Abstract TMP30: Combining the Los Angeles Motor Scale and the Muse Portable Electroencephalography System Improves the Accuracy of Large Vessel Occlusion Detection in Acute Stroke Syndrome.

Background: The prehospital scales have been developed to identify stroke patients with large vessel occlusion (LVO) to facilitate rapid transport to appropriate stroke centres. In practice, these stroke scales have moderate accuracy. There is a pressing need for adjunct easy-to-use and interpret diagnostic devices to improve prehospital stroke diagnosis and LVO detection. We aim to determine whether a machine learning algorithm using adjunct electroencephalography (EEG) Spectra can improve the accuracy of LVO detection Methods: Adult patients with suspected acute stroke were prospectively enrolled as soon as possible on arrival at the emergency department. A wearable Muse TM headband (InteraXon Inc, Canada) with an embedded 4-channel EEG was used for a resting 3-minute recording. EEG Spectra including relative alpha, beta, theta and delta spectral powers, delta-alpha ratio (DAR) and pairwise-derived brain symmetry indices (pdBSI) were calculated. These indices were compared between patients with LVO and non-LVO groups. The accuracy of LVO detection was tested with the aid of supervised machine learning(ML) algorithms including EEG Spectra, Los Angeles Motor Stroke Scale (LAMS), sex and side of stroke. Results: A total of 142 patients were included in the analysis with a mean age of 69.6±13.7 years, 60(42.2%) females, (Stroke Subtype:113[79.6%] were ischemic stroke, 22[15.5%] stroke mimics, 7[4.9%] intracerebral hemorrhage) and median NIHSS 5(2-11). Thirty-seven(26.1%) patients had LVO and EEG was acquired at a median of 6h 45m (3h 29m - 14h 15m) after symptom onset. Relative alpha spectral power was lower in both affected (p<0.0001) and unaffected hemispheres (p<0.0001) in the LVO group (Figure 1); there was no difference in the median affected hemisphere DAR (p=0.4). However, the median unaffected hemisphere DAR was higher in the LVO group compared to the no-LVO group (p=0.03) (Figure 2). The Support vector machine-based ML algorithm accuracy for detecting LVO was: 0.6 for clinical assessment (LAMS+ Side of deficit+ Sex) alone, 0.73 for EEG Spectra alone and 0.93 for clinical assessment + EEG Spectra (0.9 Sensitivity; 0.95, Positive Predictive value) Conclusion: Combining QEEG with clinical assessment significantly improves the overall accuracy of LVO detection versus LAMS alone in patients presenting with acute stroke syndrome. Future studies are ongoing to determine if a short EEG acquisition in the prehospital phase is useful for rapid triage.