Traumatic brain injury prognostic models become more informative as pillars are added
The performance of traumatic brain injury prognostic models rose as clinical data were joined by biomarker, imaging, and modifier variables in large CENTER-TBI and TRACK-TBI cohorts. The pattern was not uniform across endpoints. Some outcomes responded strongly to imaging, while others drew most of their signal from the bedside clinical examination, a reminder that the biology of head injury is not captured by one domain alone.
These figures map a layered prediction strategy built around four pillars: clinical, biomarker, imaging, and modifier. In practice, that structure mirrors how acute neurotrauma is actually assessed. A neurologic exam provides the first signal of injury severity, circulating proteins capture tissue disruption, CT identifies structural damage, and modifiers account for background factors that can shift recovery in either direction. The model comparison across diagnostic, care, path, and outcome endpoints shows why a single narrow lens leaves prognostic information on the table.
Imaging carried the heaviest independent weight for surgical and structural endpoints
The strongest signal in the pillar analysis appeared where anatomy and intervention intersected. For major cranial surgery within 72 hours, the imaging pillar added a substantial amount of previously unexplained variation once the clinical, biomarker, and modifier pillars were already in the model. In contrast, the biomarker pillar contributed little additional information after those other domains were included. One figure makes the imbalance plain: structural injury seen on CT mattered far more than circulating markers for explaining the need for early neurosurgical action.
That pattern is biologically plausible. Surgery is driven by mass effect, hemorrhage, midline shift, and other visible lesions that are directly interrogated on imaging. Biomarkers such as GFAP, UCH-L1, and S100B reflect cellular injury, but they do not specify whether the injury has crossed the threshold into a lesion that requires urgent decompression or evacuation. In other words, the blood can confirm that tissue has been damaged, but the scan often determines whether the next step is the operating room.
Clinical variables still anchored the earliest models
Across the cumulative models, the clinical pillar formed the base layer before any other data were added. That makes sense in acute care, where Glasgow Coma Scale performance, pupillary findings, and related bedside observations remain the fastest available indicators of neurologic compromise. For some endpoints, especially those tied to broad outcome prediction, the clinical model already captured a meaningful portion of the variance before biomarkers or imaging were entered.
The heatmap of pairwise correlations showed that pillar components and endpoints were related in different ways across the two cohorts. Some relationships were stronger, others weaker, and the direction of the information flow was not redundant. A low Glasgow Coma Scale score does not describe the same biological state as intracerebral hemorrhage on CT, and neither is equivalent to a modifier such as a background vulnerability that may alter recovery. The point is not that clinical assessment is outdated. The point is that it remains incomplete on its own.
Outcome prediction improved in stages rather than all at once
Cumulative model performance in both CENTER-TBI and TRACK-TBI followed a stepwise pattern. The addition of biomarker and imaging pillars increased discrimination and the proportion of explained variability for several endpoints, though the extent of improvement varied by outcome. For the diagnostic endpoint, the imaging pillar was excluded from the cumulative models, preserving the focus on the information available before structural imaging is incorporated. That design choice separates early triage from downstream anatomic confirmation.
The sunburst plots make the partition of prognostic information easier to see. Within the full CBI-M model, each pillar contributed a unique share of explained variation depending on the endpoint. For major cranial surgery within 72 hours, imaging occupied a large outer sector, while the biomarker contribution could shrink to a near-null partial pseudo-R2 after the other pillars were already in place. The inverse can happen in other settings where laboratory markers help distinguish injury burden beyond what the exam and CT already show, but the figures make clear that the incremental value is endpoint-specific rather than universal.
Existing prognostic tools were not automatically outperformed
The full CBI-M cumulative models were compared with established prognostic frameworks, including IMPACT Lab and CRASH CT. Across CENTER-TBI and TRACK-TBI, the bar plots of Nagelkerke pseudo-R2s showed that adding pillars could improve model performance relative to these reference approaches. The comparison is important because it places the pillar strategy beside tools already familiar to neurotrauma clinicians.
IMPACT Lab incorporates age, Glasgow Coma Scale motor score, pupils, hypoxia, hypotension, CT classification, traumatic subarachnoid hemorrhage on CT, epidural mass on CT, glucose, and hemoglobin. CRASH CT includes age, Glasgow Coma Scale score, pupils, extracranial injury, and a set of CT findings such as petechial hemorrhages, obliteration of the third ventricle or basal cisterns, traumatic subarachnoid hemorrhage, midline shift, and nonevacuated hematoma. The pillar approach does not discard that clinical logic. It reorganizes it, then asks whether adding biomarkers and modifiers improves predictive yield beyond the established framework.
The answer varied by endpoint, but the direction was consistent: richer models were generally better models. Still, the figures do not support a simplistic claim that more variables automatically solve prognostication. Some pillars are powerful for one decision and marginal for another, which is exactly what would be expected in a disorder as heterogeneous as traumatic brain injury.
Mixed signals from biomarkers point to the limits of blood-based prediction
The biomarker pillar clearly belongs in the model stack, but the figures also show its limitations. In endpoints closely tied to visible structural injury or acute procedural decisions, biomarker data contributed less additional explanatory power than imaging. That is not surprising. Blood proteins are indirect surrogates of neuronal and glial injury, and their value depends on timing, lesion type, and the clinical question being asked.
By contrast, modifier variables appear to help explain why patients with similar initial exams and similar scans do not always follow the same recovery path. The modifier pillar likely captures context around the injury rather than the injury itself. That distinction is central in neurotrauma, where the acute lesion is only part of the story and downstream function may depend on host factors the first CT cannot reveal.
The figures leave one practical message for neurocritical care: prognostic precision improves when bedside findings, blood-based injury markers, imaging, and modifiers are interpreted together, but the dominant pillar shifts with the endpoint. A surgical decision still leans heavily on CT anatomy. Longer-term outcome prediction draws from a broader mix. The value of the model lies in respecting that difference rather than flattening traumatic brain injury into a single score.