"Early Stratification of Risk for Poor Neurological Outcome After Cardiac Arrest Is Improved with Processed EEG Data".

AIM: To evaluate the impact of adding early processed quantitative EEG biomarkers to health record data for early neurological risk stratification after cardiac arrest using machine learning. METHODS: Data available during ICU admission after return of spontaneous circulation (ROSC) were collected from comatose patients, including the processed EEG metrics suppression ratio (SR) and bispectral index (BIS). Clinical data included demographics, Charlson Comorbidity Index, cardiac arrest and resuscitation details, admission vital signs, and initial laboratory results. Primary outcome waspoor Cerebral Performance Category score (CPC 3-5). Six machine learning models were developed to predict hospital discharge and 6-month long-term outcome with clinical data alone, EEG (BIS-SR) data alone, and combined clinical and EEG data. Two additional operating points (high specificity for poor outcome and for good outcome) were also calculated. Feature importance was analyzed to identify the most predictive variables. RESULTS: Among 913 patients, the median age was 59 years, most were male (69%), and 44% had an initial shockable rhythm. Poor outcome was observed in 70% at discharge and 72% at long-term assessment. The best-performing models for combined 6-hour EEG and clinical data revealed AUC 0.88 (0.87-0.90) for poor long-term outcome and 0.86 (0.84-0.87) for poor discharge outcome. Combining the processed EEG and clinical data significantly improved the AUC compared to either data set alone (p<0.001). CONCLUSION: The combination of early processed EEG and clinical data were best able to stratify neurological risk early after cardiac arrest using machine learning algorithms, but external validation is needed.