Abstract
AIMS: Hypoxic-ischemic brain injury drives poor outcomes after pediatric cardiac arrest, highlighting the need for early prognostication. This study evaluates whether machine learning models using a high-dimensional set of quantitative EEG (qEEG) features improve prediction of unfavorable neurologic outcome compared to a previously studied 7-feature model. We also assessed performance stability over time and the added value of clinical variables. METHODS: Single-center retrospective cohort study of children aged 3 months to 18 years who experienced cardiac arrest and received EEG monitoring within 24 h post-arrest. Patients with pre-arrest Pediatric Cerebral Performance Category (PCPC) >3 were excluded. Unfavorable outcome was defined as death or PCPC ≥4 at hospital discharge or 30 days post-arrest. We extracted 164 qEEG features and trained models using three established algorithms. Performance was evaluated using area under the ROC curve (AUROC). RESULTS: Seventy patients were included (median age 7.0 years, IQR 1.5-11.5); 53 % had unfavorable outcomes. Models using 164 qEEG features outperformed the 7-feature model: LASSO [0.81 (95 % CI: 0.69-0.91) vs 0.45 (0.31-0.58)] and Random Forest [0.80 (0.67-0.90) vs 0.65 (0.50-0.78)]. Adding clinical variables did not improve performance. AUROCs were stable across 6-h epochs from 6 to 24 h. Higher phase locking value, fractal exponent, and coherence were associated with better outcomes; higher delta power and suppression ratio variability were associated with worse outcomes. CONCLUSIONS: Data-driven models using 164 qEEG features accurately predicted neurologic outcomes after pediatric cardiac arrest, with stable performance over time. Future work includes external validation to assess generalizability.