Predicting in-hospital mortality after traumatic brain injury: External validation of CRASH-basic and IMPACT-core in the national trauma data bank

• Published in: Injury 2021-02, Vol.52 (2), p.147-153
• Main Authors: Camarano, Joseph G., Ratliff, Hunter T., Korst, Genevieve S., Hrushka, Jaron M., Jupiter, Daniel C.
• Format: Article
• Language: English
• Subjects: External validation; Mortality; Prediction models; Prognosis; Traumatic brain injury
• ISSN: 0020-1383; 1879-0267
• DOI: 10.1016/j.injury.2020.10.051

•Both the CRASH-Basic and IMPACT-Core models accurately predict in-hospital mortality after both mild and moderate-severe TBI.•By demonstrating validity in the NTDB, these models appear generalizable to new data and diverse geographical settings.•The NTDB provides a valuable tool for large-scale validation and refinement of generalizable prognostic models for TBI patients. Traumatic brain injury (TBI) prognostic prediction models offer value to individualized treatment planning, systematic outcome assessments and clinical research design but require continuous external validation to ensure generalizability to different settings. The Corticosteroid Randomization After Significant Head Injury (CRASH) and International Mission on Prognosis and Analysis on Clinical Trials in TBI (IMPACT) models are widely available but lack robust assessments of performance in a current national sample of patients. The purpose of this study is to assess the performance of the CRASH-Basic and IMPACT-Core models in predicting in-hospital mortality using a nationwide retrospective cohort from the National Trauma Data Bank (NTDB). The 2016 NTDB was used to analyze an adult cohort with moderate-severe TBI (Glasgow Coma Scale [GCS] ≤ 12, head Abbreviated Injury Scale of 2–6). Observed in-hospital mortality or discharge to hospice was compared to the CRASH-Basic and IMPACT-Core models’ predicted probability of 14-day or 6-month mortality, respectively. Performance measures included discrimination (area under the receiver operating characteristic curve [AUC]) and calibration (calibration plots and Brier scores). Further sensitivity analysis included patients with GCS ≤ 14 and considered patients discharged to hospice to be alive at 14-days. A total of 26,228 patients were included in this study. Both models demonstrated good ability in differentiating between patients who died and those who survived, with IMPACT demonstrating a marginally greater AUC (0.863; 95% CI: 0.858 – 0.867) than CRASH (0.858; 0.854 – 0.863); p < 0.001. On calibration, IMPACT overpredicted at lower scores and underpredicted at higher scores but had good calibration-in-the-large (indicating no systemic over/underprediction), while CRASH consistently underpredicted mortality. Brier scores were similar (0.152 for IMPACT, 0.162 for CRASH; p < 0.001). Both models showed slight improvement in performance when including patients with GCS ≤ 14. Both CRASH-Basic and IMPACT-Core accurately predict in-hospital mortality fo