Electroencephalography-based machine learning models for hyperacute detection of large vessel occlusion stroke

Electroencephalography (EEG) has shown potential as a prehospital triage instrument for anterior circulation large vessel occlusion (LVO-a) stroke. In this study, we determined the diagnostic accuracy of EEG-based machine learning algorithms for LVO-a stroke detection.

Anterior circulation large vessel occlusion stroke detection machine learning models were developed using EEG data from the ELECTRA-STROKE (NCT03699397) and AI-STROKE (NCT05437237) studies. Single 2- to 3-min EEG recordings were obtained using 8 dry electrodes in adults with suspected stroke in prehospital and in-hospital settings. Patients with a clean EEG segment of ≥ 5 s (visually selected) were included and EEG symmetry features were extracted. We assessed the performance of multiple machine learning models using nested cross-validation (CV), optimising models with 3-fold CV (inner loop) and evaluating them with leave-one-out CV (outer loop).

Of 1293 patients with suspected stroke, 336 (26%) were included (74% prehospital, 26% in-hospital). Median age was 72 (IQR 61-80) years, and 231 (69%) were male. Sixteen (5%) patients had LVO-a stroke, 140 (42%) non-LVO-a ischaemic stroke, 52 (15%) TIA, 13 (4%) haemorrhagic stroke and 115 (34%) a stroke mimic. Naïve Bayes achieved the highest observed area under the receiver operating characteristic curve (0.85 [95% CI, 0.77-0.93]), with a sensitivity of 69% (95% CI, 44%-86%) and specificity of 86% (95% CI, 82%-89%).

Machine learning models for LVO-a stroke detection based solely on EEG showed moderately high diagnostic accuracy. These results support the potential of EEG-based LVO-a stroke detection, but improved EEG data acquisition is needed before clinical implementation can be considered.