Improving prognostic models of six-month clinical outcomes after severe traumatic brain injury with daily inpatient biomarkers: A Bayesian modelling approach

Abstract

Background

A key limitation of the IMPACT model for prognostication after severe traumatic brain injury (TBI) is the use of predictors from hospital admission only. We sought to identify if including daily blood labs (e.g., glucose, sodium, platelets, hemoglobin) and other vitals (e.g., heart rate, mean arterial pressure [MAP], partial pressure of carbon dioxide [PaCO2]) for the first 2 weeks post-severe TBI improves prognostication compared to the IMPACT model alone.

Methods

This is a secondary analysis of a prospectively collected database of patients from a single level 1 trauma center between November 2002 and December 2018 (n = 315). All patients had severe TBI at presentation, defined as Glasgow Coma Scale (GCS) ≤8. Researchers extracted daily blood labs and vitals for the first 14 days post-injury. We used Naïve Bayes to estimate class-conditional probabilities for an “IMPACT-only” model and a “full” model with the IMPACT score plus the biomarkers measured on post-injury days 1–13. The top ten predictors were included in the full model. DeLong's test assessed whether the difference in area under the curve (AUC) were significant (p < 0.05).

Results

The full model to predict unfavorable outcomes at six-months had significantly better discrimination (AUC = 0.83) compared to the IMPACT model (AUC = 0.74; p < 0.01). The full model to predict death by six-months had significantly better discrimination (AUC = 0.83) compared to the IMPACT model (AUC = 0.75; p = 0.02).

Conclusions

Biomarkers typically collected as part of inpatient clinical workups over the first two weeks post-injury improved discrimination of unfavorable outcomes and mortality by six-months compared to IMPACT.