Objective:

To develop automated prediction models that identify post-cardiac arrest patients at risk for poor neurological outcomes using early clinical and EEG features.

Background:

Accurately predicting neurological recovery following cardiac arrest is challenging. Integrating multiple data sources through computational models may improve early prognostication and inform family counseling.

Design/Methods:

We analyzed 716 consecutive post-cardiac arrest patients admitted to two tertiary centers (2015-2022) who underwent continuous EEG within 72 hours post-arrest. Neurological outcome at discharge was classified as good (modified Rankin Scale ≤4 AND Cerebral Performance Category ≤2) versus poor. We developed six prediction models using 18 features: demographics (age, sex, insurance, ethnicity), cardiac arrest characteristics (initial rhythm, time to return of spontaneous circulation), clinical features (seizures, myoclonus), and EEG findings (background suppression, stimulus reactivity, eight hyperexcitable patterns including periodic discharges and rhythmic delta activity). Models were validated using data splitting and cross-validation with comprehensive performance assessment.

Results:

Of 716 patients, 608 (84.9%) had poor outcomes. The best-performing ensemble model achieved an AUC-ROC of 0.878 (95% CI, 0.807–0.941). When predicting poor outcome, specificity was 95.6% and sensitivity was 50.0%. Feature importance indicated older age, suppressed EEG background, prolonged time to return of spontaneous circulation, and absent EEG reactivity as the strongest adverse predictors.

Conclusions:

Automated models integrating early EEG and clinical data reliably identify cardiac arrest patients at highest and lowest risk for poor outcomes. Clinicians can use these tools alongside existing prognostic methods to counsel families about realistic recovery expectations and guide goal-concordant treatment decisions during the critical first 72 hours post-arrest.