Importance: There are currently no models that predict long-term functional dependency in patients with disorders of consciousness (DoC) after traumatic brain injury (TBI).
Objective: Fit, test, and externally validate a prediction model for 1-year dependency in patients with DoC 2 or more weeks after TBI.
Design: Secondary analysis of patients enrolled in TBI Model Systems (TBI-MS, 1988-2020, Discovery Sample) or Transforming Research and Clinical Knowledge in TBI (TRACK-TBI, 2013-2018, Validation Sample) and followed 1-year post-injury.
Setting: Multi-center study at USA rehabilitation hospitals (TBI-MS) and acute care hospitals (TRACK-TBI).
Participants: Adults with TBI who were not following commands at rehabilitation admission (TBI-MS; days post-injury vary) or 2-weeks post-injury (TRACK-TBI).
Exposures: In the TBI-MS database (model fitting and testing), we screened demographic, radiological, clinical variables, and Disability Rating Scale (DRS) item scores for association with the primary outcome.
Main outcome: The primary outcome was death or complete functional dependency at 1-year post-injury, defined using a DRS-based binary measure (DRS Depend ), indicating need for assistance with all activities and concomitant cognitive impairment.
Results: In the TBI-MS Discovery Sample, 1,960 subjects (mean age 40 [18] years, 76% male, 68% white) met inclusion criteria and 406 (27%) were dependent at 1-year post-injury. A dependency prediction model had an area under the receiver operating characteristic curve (AUROC) of 0.79 [0.74, 0.85], positive predictive value of 53%, and negative predictive value of 86% for dependency in a held-out TBI-MS Testing cohort. Within the TRACK-TBI external validation sample (N=124, age 40 [16], 77% male, 81% white), a model modified to remove variables not collected in TRACK-TBI, had an AUROC of 0.66 [0.53, 0.79], equivalent to the gold-standard IMPACT core+CT score (0.68; 95% AUROC difference CI: -0.2 to 0.2, p=0.8).
Conclusions and relevance: We used the largest existing cohort of patients with DoC after TBI to develop, test and externally validate a prediction model of 1-year dependency. The model’s sensitivity and negative predictive value were greater than specificity and positive predictive value. Accuracy was diminished in an external sample, but equivalent to the best-available models. Further research is needed to improve dependency prediction in patients with DoC after TBI.