The performance ceiling: why clinical data is insufficient for precision prognosis in concussion

The transition to active rehabilitation in concussion care requires precise tools to identify patients at risk of persistent post-concussive symptoms (PPCS). While machine learning (ML) offers the potential to personalize prognosis, current models relying on clinical history and subjective symptom reporting (e.g., SCAT5) have failed to demonstrate significant performance gains over the last decade. This perspective article argues that clinical prognostic models have reached a performance ceiling of approximately 0.85 Area Under the Curve (AUC). By reviewing key studies from 2016 to 2025, we demonstrate that increasing algorithmic complexity—from logistic regression to deep learning—yields diminishing returns when applied to subjective inputs. In contrast, models incorporating physiological data, such as neuroimaging or fluid biomarkers, consistently break this ceiling, achieving AUCs exceeding 0.95. We conclude that better mathematics cannot correct for missing biological signal, and that the advancement of precision medicine in neurotrauma requires a fundamental shift toward multimodal, biological data integration.