A cross-study analysis for reproducible sub-classification of traumatic brain injury

Abstract

Objective: To identify reproducible sub-classes of traumatic brain injury (TBI) that correlate with patient outcomes. Methods: Two TBI datasets from the Federal Interagency Traumatic Brain Injury Research (FITBIR) Informatics System were utilized, Transforming Research and Clinical Knowledge in Traumatic Brain Injury (TRACK-TBI) Pilot and Citicoline Brain Injury Treatment Trial (COBRIT). Patients included in these analyses had closed head injuries with Glasgow Comas Scale (GCS) scores of 13-15 at arrival at the Emergency Department (ED). Sparse hiearchical clustering was applied to identify TBI sub-classes within each dataset. The reproducibility of the sub-classes was evaluated by investigating similarities in clinical variable profiles and patient outcomes in each sub-class between the two datasets, as well as by using a statistical metric called in-group proportion (IGP). Results: Seven TBI sub-classes were identified in the first dataset. There were between-class differences in patient outcomes at 90 days (Glasgow Outcome Scale Extended (GOSE): p < 0.001) and 180 days (Trail Making Test (TMT): p = 0.03). Four of seven sub-classes were reproducible in the second dataset with very high IGPs (94, 100, 99, 97%). Seven TBI sub-classes were also identified in the second dataset. There were significant between-class differences in patient outcomes at 180 days (GOSE: p = 0.024; Brief Symptom Inventory (BSI) p = 0.007; TMT: p < 0.001). Three of seven sub-classes were reproducible in the second dataset with very high IGPs (100% for all). Conclusions: Reproducible TBI sub-classes were identified across two independent datasets, suggesting that these sub-classes exist in a general population. Differences in patient outcomes according to sub-class assignment suggest that this sub-classification could be used to guide post-TBI prognosis.