Mei Ngun, MBBS
The search for tools to improve the diagnosis of acute stroke is ongoing. Recently, Erani et al. explored the use of electroencephalography (EEG) in diagnosing stroke, especially in large vessel occlusion (LVO) or transient ischemic attack (TIA) in the emergency department (ED). To date, the practicalities of performing EEG have meant that there has been limited clinical use in the emergent stroke setting.
Patients with suspected/definite acute stroke were recruited from an ED of a single comprehensive stroke center. EEG was recorded using a dry-electrode system (Quick-20) with a local active amplifier and Faraday cage. At the bedside, eyes-open, resting state EEG was recorded for 3 minutes. EEG data was exported for offline analysis, filtering, and removal of noise to produce a bipolar montage of 27 bipolar lead-pairs.
A hundred patients were recruited. Lasso regression modelling was used to select a subset of electroencephalography variables. The EEG variables and certain clinical variables were used in four predictor models, which were first trained on 60 randomly selected patients and then tested on an independent validation cohort of 40 patients. The models were first evaluated using acute stroke/TIA (or not) as the dependent measure, then again with acute stroke with LVO (or not) as the dependent outcome.
Discharge diagnosis was acute stroke or TIA in 63 patients. Seven patients had a large vessel occlusion (all M1), and 14 received thrombolysis with tissue-type plasminogen activator (tPA). Median time from last-known-well to EEG was 9.4 hours, which is over the extended time window for thrombolysis in select patients. The median time from ED arrival to EEG was 3.7 hours. However, the median time from start of EEG preparation to EEG recording was only 9 minutes, with this duration reducing during the study to as low as 36 seconds. It appears then that the major contributor was not the time for setup of EEG, but other secondary factors that were not detailed in the report.
The AUC of the models ranged between 62.3 to 87.8, with the model using combined clinical and EEG variables with a deep learning neural network model performing the best. This model had a sensitivity of 80% at a specificity of 80%. All 3 models with electroencephalography were significantly (P=0.016–0.004) better predictors than the clinical-only model. The models performed similarly for acute stroke with LVO.
This study shows that it is feasible to utilize EEG as part of the armamentarium for stroke diagnosis, as it is able to detect cerebral ischemia immediately and the use of dry electrodes shortens preparation time. While it appears to increase diagnostic certainty compared to clinical criteria alone, real world implementation will have to consider the additional costs that this may incur in the form of equipment and requirement for trained staff. Even more pertinently, the median time for EEG from last-known-well and from arrival to ED were both within range of hours, and therefore would not be able to inform the decision regarding hyperacute treatment such as thrombolysis and endovascular clot retrieval. Future studies looking at the usage of EEG for hyperacute stroke and in the pre-hospital setting would be informative.
