
Seventeen articles in 1 review and a further 3 cross-sectional gating the impact of #BioEP™. He has no commercial stake holding in
reports use the same EEG data set from the University of the company.
Bonn to analyze and/or test their classification method.
Training and/or testing on a wider EEG data set would ensure Funding
that classifiers are applicable to a wider variety of EEG data The authors received no financial support for the research, authorship
as would be seen in a clinical setting. and/or publication of this article.
West et al 77

ORCID iDs 14. Coito A, Genetti M, Iannotti GR, et al. Altered directed functional
connectivity in temporal lobe epilepsy in the absence of interictal
Brendan McLean https://orcid.org/0000-0003-1419-8540
spikes: a high density EEG study. Epilepsia. 2016;57(3):402-411.
Rohit Shankar https://orcid.org/0000-0002-1183-6933
doi:10.1111/epi.13308
15. Colominas MA, Jomaa MESH, Humeau-Heurtier A, Jrad N, Van
Supplemental Material Bogaert P. Time-varying time-frequency complexity measures for
epileptic EEG data analysis. IEEE Trans Biomed Eng. 2018;
Supplemental material for this article is available online. 65(8):1681-1688. doi:10.1109/TBME.2017.2761982
16. Goksu H. EEG Based epileptiform pattern recognition inside and
outside the seizure states. Biomed Signal Process Control.
References 2018;43:204-215. doi:10.3389/fninf.2018.00095
1. Smith S. EEG in the diagnosis, classification, and management of 17. Hussain L, Aziz W, Alowibdi JS, et al. Symbolic time series anal-
patients with epilepsy. J Neurol Neurosurg Psychiatry. 2005; ysis of electroencephalographic (EEG) epileptic seizure and brain
76(Suppl 2):ii2-ii7. doi:10.1136/jnnp.2005.069245 dynamics with eye-open and eye-closed subjects during resting
2. Overview|Epilepsies: diagnosis and management|Guidance|NICE. states. J Physiol Anthropol. 2017;36(1):21.
Nice.org.uk. https://www.nice.org.uk/guidance/cg137. Published 18. Pyrzowski J, Siemiń ski M, Sarnowska A, Jedrzejczak J, Nyka WM.
2019. Accessed November 2019. Interval analysis of interictal EEG: pathology of the alpha rhythm in
3. Kramer MA, Cash SS. Epilepsy as a disorder of cortical network focal epilepsy. Sci Rep. 2015;5:16230. doi:10.1038/srep16230
organization. Neuroscientist. 2012;18(4):360-372. doi:10.1177/ 19. Rotondi F, Franceschetti S, Avanzini G, Panzica F. Altered EEG
1073858411422754 resting-state effective connectivity in drug-naïve childhood
4. Antoniades A, Spyrou L, Martin-Lopez D, et al. Detection of absence epilepsy. Clin Neurophysiol. 2016;127(2):1130-1137.
interictal discharges with convolutional neural networks using dis- doi:10.1016/j.clinph.2015.09.003
crete ordered multichannel intracranial EEG. IEEE Trans Neural 20. Sairamya NJ, Thomas GS, Subathra MSP, Balakrishnan R. An
Syst Rehabil Eng. 2017;25(12):2285-2294. doi:10.1109/TNSRE. effective approach to classify epileptic EEG signal using local
2017.2755770 neighbor gradient pattern transformation methods. Australas
5. Antoniades A, Spyrou L, Took CC, Sanei S. Deep learning for epi- Phys Eng Sci Med. 2018;41(4):1029-1046. doi:10.1007/s13246-
leptic intracranial EEG data. 2016 IEEE 26th International Workshop 018-0697-9
on Machine Learning for Signal Processing (MLSP), 2016; 1-6. 21. Schmidt H, Woldman W, Goodfellow M, et al. A computational
6. Thomas J, Comoretto L, Jin J, Dauwels J, Cash SS, Westover MB. biomarker of idiopathic generalized epilepsy from resting state
EEG classification via convolutional neural network-based EEG. Epilepsia. 2016;57(10):e200. doi:10.1111/epi.13481
interictal epileptiform event detection. 2018 40th Annual 22. Raj KV, Rajagopalan SS, Bhardwaj S, et al. Machine learning
International Conference of the IEEE Engineering in Medicine detects EEG microstate alterations in patients living with temporal
and Biology Society (EMBC), Honolulu, HI, 2018; 3148-3151. lobe epilepsy. Seizure. 2018;61:8-13. doi:10.1016/j.seizure.2018.
7. Acharya U, Vinitha Sree S, Swapna G, Martis R, Suri J. 07.007
Automated EEG analysis of epilepsy: a review. Knowl Based 23. van Diessen E, Otte WM, Braun KPJ, Jansen FE, Stam CJ.
Syst. 2013;45:147-165. doi:10.1016/j.knosys.2013.02.014 Improved diagnosis in children with partial epilepsy using a mul-
8. Adebimpe A, Aarabi A, Bourel-Ponchel E, Mahmoudzadeh M, tivariable prediction model based on EEG network characteristics.
Wallois F. Functional brain dysfunction in patients with benign PLoS ONE. 2013;8(4). doi:10.1371/journal.pone.0059764
childhood epilepsy as revealed by graph theory. PLoS ONE. 24. van Diessen E, Otte WM, Stam CJ, Braun KPJ, Jansen FE.
2015;10(10):e0139228. doi:10.1371/journal.pone.0139228 Electroencephalography based functional networks in newly
9. Adebimpe A, Aarabi A, Bourel-Ponchel E, Mahmoudzadeh M, diagnosed childhood epilepsies. Clin Neurophysiol. 2016;
Wallois F. EEG resting state functional connectivity analysis in 127(6):2325-2332. doi:10.1016/j.clinph.2016.03.015
children with benign epilepsy with centrotemporal spikes. Front 25. van Diessen E, Zweiphenning WJEM, Jansen FE, Stam CJ, Braun
Neurosci. 2016;10:143. doi:10.3389/fnins.2016.00143 KPJ, Otte WM. Brain network organization in focal epilepsy: a
10. Bao FS, Gao JM, Hu J, Lie DY, Zhang Y, Oommen KJ. Automated systematic review and meta-analysis. PloS ONE. 2014;9(12):
epilepsy diagnosis using interictal scalp EEG. Conference proceed- e114606. doi:10.1371/journal.pone.0114606
ings: Annual International Conference of the IEEE Engineering in 26. Verhoeven T, Coito A, Plomp G, et al. Automated diagnosis of tem-
Medicine and Biology Society IEEE Engineering in Medicine and poral lobe epilepsy in the absence of interictal spikes. NeuroImage
Biology Society Conference, 2009;2009:6603-6607. Clinical. 2018;17:10-15. doi:10.1016/j.nicl.2017.09.021
11. Bartolomei F, Bettus G, Stam CJ, Guye M. Interictal network 27. Bassett DS, Bullmore ET. Human brain networks in health and
properties in mesial temporal lobe epilepsy: a graph theoretical disease. Curr Opin Neurol. 2009;22(4):340-347. doi:10.1097/
study from intracerebral recordings. Clin Neurophysiol. WCO.0b013e32832d93dd
2013;124(12):2345-2353. doi:10.1016/j.clinph.2013.06.003 28. van Straaten E, Stam C. Structure out of chaos: functional brain
12. Bosl WJ, Loddenkemper T, Nelson CA. Nonlinear EEG bio- network analysis with EEG, MEG, and functional MRI. Eur
marker profiles for autism and absence epilepsy. Neuropsychiatr Neuropsychopharmacol. 2013;23(1):7-18. doi:10.1016/j.euroneuro.
Electrophysiol. 2017;3(1). 2012.10.010
13. Clemens B, Puskás S, Bessenyei M, et al. EEG Functional con- 29. Chapman M, Iddon P, Atkinson K, et al. The misdiagnosis of
nectivity of the intrahemispheric cortico-cortical network of idio- epilepsy in people with intellectual disabilities: a systematic
pathic generalized epilepsy. Epilepsy Res. 2011;96(1–2):11-23. review. Seizure. 2011;20(2):101-106. doi:10.1016/j.seizure.
doi:10.1016/j.eplepsyres.2011.04.011 2010.10.030
78 Clinical EEG and Neuroscience 53(1)

30. Juneja M, Gupta S, Thakral A. Prevalence of unrecognized autism 32. Grossi E, Buscema M, Della Torre F, Swatzyna RJ. The
spectrum disorders in epilepsy: a clinic-based study. J Pediatr “MS-ROM/IFAST” model, a novel parallel nonlinear EEG analy-
Neurosci. 2018;13(3):308-312. doi:10.4103/JPN.JPN_136_17 sis technique, distinguishes ASD subjects from children affected
31. Buckley AW, Holmes GL. Epilepsy and autism. Cold Spring with other neuropsychiatric disorders with high degree of accu-
Harb Perspect Med. 2016;6(4):a022749. doi:10.1101/cshperspect. racy. Clin EEG Neurosci. 2019;50(5):319-331. doi:10.1177/
a022749 1550059419861007

You might also like