Predicting Emergence from Post-Traumatic Amnesia Within Inpatient Rehabilitation: How Much Is Recovery and How Much Is Discharge Timing?

In this work, we address a critical yet overlooked methodological challenge in traumatic brain injury (TBI) rehabilitation research by decomposing true cognitive recovery from the care delivery timeline that bounds it. Using a nationwide cohort of 17,895 participants from the TBIMS database, we demonstrate that models predicting post-traumatic amnesia (PTA) emergence predominantly capture administrative discharge timing rather than biological recovery alone. Specifically, a model trained solely to predict rehabilitation length of stay reaches a concordance of C=0.793, compared to 0.857 for our primary survival model, revealing that roughly three-quarters of the above-chance predictive discrimination is shared with discharge timing and leaving an emergence-specific increment of only ΔC=0.064. Furthermore, we provide empirical evidence that administrative discharge censoring is strongly covariate-dependent (AUC=0.866), showing that standard reporting conventions introduce meaningful distortion: complete-case deletion underestimates the median emergence time at 19 days, while marginal Kaplan–Meier overestimates it at 26 days, compared to adjusted estimators that place it at 23–24 days. Finally, we establish the first registry-scale censoring-aware survival machine learning benchmark for PTA emergence across 20 repeated splits, documenting that while gradient-boosted trees significantly outperform tuned classical Cox models (p<10⁻²⁰), the marginal performance gain is modest (ΔC=0.012), demonstrating that algorithmic complexity provides limited advantage over standard baselines and highlighting the necessity of decomposing episode-bounded clinical outcomes before interpreting headline performance.