Updating the TBI IMPACT Prognostic Model in a Diverse Observational Cohort
Naoki Takegami1, Abel Torres Espin1, Yoshihito Imagawa2, Itsunori Watanabe3, Shannon McWeeney4, Adam Ferguson1, H. E. Hinson1
1UCSF, 2Gifu University, 3Waseda University, 4OHSU
Objective:

Our primary aim was to externally validate the IMPACT model using a modern, diverse prospective observational cohort: “PROTIPS”.

Background:
The International Mission on Prognosis and Analysis of Clinical Trials in Traumatic Brain Injury (IMPACT) prognostic model predicts 6 months outcomes after moderate to severe traumatic brain injury (TBI).
Design/Methods:
We analyzed a cohort of subjects with TBI (GCS 3-12) and evidence of intracranial hemorrhage on initial head CT at a single institution with complete outcome data. We assessed all three versions of the IMPACT model - core, extended, and laboratory - in terms of 6-month mortality and unfavorable outcomes (defined as GOSE 1-4), comparing the original model and a model where the coefficients were refitted.
Results:

Almost one third of PROTIPS subjects were female (29%), and 22% were of races other than White. Of the cohort (n=126), 88 subjects had complete 6-month outcome data. There were no missing predictor variables. Significant variables correlating with 6-month GOSE were age (45.7±21.7 years, p<0.001), GCS Motor (5 [1-5]), p<0.005), and glucose (153±63.9 mg/dl, p<0.01). Using the original model, for mortality, model discrimination was: core model (AUC 0.70), extended model (AUC 0.58), and laboratory model (AUC 0.70). For unfavorable outcomes (GOSE 1-4), the performance was: core model (AUC 0.56), CT model (AUC 0.63), and lab model (AUC 0.61). Upon refitting for mortality, model discrimination was: core model (AUC 0.87 [95% CI: 0.86-0.89]), extended model (AUC 0.82 [95% CI: 0.79-0.84]), and laboratory model (AUC 0.84 [95% CI: 0.81-0.89]). For unfavorable outcomes, the performance was: core model (AUC 0.74 [95% CI: 0.65-0.80]), CT model (AUC 0.82 [95% CI: 0.76-0.89]), and lab model (AUC 0.72 [95% CI: 0.69-0.73]).

Conclusions:
Using the same predictor variables but refitting the coefficients allowed us to achieve improved AUCs comparable to the original coefficients. Predictive accuracy can vary across cohorts, necessitating caution in interpreting model outputs.
10.1212/WNL.0000000000205847