Point-of-Care EEG Artificial Intelligence Measure of Seizure Burden Associates With Clinical Outcome at Discharge.

OBJECTIVES: Point-of-care (POC) electroencephalography (EEG) enabled with artificial intelligence (AI) algorithms hold the potential to address gaps in EEG access and interpretation. We assessed the clinical association of the patterns recognized by one of such systems.

DESIGN: Secondary cohort analysis of the retrospective multicenter Seizure Assessment and Forecasting with Efficient Rapid-EEG (SAFER-EEG) study.

SETTING: EEG and clinical data were gathered from three academic centers with access to POC EEG (Ceribell, Sunnyvale, CA) as part of their standard of care. We used a bedside seizure burden (SzB) monitor ("bedside" algorithm) and a more sensitive online algorithm ("portal" algorithm) to determine SzB. The modified Rankin Scale (mRS) score at discharge was selected as the outcome.

PATIENTS: Four hundred enrolled adult patients, 359 with complete outcome and clinical data.

INTERVENTIONS: Time in seizure and peak SzB assessed by AI.

MEASUREMENTS AND MAIN RESULTS: Per the bedside algorithm, 39.8% of patients had peak 5-minute SzB greater than 0%. With every additional hour of seizure detected by the bedside algorithm, patients were more likely to have unfavorable outcome (mRS > 3) at discharge (adjusted odds ratio [aOR], 1.98; 95% CI, 1.11-4.29). Compared with 5-minute SzB = 0%, prolonged activity per the bedside algorithm (peak 5-min SzB ≥ 90%) was associated with a 3.4-fold increase in aOR of poor outcome. Every 30 seconds of activity in the maximum hourly SzB of the bedside algorithm was associated with increased odds of unfavorable outcome (aOR, 1.02; 95% CI, 1.00-1.03). Combining outputs from the algorithms increased the strength of associations with outcome, particularly in patients with peak 5-minute SzB greater than or equal to 90% (aOR, 4.4; 95% CI, 1.66-12.69).

CONCLUSIONS: AI-estimated SzB is associated with functional outcomes at discharge in a dose-response fashion, even after controlling for clinical cofounds. This provides initial evidence of the clinical utility of AI algorithms for detecting clinically important EEG patterns.