The study team evaluated the CBI-M Framework using data from two contemporary prospective observational studies: CENTER-TBI (4,509 participants) and TRACK-TBI (2,552 participants). The cohorts provided a broad clinical substrate across acute-care pathways in high-income settings. Investigators restricted analysis to endpoints that reflect routine clinical decision-making and longer-term patient-centered outcomes. The authors reported that the datasets did not contain three elements of the Modifier pillar; those missing elements were not inventored by the analysis and thus absent from model inputs.
Data indicated that full CBI-M models discriminated well for initial diagnostic and care-path endpoints. Discrimination metrics (AUC) for outcomes such as traumatic intracranial injury on CT, hospital admission, ICU admission, intracranial pressure monitoring, and major cranial surgery fell in a high-performance range in both cohorts. The proportion of explained variability, expressed with Nagelkerke’s pseudo-R2, was also substantial for these acute endpoints. The study team observed that individual pillar contributions varied: the Biomarker pillar showed the strongest association with CT-defined abnormalities, while the Imaging pillar carried the largest signal for predicting major cranial surgery. Clinical variables influenced admission and monitoring decisions in expected directions.
Investigators compared CBI-M prognostic output to established prediction models appropriate to each cohort’s target population — namely IMPACT-Lab and CRASH-CT. The full CBI-M model demonstrated larger pseudo-R2 values for 6-month mortality and unfavorable outcome (GOSE ≤ 4) than either comparator within their respective target populations. Reported discrimination for mortality and unfavorable GOSE was high: AUC estimates were near 0.89–0.95 across CENTER-TBI and TRACK-TBI, with accompanying pseudo-R2 values indicating substantial explained variability. The authors emphasized that the CBI-M approach is multimodal—integrating clinical, biomarker, and imaging data—potentially accounting for richer prognostic association compared with single-model inputs.
The cohort analyses showed divergent contributions across pillars when predicting 6-month outcomes. Clinical and patient-related factors were relatively more influential for longer-term recovery measures, including functional status captured by the Glasgow Outcome Scale-Extended. Biomarkers and imaging retained strong early predictive utility, particularly for diagnostic CT changes and acute surgical decision-making. For mortality prediction, the combined information across pillars yielded the highest explanatory power. The authors reported these relationships quantitatively using AUC and pseudo-R2 metrics; specific numerical values are provided in the source article for readers who require exact estimates.
The full CBI-M framework produced robust discrimination for mortality and unfavorable GOSE at 6 months. By contrast, model performance for health-related quality of life (HRQoL) was more modest. The study team observed that predictors relevant to HRQoL differed from those driving early clinical decisions and hard endpoints like mortality; patient-related modifiers and longer-term social and health determinants appeared important but were incompletely represented in the available datasets. The source article documents these patterns and reports that HRQoL remains harder to predict with the current set of pillar elements.
The authors noted several limitations inherent to the available data and analytic approach. Analyses were restricted to high-income settings, which limits generalizability to different health systems and resource environments. Three components of the Modifier pillar were absent from the datasets accessed and therefore omitted from the framework implementation in this work; the source did not report data for these items. The retrospective analytic framing within prospectively collected cohorts and some analyses raised potential for reverse causality. Finally, the authors called for further prospective validation and refinement — including prospective weighting or optimization of pillar elements — before recommending routine clinical implementation. These specific limitations close the report and mark the immediate next steps the study team identified for future work.