(728) Association Between Discharge Outcomes and Seizure Burden From a Point-of-Care EEG AI Algorithm

Introduction: There are known associations between the extent of seizures and epileptiform activity on electroencephalography (EEG) recordings and worse clinical outcomes. Point-of-care (POC)-EEG is a rapid diagnostic tool that allows for fast seizure triage and can run Clarity (Ceribell, Inc.), an AI algorithm that monitors seizure burden (SzB) in 5-min windows. We hypothesized that there will be a direct relationship between SzB and poor functional outcomes at discharge.

Methods: This was a cohort subanalysis of the retrospective SAFER-EEG trial, which included adult patients who received EEG monitoring during their hospital stay. For this report, we included only POC-EEG cases from three sites that had valid modified Rankin Scale (mRS) scores pre-EEG and at discharge. Two algorithms were run post-hoc, Clarity and a more seizure-sensitive algorithm, which is part of the recently approved Clarity Pro. A maximum 5-min SzB was obtained for Clarity and the combination of the two algorithms. We evaluated the odds of poor outcome (discharge mRS ≥ 4) at different SzB thresholds (1%, 5%, 10%, 20%, 50% and 90%) and as a continuous variable. We adjusted for clinical covariates that may affect outcome.

Results: We included 359 patients in the analysis, 212 of these (59.1%) had 0% max SzB per Clarity and 103 (28.7%) had 0% max SzB when combining both algorithms. In the multivariate models, we found that as the SzB threshold increased the odds of poor outcome also increased, with respect to patients with 0% SzB. Per Clarity, the adjusted odds ratio (OR) ranged from 1.3 to 4.9 but it was only significant for SzB ≥50% (OR = 3.5 [95% C.I: 1.5-8.9]; p = 0.002) and ≥90% (OR = 4.9 [1.6-17.9]; p = 0.009), after multiple-comparison correction. When combining both algorithms, the association was more robust and the OR ranged from 1.2 to 6.1, with SzB ≥50% (OR = 4.2 [1.6-11.8]; p = 0.005) and ≥90% (OR = 6.1 [2.1-20.3]; p = 0.002) remaining significant after multiple-test correction.

Conclusions: We showed a monotonical relationship between seizure burden, per automated algorithms on POC-EEG, and an increased odds of unfavorable outcomes at discharge. Combining the outputs of the two different algorithms offered a more robust association. Future prospective studies should be done to explore this relationship.