Prehospital Detection of Large Vessel Occlusion Stroke With EEG

Background and Objectives Endovascular thrombectomy (EVT) is standard treatment for anterior large vessel occlusion stroke (LVO-a stroke). Prehospital diagnosis of LVO-a stroke would reduce time to EVT by allowing direct transportation to an EVT-capable hospital. We aim to evaluate the diagnostic accuracy of dry electrode EEG for the detection of LVO-a stroke in the prehospital setting. Methods ELECTRA-STROKE was an investigator-initiated, prospective, multicenter, diagnostic study, performed in the prehospital setting. Adult patients were eligible if they had suspected stroke (as assessed by the attending ambulance nurse) and symptom onset <24 hours. A single dry electrode EEG recording (8 electrodes) was performed by ambulance personnel. Primary endpoint was the diagnostic accuracy of the theta/alpha frequency ratio for LVO-a stroke (intracranial ICA, A1, M1, or proximal M2 occlusion) detection among patients with EEG data of sufficient quality, expressed as the area under the receiver operating characteristic curve (AUC). Secondary endpoints were diagnostic accuracies of other EEG features quantifying frequency band power and the pairwise derived Brain Symmetry Index. Neuroimaging was assessed by a neuroradiologist blinded to EEG results. Results Between August 2020 and September 2022, 311 patients were included. The median EEG duration time was 151 (interquartile range [IQR] 151–152) seconds. For 212/311 (68%) patients, EEG data were of sufficient quality for analysis. The median age was 74 (IQR 66–81) years, 90/212 (42%) were women, and the median baseline NIH Stroke Scale was 1 (IQR 0–4). Six (3%) patients had an LVO-a stroke, 109/212 (51%) had a non–LVO-a ischemic stroke, 32/212 (15%) had a transient ischemic attack, 8/212 (4%) had a hemorrhagic stroke, and 57/212 (27%) had a stroke mimic. AUC of the theta/alpha ratio was 0.80 (95% CI 0.58–1.00). Of the secondary endpoints, the pairwise derived Brain Symmetry Index in the delta frequency band had the highest diagnostic accuracy (AUC 0.91 [95% CI 0.73–1.00], sensitivity 80% [95% CI 38%–96%], specificity 93% [95% CI 88%–96%], positive likelihood ratio 11.0 [95% CI 5.5–21.7]). Discussion The data from this study suggest that dry electrode EEG has the potential to detect LVO-a stroke among patients with suspected stroke in the prehospital setting. Toward future implementation of EEG in prehospital stroke care, EEG data quality needs to be improved. Trial Registration Information ClinicalTrials.gov identifier: NCT03699397. Classification of Evidence This study provides Class II evidence that prehospital dry electrode scalp EEG accurately detects LVO-a stroke among patients with suspected acute stroke.


Introduction
6][7][8][9] On average, this so-called drip-and-ship workflow delays the initiation of EVT by 39-114 minutes and is associated with a 7.8%-21.4%lower chance of functional independence at 90 days. 5,7,91011triage instrument that can quickly and reliably identify patients with an LVO-a stroke in the prehospital setting would allow direct transportation of these patients to a comprehensive stroke center, which could substantially shorten EVT treatment initiation times.3][14][15][16][17][18] However, none of these methods have been implemented on a widespread scale, and inadequate prehospital stroke triage remains an important issue in acute care organization.0][21][22][23] Previous studies have shown that EEG can discriminate between LVO-a stroke patients and other suspected stroke patients in an in-hospital setting, 242526 but studies in the prehospital setting have not been performed.The primary aim of the ELECTRA-STROKE study was to determine the diagnostic accuracy of dry electrode EEG for LVO-a stroke detection in the prehospital setting, expressed as the area under the receiver operating characteristic (ROC) curve (AUC) of the theta/alpha ratio.

Methods
Study Design and Population ELECTRA-STROKE was a prospective, investigatorinitiated, multicenter, diagnostic study, consisting of an inhospital and a prehospital phase.The results of the in-hospital phase have been published previously. 26Here, we report the results of the prehospital phase.
The study was performed by 2 ambulance services (Ambulance Amsterdam and Witte Kruis Ambulancezorg Alkmaar) and 3 receiving hospitals in the Dutch province of North Holland (Amsterdam UMC, OLVG Hospital, and Noordwest Ziekenhuisgroep).The patient recruitment took place between August 2020 and September 2022.
We included adult patients with a clinically suspected stroke as judged by the ambulance nurse and onset of symptoms or time last seen well less than 24 hours before start of the EEG recording.Patients with a wound or active infection of the scalp in the dry electrode cap placement area and patients with a (suspected) COVID-19 infection were excluded.

Standard Protocol Approvals, Registrations, and Patient Consents
The study protocol (eSAP, links.lww.com/WNL/D121) was approved by the institutional review board of the Amsterdam UMC and by each participating hospital.The study protocol was published previously. 27The study was monitored by the Clinical Research Unit of the Amsterdam UMC.Because of the emergency setting of the study, we used a deferred informed consent procedure in line with national legislation and with approval of the institutional review board. 28Written informed consent was obtained from all patients or their legal representatives after EEG data acquisition.

Study Procedures
A single EEG recording was performed using a dry electrode EEG cap with 8 electrodes covering the vascular territory of the middle cerebral artery (FC3, FC4, CP3, CP4, FT7, FT8, TP7, and TP8; Waveguard touch, Eemagine, Berlin, Germany) and a compatible EEG amplifier (eego amplifier EE-411; Eemagine, Berlin, Germany).EEG data were acquired at a sample frequency of 500 Hz using clinical EEG software (NeuroCenter EEG, Clinical Science Systems, Leiden, The Netherlands).EEG recordings were performed in the prehospital setting by ambulance personnel who attended a 1-hour training in performing dry electrode EEG recordings for the purpose of this study.A portable and lightweight bag was used to store all equipment required for the EEG recording (Figure 1).In December 2020 (Witte Kruis Ambulancezorg Alkmaar) and February 2021 (Ambulance Amsterdam), we adapted the acquisition software to provide visual feedback on the electrode-skin impedances to guide optimization of the electrode-skin contact.We time-limited both the impedance optimization step (1.5 minutes) and the EEG recording step (2.5-3.0 minutes) to minimize delay in the prehospital workflow.As of January 2021, we collected information on the length of the hair of the patients in whom an EEG recording was performed by ambulance personnel of Witte Kruis Ambulancezorg Alkmaar.
Routine stroke workup included a noncontrast CT, and, if indicated, CT angiography and CT perfusion were performed in the receiving hospital.

Definitions and Outcomes
The primary endpoint of the study was the diagnostic accuracy of dry electrode EEG to discriminate LVO-a stroke from all other strokes and stroke mimics, expressed as the AUC of the theta/alpha ratio.The theta/alpha ratio was defined as: PðθÞ − PðαÞ PðθÞ + PðαÞ ; with PðθÞ the power in the theta frequency band (4-8 Hz) and PðαÞ the power in the alpha frequency band (8-13 Hz).
The theta/alpha ratio was chosen as this EEG feature had highest diagnostic accuracy for LVO-a stroke detection in the first 100 patients included in the in-hospital phase of the study. 26Besides the AUC, sensitivity, specificity, positive predictive value, negative predictive value, and positive likelihood ratio were determined.
Secondary endpoints, which should be considered exploratory only, were the diagnostic accuracies of the relative delta (1-4 Hz), theta (4-8 Hz), alpha (8-13 Hz), and beta (13-18 Hz) power; the delta/alpha ratio, the (delta + theta)/(alpha + beta) ratio, and the pairwise derived Brain Symmetry Index for LVOa stroke detection; and the logistical and technical feasibility of ambulance personnel performing an EEG recording with a dry electrode cap in the prehospital setting in patients with a suspected stroke.Definitions of the EEG features are included in the eAppendix 1 (links.lww.com/WNL/D119).We quantified the logistical and technical feasibility by calculating the percentage of patients with EEG data of sufficient quality for analysis and the percentage of clean EEG data per patient.
To account for learning effects and hardware and software improvements, we compared the percentage of clean EEG data of the first 50 EEG recordings with the percentage of clean EEG data of the last 50 EEG recordings performed in this study.The principal safety outcome was the occurrence of device-related adverse events.
We defined an LVO-a stroke as an occlusion of the intracranial part of the internal carotid artery, the first or proximal second segment of the middle cerebral artery (M1 and proximal M2, respectively), or the first segment of the anterior cerebral artery (A1).The occlusion location was determined by an adjudication committee-consisting of a neuroradiologist (M.S.K.) and a stroke neurologist (J.M.C.)-based on CT angiography data.The adjudication committee was blinded for EEG results.In case a patient was not clinically suspected of having an LVO-a stroke and therefore did not undergo CT angiography, the patient was scored as not having an LVO-a stroke.

Data Analysis
EEG data were re-referenced to a 12-channel bipolar montage with 6 bipolar channels located at each hemisphere and bandpass filtered between 0.5 and 35 Hz using a third order Butterworth filter.Artifacts were detected and rejected per 2-second bipolar channel segment using an automatic Convolutional Neural Network algorithm specifically developed for the detection of artifacts in dry electrode EEG data. 29The pairwise derived Brain Symmetry Index was calculated using EEG channels located on both hemispheres.All other EEG features were calculated per hemisphere.For patients with a stroke, we kept the ipsilesional EEG features, and for patients with a stroke mimic, we averaged the EEG features over both hemispheres (if available).Further details are provided in the eMethods (links.lww.com/WNL/D120).

Statistical Analysis
The sample size was calculated based on an expected specificity of dry electrode EEG for LVO-a stroke of 70%.At the start of the study, the sample size was set at 222, based on an expected incidence of LVO-a stroke of 7% and a dropout rate of 20%.An interim analysis, based on 137 suspected stroke patients in whom a dry electrode EEG recording was performed in the prehospital setting, showed an actual incidence of LVO-a stroke of 5% and a dropout rate of 58%.Patients were considered a dropout if EEG data were of insufficient quality, time of symptom onset was more than 24 hours before the start of the EEG recording, or no informed consent was provided.A revised incidence of LVO-a stroke of 5%, an alpha of 0.05 and a maximum margin of error of 7% indicated that 174 patients with a suspected stroke were required. 30As EEG data quality improved over time, 26 we assumed a dropout rate of 55% when recalculating the sample size.Considering this dropout rate, our revised estimation resulted in a sample size of 386 patients with a suspected stroke.
The diagnostic accuracies of all EEG features were evaluated by calculating areas under the ROC curves.Two cutoff values were determined as the highest sensitivity at a specificity of ≥70% and ≥80% for LVO-a stroke, respectively.The latter is defined as optimal cutoff.For both cutoff values, sensitivity, specificity, positive predictive value, negative predictive value, and positive likelihood ratio are reported with 95% confidence intervals.The 95% confidence intervals were estimated using the method of Hanley and McNeil 31 for the AUC; the Wilson interval 32 for the sensitivity, specificity, positive predictive value, and negative predictive value; and the method of Simel et al. 33 for the positive likelihood ratio.

Data Availability
Individual patient data cannot be made available under Dutch law because we did not obtain patient approval for sharing individual patient data, even in coded form.

Results
Ambulance personnel performed a dry electrode EEG recording in 386 patients, of whom 75/386 (19%) were subsequently excluded because no informed consent could be obtained (n = 65), symptom onset was >24 hours (n = 9), or symptoms were absent during EEG recording (n = 1; Figure 2).Of the 311 patients included in the study, 99 (32%) were excluded from analysis because of insufficient quality of  1. Excluded patients were more often female (59% vs 42%, p = 0.01) and more often had an LVO-a stroke (9% vs 3%, p = 0.02).Patients who were excluded from analysis also more often had long hair (32% vs 15%, p = 0.07).
Of the 212 patients included in data analysis, 6 (3%) had an LVO-a stroke, 109 (51%) a non-LVO-a ischemic stroke, 32 (15%) a TIA, 8 (4%) a hemorrhagic stroke, and 57 (27%) a stroke mimic.Patients with an LVO-a stroke had a higher Abbreviations: EVT = endovascular thrombectomy; IQR = interquartile range; IVT = IV thrombolysis; LVO-a stroke = anterior circulation large vessel occlusion stroke; NA = not applicable; NIHSS = NIH Stroke Scale.a p value for the difference between patients with and without an LVO-a stroke.b Of the 109 patients with a non-LVO-a ischemic stroke, 2 had an LVO stroke of the posterior circulation, located in the first segment of the posterior cerebral artery (n = 1) and basilar artery (n = 1).Seventy (64%) non-LVO-a ischemic strokes were located in the vascular territory of the anterior circulation.c All were intraparenchymal hemorrhages.f The patient with an LVO-a stroke who did not receive EVT had an occlusion of the extracranial internal carotid artery and proximal second segment of the middle cerebral artery (both left).The risk of EVT was deemed to great considering the minimal neurologic deficit (NIHSS: 3).h The patients without an LVO-a stroke who received EVT had a distal occlusion of the second segment of the middle cerebral artery (n = 3) and an occlusion of the third segment of the middle cerebral artery (n = 1).Number of patients who received IVT before EVT: g 3; i 5. Number of missing values: d 9; e 10; j 15.

Discussion
In this prehospital study, dry electrode EEG detected LVOa stroke with a high diagnostic accuracy in a population of suspected stroke patients.However, EEG data were of insufficient quality for analysis in around one-third of patients.
In ELECTRA-STROKE, an ambulatory LVO-a stroke detection device was tested by ambulance personnel in the prehospital setting.In line with other prehospital stroke trials, 34,35 we demonstrate that successful conduct of a prehospital stroke study is feasible but has its challenges.Consistent with the MR ASAP trial, 35 obtaining written deferred consent proved to be more challenging in a prehospital setting than in stroke trials that ran in an in-hospital setting, such as the MR CLEAN NO-IV trial. 36Lack of written informed consent led to exclusion of 17% of the patients.Still, without a deferred informed consent procedure, performing the study would have been logistically more challenging and could have resulted in substantial treatment delays.Abbreviations: AUC = area under the receiver operating characteristic curve; LVO-a stroke = anterior circulation large vessel occlusion stroke; NA = not available; NPV = negative predictive value; pdBSI = pairwise derived Brain Symmetry Index; PLR = positive likelihood ratio; PPV = positive predictive value.The primary classifying metric, the AUC, is bolded.a 194 patients, of whom 6 had an LVO-a stroke, used for analysis.
Prehospital detection of patients with an LVO-a stroke would enable direct transport of these patients to a comprehensive stroke center while still allowing other suspected stroke patients to be transported to the closest primary stroke center.Data from nonrandomized studies suggest that this workflow can lead to shorter EVT treatment initiation times and better clinical outcomes for patients with an LVO-a stroke. 5,7,91011y contrast, the RACECAT trial-performed in a nonurban area with relatively long initial transport times-did not find a difference in clinical outcome in suspected LVO stroke patients between a mothership and drip-and-ship model. 37This discrepancy between studies may be attributed to geographic differences and emphasize the need for regional solutions to optimize stroke care.
Recent in-hospital studies have demonstrated the potential of EEG for LVO-a stroke detection.In the first 100 patients enrolled in the in-hospital phase of the ELECTRA-STROKE study, the diagnostic accuracies of the frequency band powers were comparable with this study, with the theta/alpha ratio as the best performing EEG feature (AUC = 0.83). 26The pairwise derived Brain Symmetry Index, however, did not have a good diagnostic accuracy in that study (AUC = 0.38), while in this study, this EEG feature showed the best performance (AUC = 0.91).This disparity may be due to a difference in EEG data quality, as the pairwise derived Brain Symmetry Index is known to be highly susceptible to artifacts and this study included an improved, more sensitive, artifact detection algorithm than the in-hospital study.Erani et al. 24 performed dry electrode EEG recordings in 100 suspected stroke patients and found moderately high diagnostic accuracy of EEG for LVO-a stroke detection (AUC = 0.69) when using a combination of the ipsilesional relative theta and alpha power.We assume that the diagnostic accuracy is lower compared with our study because none of the utilized EEG features contained information on both hemispheres-as is the case for the pairwise derived Brain Symmetry Index-and the diagnostic accuracy was validated on unseen data.In our study, we did not validate the results in unseen data, which may have led to an overestimation of the diagnostic accuracy.Sergot et al. 25 investigated LVO detection by means of a portable wet electrode EEG and somatosensory-evoked potentials device.They used data of 109 suspected stroke patients and found a sensitivity and specificity for LVO stroke of both 80%.
The value of using prehospital stroke scales, especially the RACE, G-FAST, and ACT-FAST, for LVO-a stroke detection has recently been demonstrated. 15,38The best performing  scale, the RACE, reached an AUC of 0.83 in a group of 1,039 suspected stroke patients.Although clinical scales have the potential to be easily implemented in the prehospital setting and can be applied to most patients, assessment remains subjective.An important aspect of progress may lie in combining EEG with clinical scales in future studies.The improvement of diagnostic accuracy for LVO-a stroke detection, when combining EEG and clinical data, has been shown by Erani et al. 24 Unfortunately, both the RACE and G-FAST scales are not routinely used by ambulance personnel in the regions where ELECTRA-STROKE was performed; thus, we could not study the value of combining both techniques in this study.
Our study has several limitations.First, EEG data quality was insufficient in approximately one-third of patients.The fact that ambulance personnel had to perform the measurement in a limited timeframe and while under pressure of transporting a suspected stroke patient as quickly as possible undoubtedly contributed to this.Future innovations in EEG hardware and software should specifically address this issue, especially among groups with higher dropout rates due to insufficient EEG data quality (women, those with longer hair and those with more severe neurologic deficits).Second, we only had a limited number of LVO-a stroke patients in the study, increasing the statistical uncertainty of the results.Third, we had insufficient power to determine whether EEG can distinguish between LVO-a stroke and hemorrhagic stroke.This should be the focus of future studies because hemorrhagic stroke is an important mimic of LVO-a stroke and there are data that suggest that subjecting patients with a hemorrhagic stroke to longer transportation times may negatively affect their outcome. 37 conclusion, the data from this study suggest that dry electrode EEG has potential to detect LVO-a stroke among patients with suspected stroke in the prehospital setting.Toward future implementation of EEG in prehospital stroke care, EEG data quality needs to be improved and additional validation in a larger cohort of patients is necessary.

Figure 1 EEG
Figure 1 EEG Setup as Used in the ELECTRA-STROKE Study

Figure 2
Figure 2 Patient Flowchart

Figure 3 ROC
Figure 3 ROC Curves for LVO-a Stroke Detection

Table 2
Baseline Characteristics of All Patients Included in the Final Analysis

Table 3
Diagnostic Accuracy of EEG Features for LVO-a Stroke Detection at Optimal Cutoff