Cognitive-motor dissociation and time to functional recovery in patients with acute brain injury in the USA: a prospective observational cohort study.
Level 3 - non-randomized controlled study
Prospective observational derivation and validation cohort study of prognostic factors
PubMed 35841909 · doi:10.1016/S1474-4422(22)00212-5
What was done
Researchers prospectively enrolled clinically unresponsive adult patients with acute brain injury across two cohorts (100 derivation, 93 validation; screened n=598, analyzed n=193). Machine learning was applied to EEG recordings during spoken motor commands to diagnose cognitive-motor dissociation (CMD; brain activation without behavioral response). Functional outcomes were assessed via the Glasgow Outcome Scale-Extended (GOS-E) at discharge and at 3, 6, and 12 months. Survival models and shift analyses evaluated the association between CMD and recovery, treating death as a competing risk and censoring withdrawal of life-sustaining therapy.
What was found
CMD was identified in 27 of 193 patients (14%). At 12 months, 28 of 193 patients (15%) achieved a GOS-E score of 4 or higher. CMD independently predicted a shorter time to good recovery (hazard ratio 5.6 [95% CI 2.5-12.5]), alongside traumatic brain injury or subdural hematoma (HR 4.4 [95% CI 1.4-14.0]), admission Glasgow Coma Scale score >= 8 (HR 2.2 [95% CI 1.0-4.7]), and younger age (HR 1.0 [95% CI 1.0-1.1]). Among patients discharged home or to rehabilitation, CMD was associated with higher GOS-E scores as early as 3 months post-injury (odds ratio 4.5 [95% CI 2.0-33.6]).
Why it matters
Early machine-learning EEG detection of covert command-following provides a strong prognostic indicator in clinically unresponsive patients with acute brain injury, helping clinicians identify candidates likely to benefit from aggressive rehabilitation and aiding family counseling.
Limits
Only 193 of 598 screened patients (32%) were included, introducing potential selection bias. The total number of patients with CMD was small (n=27), yielding wide confidence intervals for key effect estimates. The specialized machine-learning EEG methodology may limit generalizability across centers lacking advanced neuro-monitoring infrastructure.