Predicting ordinal GOSE using dynamic NICU physiological variables in moderate-to-severe TBI: A Temporal and Stratified Split Study

This is the authors' abstract. We don't add key points for this paper.

Results: Mortality discrimination was significantly higher in younger than in older adults (AUC 0.905 vs 0.791), regardless of the model.Differences between age groups were smaller for unfavorable outcome prediction (AUC 0.867 vs 0.800).In younger adults, calibration plots showed that IMPACT and CRASH models overestimated mortality and unfavorable outcomes, whereas the Helsinki model slightly underestimated risk.In older adults, all models except the CRASH refitted model for unfavorable outcome overestimated prognosis (optimistic predictions).Laboratory variables contributed less to outcome prediction than CT findings.Models based on individual CT features (CRASH original, Nijmegen, Helsinki) showed advantages in older adults.After local refitting of coefficients, model performance was comparable across approaches.Conclusions: All models demonstrated good discrimination, though performance declined in older adults.Observed outcomes were generally worse than predicted in this group.Locally adapted (refitted) versions improved calibration, supporting their use in clinical practice.