Surface Electroencephalography (EEG) During the Acute Phase of Stroke to Assist With Diagnosis and Prediction of Prognosis: a Scoping Review


 BackgroundStroke is a common medical emergency responsible for significant mortality and disability. Early identification improves outcomes by promoting access to time-critical treatments such as thrombectomy for large vessel occlusion (LVO), whilst accurate prognosis could inform many acute management decisions. Surface electroencephalography (EEG) shows promise for stroke identification and outcome prediction, but evaluations have varied in technology, setting, population and purpose. This scoping review aimed to summarise published literature addressing the following questions:1. Can EEG during acute clinical assessment identify:a) Stroke versus non-stroke mimic conditionsb) Ischaemic versus haemorrhagic strokec) Ischaemic stroke due to LVO. 2. Can these states be identified if EEG is applied <6hrs since onset. 3. Does EEG during acute assessment predict clinical recovery following confirmed stroke.MethodsWe performed a systematic search of five bibliographic databases ending 19/10/2020. Two reviewers assessed eligibility of articles describing diagnostic and/or prognostic EEG application <72hrs since suspected or confirmed stroke. ResultsFrom 5892 abstracts, 210 full text articles were screened and 39 retained. Studies were small and heterogeneous. Amongst 21 reports of diagnostic data, consistent associations were reported between stroke, greater delta power, reduced alpha/beta power, corresponding ratios and greater brain asymmetry. When reported, the area under the curve (AUC) was at least good (0.81–1.00). Only one study combined clinical and EEG data (AUC 0.88). There was little data found describing whether EEG could identify ischaemic versus haemorrhagic stroke. Radiological changes suggestive of LVO were also associated with increased slow and decreased fast waves. The only study with angiographic proof of LVO reported AUC 0.86 for detection <24hrs since onset. Amongst 26 reports of prognostic data, increased slow and reduced fast wave EEG changes were associated with future dependency, neurological impairment, mortality and poor cognition, but there was little evidence that EEG enhanced outcome prediction relative to clinical and/or radiological variables. Only one study focussed solely on patients <6hrs since onset, for predicting neurological prognosis post-thrombolysis.ConclusionsAlthough studies report important associations with EEG biomarkers, further technological development and adequately powered real-world studies are required before recommendations can be made regarding application during acute stroke assessment.


Abstract Background
Stroke is a common medical emergency responsible for signi cant mortality and disability. Early identi cation improves outcomes by promoting access to time-critical treatments such as thrombectomy for large vessel occlusion (LVO), whilst accurate prognosis could inform many acute management decisions.
Surface electroencephalography (EEG) shows promise for stroke identi cation and outcome prediction, but evaluations have varied in technology, setting, population and purpose. This scoping review aimed to summarise published literature addressing the following questions: 1. Can EEG during acute clinical assessment identify: a) Stroke versus non-stroke mimic conditions b) Ischaemic versus haemorrhagic stroke c) Ischaemic stroke due to LVO.
2. Can these states be identi ed if EEG is applied <6hrs since onset.
3. Does EEG during acute assessment predict clinical recovery following con rmed stroke.

Methods
We performed a systematic search of ve bibliographic databases ending 19/10/2020. Two reviewers assessed eligibility of articles describing diagnostic and/or prognostic EEG application <72hrs since suspected or con rmed stroke.

Results
From 5892 abstracts, 210 full text articles were screened and 39 retained. Studies were small and heterogeneous. Amongst 21 reports of diagnostic data, consistent associations were reported between stroke, greater delta power, reduced alpha/beta power, corresponding ratios and greater brain asymmetry. When reported, the area under the curve (AUC) was at least good (0.81-1.00). Only one study combined clinical and EEG data (AUC 0.88). There was little data found describing whether EEG could identify ischaemic versus haemorrhagic stroke. Radiological changes suggestive of LVO were also associated with increased slow and decreased fast waves. The only study with angiographic proof of LVO reported AUC 0.86 for detection <24hrs since onset. Amongst 26 reports of prognostic data, increased slow and reduced fast wave EEG changes were associated with future dependency, neurological impairment, mortality and poor cognition, but there was little evidence that EEG enhanced outcome prediction relative to clinical and/or radiological variables. Only one study focussed solely on patients <6hrs since onset, for predicting neurological prognosis post-thrombolysis.

Conclusions
Although studies report important associations with EEG biomarkers, further technological development and adequately powered real-world studies are required before recommendations can be made regarding application during acute stroke assessment.

Background
Stroke is responsible for a high disability, mortality and economic burden worldwide. Emergency treatments can improve outcomes [1,2], particularly intravenous thrombolysis and mechanical thrombectomy for selected patients with ischaemic stroke. These highly time-sensitive treatments reduce long-term disability when administered < 4.5 and < 6 hours respectively, but urgent clinical assessment including brain Computed Tomography (CT) or Magnetic Resonance Imaging (MRI) must rst determine eligibility. For mechanical thrombectomy, additional angiography (CTA or MRA) is needed to con rm the presence of large vessel occlusion (LVO), with subsequent transfer of treatable patients if they are not already at a comprehensive stroke centre [3]. Earlier identi cation of individual patients most likely to bene t from speci c emergency treatments will improve outcomes, especially if this is possible in the prehospital setting so that ambulance admissions can be directed to the most appropriate facility.
Accurate initial identi cation of stroke patients is complicated by 'mimic' conditions that produce the same symptoms as stroke, such as epileptic seizures, migraine and infections. A literature review of 79 studies reported that, despite routine use of symptom checklists like the Face Arm Speech Test, an average of 27% (range: 4-43%) prehospital suspected stroke admissions and 10% (range: 1-25%) thrombolysis patients were later re-categorised as stroke mimics [4]. More complex symptom checklists have been developed to identify LVO, but these have not been widely adopted due to the unfavourable balance between speci city and sensitivity [5,6]. Point-of-care tests to distinguish stroke from mimic patients, haemorrhagic from ischaemic stroke and/or identify LVO would allow earlier access to appropriate emergency care, but none are available currently [7]. Similarly, portable technologies providing early information about prognosis could assist clinicians whilst making a range of acute management decisions, such as whether treatment of early complications would be likely to in uence recovery or might possibly be futile. Electroencephalography (EEG) is a non-invasive clinical tool frequently used in hospital-based diagnosis and management of seizures, but has also been evaluated for stroke identi cation and prognostication. An increase in slow-wave (delta) versus faster (alpha/beta) activity has long been recognised following a recent stroke, although the exact mechanism is uncertain [8-10]. Quantitative EEG (qEEG) has been used as a biomarker to predict outcomes in ischaemic stroke in acute and sub-acute settings [11,12]. Its ability to detect and size lesions [13,14] suggests that it could be used as a diagnostic tool and a clinical decision aid during treatment decisions. Advances in qEEG analysis methods and algorithms such as the Brain Symmetry Index [15], and introduction of portable systems using a minimal number of electrodes [16,17], have increased the practical potential for use in emergency department (ED) and prehospital settings [18]. We undertook a literature review to describe the use of EEG during the acute phase of stroke for strati cation of unselected patients into important clinical groups, and as an aid for clinical decision-making through early estimation of prognosis. A scoping review approach was applied due to signi cant heterogeneity in technology and setting in this emerging eld.

Methods
The Preferred Reporting Systems for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR) framework was applied [19].

Aim
The aim was to report evidence describing the capability of EEG technologies for strati cation (identi cation and prognostication) when applied within 72 hours of stroke symptom onset.
Objectives By classifying and describing clinical studies of EEG technologies applied soon after stroke symptom onset (< 72hrs), we addressed the following questions: 1. Can EEG during acute clinical assessment identify: a. Stroke versus non-stroke mimic conditions b. Ischaemic versus haemorrhagic stroke c. Ischaemic stroke due to LVO.
2. Can these states be identi ed if EEG is applied < 6 hours of symptom onset.
3. Does EEG during acute assessment predict clinical recovery following con rmed stroke.

Search strategy
Following exploratory searches, a systematic strategy combining MeSH/Web of Science categories and keywords was developed and executed in Ovid (selecting Medline, Embase and PsycINFO databases), Web of Science and Scopus databases up until the 19th October 2020 inclusive. Hand searching of reference lists and citation searches of included studies were undertaken. Only published peer-reviewed literature was retained, including conference abstracts if there was su cient information reported, but case studies were excluded. It was not necessary to contact the authors of any articles for clari cation. The search strategies are listed under 'Supplement A' in the Supplementary Material.

Study Inclusion Criteria
Research studies and review articles, including feasibility and pilot studies, with abstracts published in English from any country were eligible for inclusion if they presented original data and appropriate statistical comparison describing the application of EEG technology for stroke identi cation or prognosis. It was necessary for the test population to include patients with suspected or con rmed stroke, where the EEG technique was commenced (but not necessarily completed) within 72 hours. Although this time window extended beyond the interval for delivery of emergency stroke treatments, it enabled inclusion of information from studies with a range of onset to EEG times. Studies that focused mainly or solely on seizures (including prediction of post-stroke epilepsy) or Transient Ischaemic Attack (TIA) (stroke symptoms resolved within 24hrs) were excluded.
Any EEG-based assessment was permissible, including but not limited to: qualitative visual analysis of EEG, qEEG, continuous EEG monitoring, the Brain Symmetry Index (BSI) and frequency-speci c power measures such as delta/alpha power ratio (DAR) or (delta + theta)/(alpha + beta) power ratio (DTABR).
The study setting could be in hospital or in an ambulance, including situations where patients were conveyed to a specialist laboratory from hospital for EEG recording.
Any diagnostic process was accepted for the stroke reference standard i.e. MRI/A, CT/A and/or specialist opinion. Comparisons against mimic conditions and non-stroke/healthy controls were included when the origin of the source data was stated. However, studies were not included if stroke patient data were being compared only to standard de nitions of 'healthy/normal' EEG parameters, without description of a reference data source.
Studies examining detection of LVO were included if there was direct evidence of large artery occlusion (e.g. CT angiography) or, because not many studies were expected to use this reference standard, we also considered studies reporting indirectly associated radiological features (e.g. large infarct size).
For prognostic studies we included those using any previously described clinical stroke outcome measure, or survival/death. For these studies, we reported only the main outcome of interest as stated by the authors.

Study selection
Duplicate articles were excluded. Two members of the study team (LSu + RF) reviewed titles and abstracts and selected full text articles to con rm inclusion with arbitration by a third reviewer if required (CP and/or LSh). Templates for review, extraction and quality assessment can be found under 'Supplement B' in the Supplementary Material.

Data extraction
Data were independently extracted by two reviewers (LSu and HL), with discrepancies resolved via group discussion.
A data extraction framework was developed and piloted by the reviewers before use, which included elds for: Year of publication, country of origin, study aims, study design, setting, inclusion/exclusion criteria, EEG technology, EEG data processing methodology, reference standard information, outcome measures, blinding, sample size, time from stroke onset to rst EEG measure, major ndings (including statistical signi cance and diagnostic accuracy) and whether all patients were represented in the data with any exclusions explained.
To assess study quality, a simple scoring system (0-5) was created which re ected the main indicators of good research design i.e. clear eligibility criteria; clearly de ned technology; clearly de ned reference standard and/or outcome measure; blinding; whether all participants were accounted for in the results presented. Studies were not excluded based on quality, but quality and design were considered during recommendations based upon strength of evidence.

Data synthesis
As this was a scoping review, there was no a-priori plan for data meta-analysis and a narrative description is provided. Data are presented in tables according to reference standard or outcome measure in ascending order of publication date.

Results
Databases searches identi ed 7624 articles, with 20 more from hand-searching relevant review publications. After removal of duplicates, 5892 abstracts remained. Of these, 5682 abstracts did not meet the inclusion criteria. The remaining 210 full text articles were assessed (Fig. 1) and 171 articles were excluded: 24 did not meet study design criterion, 70 did not meet the participants criterion, 16 did not address the review question, 59 did not meet multiple criteria, and 2 were republished as another included study. After full text review, 39 articles were included for data extraction and quality assessment: 13 reporting diagnostic data only, 18 reporting prognostic data only and eight articles reporting both.

Included studies
Study designs were diverse. The majority were cohort (n = 24) or case-control (n = 14) studies, although very few speci cally used these terms. Only one study [20] was considered a true diagnostic accuracy study, as the investigators performing the EEG were blinded to patients' clinical status and the reference standard was determined in advance (clinical specialist opinion).

Population
There was a wide geographic distribution of studies: Eight in China; seven in Australia; ve in USA; three each in Belgium and Cuba; two each in Portugal and Israel and one each in Germany, Indonesia, Ukraine, Italy, Brazil, Finland, France and Hungary. Nationality was unclear for one conference abstract. Apart from one study where the setting was unclear, most were conducted in acute care settings in hospital (two in Emergency Departments; four in an Intensive Care Unit; seven in a neurology department; 10 in a stroke unit and 15 in hospital with no clear department). No studies were conducted in an ambulance or in the prehospital setting.
The median number of patients across the 39 articles was 33 (range = . Inclusion and exclusion criteria were extremely variable, with some studies requiring extensive lists of exclusions and others giving limited or no information beyond a diagnosis of "stroke". Two diagnostic studies [21,22] included TIA as part of the stroke patient sample, whereas others excluded TIA. Inclusion/exclusion criteria that appeared frequently are listed in Supplement C (Table S1) in the Supplementary Material.
Median time from stroke onset to EEG application was 48hours (range = 4.5-72 hours) when this information was available. There was only one study where all patients were within six hours of symptoms onset, which was examining EEG indicators for recovery of neurological impairment after thrombolysis [23].

Quality of Studies
Only two articles showed evidence of a sample size calculation [24,25]. One other article included a post-hoc power calculation and ascertained that only some of their EEG parameters/sub-analyses had adequate statistical power [14].
Sixteen articles had some evidence of outcome blinding. It was stated that the EEG assessor was blinded to clinical data in the reference standard for three diagnostic articles, but there was no explicit indication that the clinician assessing the reference standard was blinded to EEG data. Eight articles with a prognostic aim reported a variety of blinding methods: EEG and outcome assessors blinded (n = 5), only EEG assessor blinded (n = 3), only outcome assessor blinded (n = 1) and patients blinded (n = 1). For three articles that had both diagnostic and prognostic aims and any form of blinding, there was evidence that the EEG assessor (or secondary EEG assessor) or outcome assessor was blinded to clinical data.
There were ve articles where it was not possible to account for all the participants due to unclear text, gures, or presentation of data representing only individual patients.
Of the 21 articles with a diagnostic aim, 15 had evidence of a predetermined reference standard including specialist opinion. Of the 26 articles with a prognostic aim, the outcome measure was clearly de ned for ve studies but the majority were unclear as to whether a measure had been selected before commencing recruitment.
Study data 1a) Identi cation of stroke versus non-stroke Seventeen articles considered whether EEG could distinguish stroke from non-stroke; two of which speci cally aimed to distinguish between stroke and TIA. Studies are summarised in Table 1, grouped by year of publication and reference standard.
Fifteen articles examined differences between stroke from healthy controls, or an identi ed healthy control dataset, and two compared stroke with stroke mimic conditions [21,22]. Median article quality score was 3 (range 2-5), but even higher quality reports included only modest numbers of patients (e.g. ischaemic stroke cases ranged from 6-65 patients).
Amongst eight articles that calculated any summary indicator of diagnostic accuracy, performance was generally good or high [20,22,27,28,31,[33][34][35]. Two articles in particular displayed very high accuracy for individual EEG frequency bands but were not in complete agreement [20,31]. Within 24hrs of symptom onset, Finnigan (2016) [31] reported ischaemic stroke could be detected by greater delta (AUC 0.99) and theta (AUC 0.81) activity, but less alpha (AUC 0.97) and beta (AUC 0.90). However, although Rogers (2019) [20] also reported accurate prediction by greater delta activity (AUC 0.87) within 72hrs onset, there was no difference between stroke and controls for alpha and beta, and controls had greater theta activity (AUC 0.93). Finnigan (2016) [31] also reported very high AUC from higher DAR (AUC 1.0) and DTABR (AUC 0.99). Subsequent analysis con rmed that the DAR result could be replicated by using just two frontal electrodes (AUC 0.99) [34]. A more recent article used deep learning network-based modelling of clinical information and EEG data from electrode pairs selected by lasso regression within 24hrs of symptom onset, and showed the AUC was higher (0.88) than could be achieved by standard analysis of clinical and/or EEG data [22].
For two articles also aiming to distinguish stroke from TIA, median quality score was 4 (range 3-5), with small numbers of participants. One article [20] distinguished between stroke and TIA (as well as control) with high diagnostic accuracy using evoked potentials and spectral power across all bands, with greater delta less alpha and less beta in stroke versus TIA. The other [25] distinguished stroke from TIA using a modi ed BSI but did not nd any difference in slow:fast wave ratios. Only two studies considered differences between ischaemic and haemorrhagic stroke aetiologies, with differing methodologies and results. Studies are summarised in Table 2, grouped by year of publication and reference standard.
Both studies were of medium quality (median score 3.5, range 3-4). One was an examination of post-stroke seizures during EEG monitoring and found a higher incidence of these was predictive of haemorrhagic stroke [37]; extrapolated speci city was high but sensitivity low. The other used relative spectral power methods and found differences in global frequencies (i.e. a more abrupt decrease of higher frequencies in haemorrhage), but did not nd any useful diagnostic value in ratios such as PRI or DAR [30]. Five studies reported whether EEG data was associated with direct (angiographic; n = 1) or indirect (infarct volume; n = 4) radiological evidence that LVO was likely to be responsible for ischaemic stroke. Studies are summarised in Table 3, grouped by year of publication and reference standard.
The quality of these studies was mixed, with a median score of 3 (range 2-4). Four reported that relative spectral power detected large infarct volume (more common in LVO), either by identifying areas of increased slower-waves (delta [11,21] and theta [22]) and/or decreased fast-waves (beta [21,38] and alpha [22]). Epileptiform activity (including slowing of frequencies) differentiated between territorial infarcts more typical of LVO and sub-cortical infarcts more likely to result from small vessel ischaemia [39]. Two studies comparing activity between hemispheres showed a general trend towards increased slow waves in the affected hemisphere but also a reduction in faster waves in the contralesional hemisphere when infarct size was greater [21,39]. The only study with direct angiographic evidence of LVO [22] used deep learning models combining clinical and EEG data, showing that the combination could achieve a high level of accuracy to detect 7 cases of LVO amongst 100 cases of suspected stroke (AUC 0.86, sensitivity = 76%, speci city = 80%).

2) Diagnostic accuracy within 6 hours of symptom onset
No diagnostic studies were found which consisted purely of patients within six hours of symptom onset. For identi cation of stroke versus non-stroke, only one small study included patients who were all within nine hours of onset [11], showing signi cantly greater mean delta power for stroke versus control.
Studies which considered ischaemic versus haemorrhagic stroke only included patients within 24 [37] and 72 [30] hours of onset, and no conclusion can be drawn about early EEG application for this purpose. For detection of LVO, two out of ve studies involved participants who were potentially within time windows for thrombectomy treatment; within nine [11] and ten [38] hours. These studies did not have high quality scores, but both showed associations with large volume infarction (loss of beta power and higher aDCI respectively) which may indicate that early changes associated with LVO are detectable.

3) Prediction of outcome following con rmed stroke
Twenty-six articles investigated the use of EEG biomarkers in predicting clinical recovery following con rmation of a stroke diagnosis within the previous 72 hours. Studies are summarised in Table 4 grouped by year of publication and outcome of interest. Prognostic articles had a median quality score of 4, re ecting a range of scores from poor to excellent (2-5).
Seven articles assessed later neurological impairment using the NIHSS. Of three studies seeking associations with abnormal EEG patterns, one found an association with epileptiform activity [37] and one generalised EEG slowing [39]. One study [52] did not report an association between epileptiform activity and the NIHSS, nding this was only useful in predicting seizure incidence. Three studies found associations between poor outcome and relative band power or ratios using various biomarkers such as less relative alpha power and greater DAR [53], greater interhemispheric alpha peak frequency asymmetry [54] and greater aDCI [11]. One study showed associations between a more favourable NIHSS after thrombolysis for ischaemic stroke and early decreases in BSI, DAR and DTABR [23]; this was the only study to focus solely on patients within six hours of symptoms onset. None of these studies calculated summary statistics to re ect accuracy.
Four studies assessed outcome by cognitive function (MoCA or diagnosis of dementia). Three found associations between spectral power and poorer cognitive outcome: lower theta, higher delta, greater DTR and DAR [24], lower beta [38], and greater theta with high background rhythm frequency [55]. Four studies considered prediction of mortality at various time intervals after stroke. At hospital discharge, greater contralateral theta power [57] and greater asymmetry measured by the Bilateral Brain Symmetry Index (BBSI) [35] were associated with poorer outcome. Higher DAR at day 90 [32], greater contralateral theta power at 6 months [57], and epileptiform activity, background slowing and overall asymmetry at 12 months [43] were associated with poorer outcome.
Prognostic accuracy was moderate, but with poor sensitivity, in two studies [43,57] and appeared high for two other studies [32,35]. There is evidence to support potentially valuable diagnostic accuracy of EEG approaches for differentiating stroke from non-stroke states including healthy controls, mimic conditions, and TIA patients, using biomarkers such as spectral power (e.g. [20]), DAR and DTABR (e.g. [31]) and BSI (e.g. [36]). Generally, there were statistical associations between a diagnosis of stroke, increased slow-wave EEG activity (delta in particular) and decreased fast-wave activity (alpha and beta). Although theta activity was often increased for stroke relative to control subjects, this was not a consistent nding and it appears to be the least useful frequency for diagnosis in this context, probably due to its intermediate speed between alpha (fast) and delta (slow). However, despite these promising early studies, it is important to recognise that most were small and included selected patients who were beyond six hours since symptom onset, so there is relatively little evidence that the potential EEG biomarkers identi ed would be present in the very early stages of stroke when the impact for emergency care decisions would be greatest (e.g. to initiate direct ambulance transfer to a stroke centre).
Evidence for the ability of EEG to distinguish between haemorrhagic and ischaemic stroke was limited to two studies, which again were not focussed upon the early hours when this information would be of greatest clinical value e.g. for administration of thrombolytic therapy within 4.5 hours of symptom onset. One study predicted haemorrhage based upon a greater frequency of acute seizures with relatively poor accuracy [37], whilst the other reported differences in the alpha-beta range but no detailed data were provided [30]. Currently it appears unlikely that EEG has a role to play in the differentiation between ischaemic and haemorrhagic stroke that would change patient management.
On the basis that indirect radiological evidence is a reliable indicator of LVO, a small number of studies support the further development of EEG biomarkers for this purpose, mainly using frequency band power to indicate areas of increased slow or decreased fast activity correlating with larger areas of acute infarction. Only two studies focused purely on a suspected stroke population during the standard time interval of maximal clinical value for thrombectomy [11,38]. For all studies, it was unclear whether participants were representative of an unselected suspected stroke population, which is where prediction of LVO would have the greatest clinical impact. Only one study used angiography as the reference standard and reported that the most promising AUC (0.86) was achieved by combining clinical and EEG (lower alpha and greater theta) data with a deep learning algorithm [22]. In addition, a prospective study published since completion of our search has also con rmed that amongst 109 patients within 24 hours of symptom onset (25 angiography-proven LVO, 38 non-LVO ischemic, 14 haemorrhages, and 32 stroke mimics) an AUC of 0.88 was achieved using a portable LVO-detection device which combined EEG and somatosensory-evoked potentials [58].
When used to provide an early estimate of prognosis, EEG biomarkers recorded within 72 hours of stroke onset had associations with later clinical outcomes which could be useful to inform acute management decisions including future dependency, neurological impairment, cognitive function and mortality. In particular, greater delta and theta activity, less alpha and beta activity, greater interhemispheric asymmetry and greater DAR and DTABR ratios appear to be predictors for both long-, and short-term neurological function and dependency. Apart from one study showing poor predictive value of a non-speci c EEG biomarker for post-stroke dementia [56], accuracy was at least moderate. Such associations are not unexpected as EEG changes re ect the volume of cerebral tissue injury, which itself directly correlates with dependency and survival [59]. Only one cohort was identi ed where EEG information improved upon the accuracy of predictions for dependency and mortality made using simple clinical assessments and/or brain imaging to con rm the number and location of vascular lesions [43,45]. There are, however, validated clinical scores already available to estimate various aspects of physical stroke recovery (e.g. arm function [60]; independent walking ability [61]) which are not widely used in practice because of concerns that they could restrict access to nite, but potentially bene cial, care resources [62]. Therefore, in parallel with further research focussed upon whether surface EEG can re ne early clinical prediction of future survival and dependency, it is necessary to understand whether using such technology as a decision support aid is an acceptable concept and how the results would be communicated to patients and their families. Likewise, although clinical models to predict cognitive impairment after stroke have been created, currently there is insu cient evidence of validity and/or accuracy for routine use [63]. Neuroimaging variables such as white matter lesions have separately been found to be risk factors for dementia after stroke [64] and so it is feasible that EEG biomarkers will be helpful in identifying patients with a subclinical risk. However, studies identi ed by our search did not combine EEG data with neuroimaging variables, or compare to age matched controls, and it will be necessary to undertake additional longitudinal studies of well described cohorts before it is clearer whether EEG has a role to play as a clinical decision aid by providing prognostic estimates during acute stroke care.
According to the basic scoring system we employed, most studies were not high quality, usually due to a lack of clarity about populations, reference standards and adjudication. Few studies produced power calculations or seemed to be adequately powered given required case:variable ratios. Many studies did not calculate prognostic or diagnostic accuracy or provide information that would be important to determine clinical utility, such as the number of patients who could not tolerate the procedure and the time required to obtain an EEG recording. Techniques using large numbers of electrodes are unlikely to be deployed during emergency assessment of suspected stroke if application requires additional training and signi cantly delays routine care, but it is encouraging that diagnostic value was reported by studies using six electrodes or fewer [20,34,36]. Clinical feasibility will be further facilitated by easily applicable dry (sans electroconductive gel) electrode systems, and ongoing development of machine learning approaches to automatically select electrode pairs and rapidly identify multi-wave activity patterns predictive of a stroke diagnosis or LVO [22,58]. Rapid application is less essential for collecting information to inform prognosis and could be done after hospital arrival, but it is still important to consider that some patients may not be able to tolerate a lengthy EEG procedure and e cient portable systems will minimise disruption of acute clinical care.
Finally, our review has some limitations which should be acknowledged. It was not possible to include studies written in a non-English language, which may have excluded relevant reports that did not already have an English translation available. There was a wide variation in EEG technique (e.g. lters and electrode placement) and outcome measures which prevented data meta-analysis and hinders recommendation of a speci c technical approach for diagnosis or prognosis. Additionally, most studies had strict inclusion criteria to minimise interference with the EEG signal, limiting the generalisability of ndings to the wider population. There have been recent advances in commercial EEG technology for use in stroke diagnosis, notably for early identi cation of LVO [17], but our review was limited to published studies. Other portable technologies are also in development for emergency detection of stroke and LVO, including blood assays and non-ionising imaging [7], and the future clinical value of surface EEG should be considered alongside alternative biomarkers used separately and in combination.

Conclusion
Reports identi ed during this review show that surface EEG techniques have promise for assisting with stroke diagnosis and prognosis during the acute phase. However due to the small size of studies and variations in technology, populations and settings, it is not yet possible to make recommendations regarding EEG use to guide early diagnostic and prognostic management decisions. Further research is required to determine which combinations of

Availability of Data and Materials
Data sharing is not applicable to this article as no datasets were generated or analysed during the current study.
Author Contributions CP and LSh conceived the review. CP, LSh, LSu and RF developed the review methods. LSu and CP designed and conducted the search strategy. LSu and RF, with assistance of CP and LSh, assessed studies for inclusion. LSu, HL and RF extracted data from included studies. LSu, HL, CP and RF drafted the manuscript. LSu, HL, CP, LSh and RF were involved in the interpretation of data, critically reviewed the manuscript for intellectual content and approved the nal version of the manuscript.
Funding CP and LSu received salary funding from the UK Stroke Association during production of the review.
Competing Interests HL, LSu and RF declare no con icts of interest. CP and LSh declare interests as investigators for two non-commercial studies of new technologies to assist with stroke assessment (PRISM: http://www.isrctn.com/ISRCTN22323981 and ABACUS: http://www.isrctn.com/ISRCTN79169844).