DeepCRI: Real-time EEG-based Prognostication after Cardiac Arrest
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Rationale
Accurate early prediction of neurological outcome after cardiac arrest is essential for guiding intensive care decisions. Visual EEG interpretation can support prognostication but is time-consuming and subjective.
Methods
We developed DeepCRI, a deep learning model trained on EEG recordings obtained 24 h after arrest from a prospective cohort of 522 patients across two centers. Labels were based on Cerebral Performance Category (CPC) scores, restricted to neurologically determined outcomes. Calibration was assessed using the expected calibration error (ECE). To enable dynamic prog-nostication, we derived boundaries for the 95% and 100% specificity thresholds (good and poor outcome, respectively) by applying the trained model to the training cohort at all time points between 0–36 h after arrest. Validation was performed in an independent cohort of 222 patients (60% good outcome), after automatic artifact rejection. Predictions were locked once DeepCRI values consistently exceeded the predefined boundaries for good or poor outcome. DeepCRI was implemented in Neuro-Center EEG for real-time bedside monitoring.
Results
In training, area under the receiver operating characteristic curve (ROC AUC) at 24 h was 0.97 with ECE = 0.049. In the validation cohort, outcome classification was achieved in 184/222 patients (83%) within 24 h; 38 patients (17%) remained in the gray zone. Sensitivity for good outcome was 93.9% (95% confidence interval [CI]: 88.5–96.9%) at 78.9% (95% CI: 67.1–84.6%) specificity. Sensitivity for poor outcome was 56.6% (95% CI: 47.1–65.7%) at 100% (95% CI: 97.1–100%) specificity. ROC AUC at 24 h was 0.91 (95% CI: 0.87–0.94). Median lock-in time was 11.2 h for good and 14.4 h for poor outcome.
Conclusion
DeepCRI is the first real-time deep learning model for prognostication of neurological outcome within 12-24 h after cardiac arrest. Its lock-in mechanism ensures individualized, stable, classifications. By embedding DeepCRI into routine EEG infrastructure, we demonstrate the technical feasibility of continuous bedside integration for comatose patients after cardiac arrest.
