An Uncertainty-Aware Neuroprognostication Pipeline Using Longitudinal Brain Continuity Index and Leave-One-Center-Out Validation

Abstract Neuroprognostication in intensive care is high-stakes, where overconfident errors may contribute to premature withdrawal of life-sustaining therapy. To improve reproducibility and efficiency over raw EEG modeling, we propose an uncertainty- aware multimodal framework based on the Brain Continuity Index (BCI), a compact EEG-derived representation of background continuity. A dual-branch late-fusion CNN–BiGRU integrates 72-hour longitudinal BCI trajectories with static clinical covariates. In a hospital-level leave-one-center-out evaluation on 607 patients, the model achieved a pooled AUROC of 0.76 and F1-score of 0.80. We observe a performance ceiling associated with the compressed BCI represen- tation: deep sequence modeling was statistically equivalent to classical ensembles ( p = 0 . 413 ), suggesting most prognostic signal is captured by summary-level dynamics. Using Monte Carlo dropout, predictive entropy enabled selective prediction, reaching ≥ 85% accuracy on high-confidence cases while deferring uncertain cases for expert review. These results highlight multicenter robustness and uncertainty-aware triage as key elements for trustworthy AI in neurocritical care.