Predicting Outcome After Traumatic Brain Injury in Bouaké, Côte d’Ivoire: Performance of IMPACT, CRASH-Basic and a Simplified Glasgow-Pupils Score in a Hospital Series of 772 Patients

External Validation of Prognostic Models in a Resource-Limited Neurosurgical Setting

Abstract

Background: Prognostic scores for traumatic brain injury (TBI) were built in wealthy countries. Few have been tested in sub-Saharan
Africa.
Objective: To test, in Bouaké, IMPACT (computed from available variables), CRASH-Basic and a simplified score, the GCSPupils adapted to anisocoria (adapted GCS-P), and to explore local recalibration.
Methods: Retrospective analysis of the consecutive-admissions register of 772 patients (Bouaké University Hospital, 2025–2026), following TRIPOD. Primary endpoint: in-hospital death (36 deaths, all by day 14). Secondary endpoint: unfavourable outcome (GOS 1–3, median follow-up 3.2 months). Discrimination: AUC, DeLong test. Calibration: intercept, slope, expected/observed ratio, Brier score compared with a null model. Analyses: whole cohort, moderate-to-severe TBI (IMPACT’s domain), adults; post hoc additions.
Results: In-hospital mortality was 4.7%. IMPACT overestimated risk (expected/observed ratio 3.4; 2.0 in moderate-to-severe TBI), discriminated poorly (AUC 0.64; 0.66), with a Brier score above the null model’s. The adapted GCS-P (AUC 0.79), GCS alone (0.78) and CRASH-Basic (0.78; expected/observed 1.46) discriminated better. Anisocoria added almost nothing to the GCS. In moderate-to-severe TBI the difference with IMPACT was not significant (p = 0.16).
Conclusions: IMPACT clearly overestimates risk: its horizon (six months) is not that of our endpoint, and 56% of patients have mild
TBI. The GCS and CRASH-Basic rank patients better, CRASH-Basic being less miscalibrated. Calibration of the adapted GCS-P remains to be evaluated. Multicentre validation is needed.