Evaluating the Performance of the IMPACT Model in the Prehospital TXA for TBI Trial
H. E. Hinson1, Shannon McWeeney2, Adam Ferguson4, Martin Schreiber3, Susan Rowell5
1Neurology and Emergency Medicine, 2Medical Informatics and Clinical Epidemiology, 3Surgery, OHSU, 4Neurosurgery, UCSF, 5Surgery, University of Chicago
Objective:
To test the IMPACT model’s performance in a contemporary clinical trial population.
Background:
The International Mission on Prognosis and Analysis of Clinical Trials in Traumatic Brain Injury (IMPACT) prognostic model predicts functional outcome after moderate-severe traumatic brain injury (TBI). Our previous work suggested that glial fibrillary acidic protein (GFAP) may improve the discrimination of outcome when added to prognostic models.
Design/Methods:
We analyzed data from a subset subjects from the phase II double-blind, multicenter randomized controlled trial, “Prehospital Tranexamic Acid Use for Traumatic Brain Injury.” This subset was limited to subjects in the placebo arm of the parent trial with evidence of hemorrhage on initial head computed tomographic (CT) scan. All 3 versions (core, extended, lab) of the IMPACT model were evaluated against 6-month mortality, plus an exploratory 4th version (lab+GFAP). We compared our results to previously published validation results from the CENTER-TBI cohort.
Results:
Within the subset (n=166), 127 subjects had complete 6-month outcome data. Predictor variables were missing <10% (0-9.4%) of cases, and were imputed (GCS motor score, pupillary reactivity, serum glucose, serum hemoglobin, hypoxia, and Marshall CT score). GFAP was associated with mortality (p=0.004). The core model discriminated poorly in the subset (AUC 0.60 [95% CI: 0.48-0.71], p=0.096). The extended model discriminated well (AUC 0.73 [95% CI: 0.64-0.82], p<0.001). The lab model discriminated non-significantly better (AUC 0.76 [95% CI: 0.66-0.86], p=0.167). Adding GFAP to the lab model improved discrimination further (AUC 0.83 [95% CI: 0.76-0.90], p=0.004). All 4 models overestimated mortality compared with the observed rate (n=31, 24% v. 26-82%).
Conclusions:
In comparison with the observational CENTER-TBI validation cohort, the IMPACT models (core/extended/lab) discriminated mortality less well in a subset of the TXA trial. The addition of GFAP improved discrimination for mortality, but further work is needed to optimize its addition to the lab model.