Picking up exactly where tiered severe-TBI management stopped, the central question changes from whether we can lower the ICP to what the patient might become if they survive. This is the most consequential transition in traumatic brain injury care because treatment decisions are now being made in a system with asymmetric error. Continuing life support through a period of uncertainty may prolong treatment that the patient would ultimately have declined. Withdrawing life support after an inaccurately pessimistic prediction, however, removes the possibility of discovering that the prediction was wrong. Neuroprognostication after severe TBI must therefore be treated as a longitudinal, probabilistic process rather than a verdict produced by one examination, one CT, or one calculator.
The first discipline is to define the outcome being predicted. Survival, recovery of consciousness, functional independence, cognitive recovery, return to work, and a life the patient would regard as acceptable are not interchangeable. A patient may recover consciousness yet remain dependent for basic activities. Another may walk and feed themselves while losing executive function, emotional regulation, or the ability to resume a valued professional role. Conversely, clinicians and families may rate a physical disability as intolerable while survivors later report satisfactory quality of life. Even “favorable outcome” is unstable across the literature: some studies define it as a Glasgow Outcome Scale–Extended score of at least 4, meaning the patient can be left alone at home for part of the day but may remain dependent outside the home, whereas others require a GOSE of at least 5. Prognosis is meaningless unless the endpoint and time horizon are stated explicitly.
The physiology explains why the early behavioral examination is such an imperfect proxy for outcome. Following a command requires intact arousal, language comprehension, attention, working memory, motor planning, descending motor pathways, peripheral neuromuscular execution, and sufficient motivation and endurance to produce an observable response. Severe TBI can interrupt any link in that chain. Traumatic axonal injury can disconnect thalamocortical and frontoparietal networks without destroying all cortical function. Mesencephalic or thalamic injury can impair arousal. Aphasia, apraxia, abulia, akinetic mutism, critical-illness weakness, cervical cord injury, peripheral trauma, or hearing loss can prevent a conscious patient from demonstrating command-following. Sedatives, antiseizure medications, fever, metabolic derangement, hydrocephalus, seizures, sleep disruption, and ongoing intracranial hypertension further suppress the observable examination. “Not following commands” is therefore a behavioral description, not a diagnosis of unconsciousness and certainly not a prediction of permanent dependence.
This is why timing matters. A patient examined twelve hours after a long operation under propofol, fentanyl, neuromuscular blockade, and hyperosmolar therapy is not simply the same patient examined without those exposures. Midazolam, fentanyl, and especially pentobarbital can accumulate after prolonged infusions, with clearance further delayed by hypothermia, obesity, hepatic dysfunction, renal dysfunction, and drug interactions. Residual paralysis can abolish motor output without altering consciousness. Pupils are relatively resistant to many sedatives but can still be confounded by opioids, anticholinergic exposure, ocular trauma, third-nerve injury, prior ophthalmologic disease, seizures, and direct brainstem injury. The prognostic examination begins only after asking what physiology and pharmacology are still masking it.
The 2024 Neurocritical Care Society guideline formalized this caution. It recommends considering the complete clinical condition rather than any single variable and avoiding definitive prognostication based only on ultra-early characteristics during the first three days. Unless there is irreversible brainstem destruction with imminent death or documented prior wishes against treatment under uncertainty, the panel suggests at least three days of full support and, when possible, one to two weeks before attempting a definitive poor-outcome prognosis. Even after two weeks, substantial uncertainty may remain. This does not mean that clinicians should avoid prognosis conversations for two weeks. It means that early conversations should describe injury severity, immediate threats, and uncertainty without pretending that an irreversible long-term outcome has already declared itself. Among the individual variables evaluated, only accurately assessed bilateral pupillary nonreactivity on admission reached the guideline’s category of a moderately reliable predictor, and even that was not considered fully reliable or sufficient by itself. Age, admission GCS or motor score, unilateral pupillary nonreactivity, hypotension, hypoxemia, elevated ICP, acute kidney injury, hypernatremia, and post-traumatic infarction were not considered reliable stand-alone predictors. The CRASH and IMPACT models were classified as moderately reliable estimates, but their results must be presented with substantial uncertainty. The guideline’s most important contribution is therefore not a new calculator but a higher evidentiary threshold for declaring poor outcome.
At the bedside, the serial examination remains more informative than the worst isolated examination. The admission GCS may capture the combined effects of primary injury, intoxication, shock, hypoxemia, seizures, medications, and prehospital paralysis. Its motor component becomes more meaningful after resuscitation and clearance of confounders, but even a true post-resuscitation motor score of 1 is a severity marker rather than a deterministic endpoint. What matters is the trajectory: recovery of pupillary reactivity, transition from extension to flexion or localization, emergence of visual fixation or pursuit, reproducible movement to command, and increasing consistency across examinations. Conversely, loss of previously present responses, new bilateral pupillary abnormalities, or progressive loss of brainstem reflexes should prompt a search for a new lesion or physiologic deterioration before being interpreted as natural prognostic evolution.
Once the patient remains behaviorally unresponsive beyond the initial unstable period, the Coma Recovery Scale–Revised is more appropriate than repeatedly documenting a GCS. It separates reflexive from cognitively mediated behavior across auditory, visual, motor, verbal, communication, and arousal domains. It also forces the examiner to test visual pursuit, object use, localization, and communication systematically rather than equating absent command-following with an unresponsive wakefulness syndrome. Because arousal fluctuates, repeated assessments at different times are more reliable than a single examination. The established disorders-of-consciousness guideline continues to recommend treating confounders, optimizing arousal, and using serial standardized assessments; traumatic etiology and a minimally conscious state are associated with more favorable recovery than nontraumatic injury and unresponsive wakefulness, but neither guarantees a particular outcome. These principles remain clinically relevant even though the formal guideline addresses prolonged disorders of consciousness rather than the first ICU week.
CT describes anatomy urgently but predicts the future only incompletely. Obliterated cisterns, substantial midline shift, intraventricular or subarachnoid blood, a nonevacuated mass lesion, and diffuse swelling all increase population-level risk. Yet the same CT feature can represent very different biologic states. Midline shift from a rapidly evacuated epidural hematoma is not equivalent to shift from bilateral contusions, diffuse swelling, and brainstem injury. The Marshall classification describes lesion phenotype and surgical mass effect; the Rotterdam, Helsinki, and Stockholm scores improve statistical risk stratification by combining additional CT features. The 2024 guideline nevertheless found none of these CT systems reliable as a stand-alone predictor of six-month mortality or functional outcome. Their calibration is imperfect, many validation cohorts included milder injuries, and most studies could not separate death caused by the injury from death following treatment withdrawal.
MRI is most valuable when it answers why the examination is worse than the CT would suggest. Susceptibility-weighted imaging can reveal hemorrhagic axonal lesions, diffusion imaging can identify nonhemorrhagic axonal injury and acute ischemia, and brainstem, thalamic, callosal, internal-capsule, or widespread white-matter injury can refine the lesion phenotype. But “grade III diffuse axonal injury” should not be translated into inevitable nonrecovery. Lesion location, bilaterality, burden, accompanying hypoxic–ischemic injury, and the integrity of arousal and motor networks matter more than the label alone. Quantitative diffusion-tensor approaches are promising: a multicenter MRI-COMA and CENTER-TBI study using deep-white-matter diffusion metrics obtained between days 7 and 35 achieved high specificity for unfavorable and favorable one-year outcomes in validation cohorts. That result supports the physiology that distributed white-matter integrity matters, but scanner harmonization, specialized processing, transport selection, dichotomized outcomes, and incomplete sensitivity prevent DTI from functioning as an independent bedside withdrawal criterion. MRI currently improves phenotyping more reliably than it provides certainty.
EEG contributes another physiologically distinct layer. A continuous background, preserved reactivity, sleep architecture, posterior dominant rhythm, and greater alpha variability imply that thalamocortical networks retain organization. Suppression, discontinuity, and absent reactivity may reflect more severe network failure, but they can also be produced by sedatives, barbiturates, hypothermia, or metabolic illness. A 2025 systematic review found associations between better outcome and continuity, reactivity, sleep features, and alpha variability, whereas seizures and epileptiform discharges were not consistently associated with long-term outcome. Every included study, however, carried moderate-to-high risk of bias because EEG definitions, timing, outcome definitions, and adjustment for confounders varied substantially. EEG can narrow uncertainty and detect treatable seizures, but no isolated TBI EEG pattern currently has the validated specificity required for an irreversible prognosis.
Somatosensory evoked potentials require similar restraint. Bilateral absence of cortical N20 responses has a different evidentiary foundation in post–cardiac arrest hypoxic–ischemic injury than in heterogeneous trauma. Peripheral nerve injury, plexopathy, cervical cord injury, technical factors, decompressive surgery, and focal thalamocortical disruption can all complicate interpretation. Present responses demonstrate preserved pathway conduction and may be reassuring, but absent responses after TBI should not be imported uncritically into the cardiac-arrest prognostication framework.
Quantitative pupillometry converts a subjective examination into a reproducible trajectory, which is useful, but reproducibility does not automatically create prognostic certainty. A 2025 single-center study of 131 patients found that a greater frequency of abnormal pupil measurements during the first 72 hours was associated with unfavorable discharge disposition and modestly improved the fit of an IMPACT-based model. The cohort included the full TBI severity spectrum, the outcome combined death, hospice, and long-term care, and the study did not establish a threshold capable of reliably predicting long-term function. The practical interpretation is that repeated bilateral abnormalities deserve more weight than one isolated low NPi, particularly when the abnormalities persist after correction of ICP and other confounders. They remain a dynamic severity signal, not a permission slip for withdrawal. The study is hypothesis-generating rather than a validated individual decision rule.
Serum biomarkers occupy a similar position. GFAP, UCH-L1, neurofilament light, S100B, and tau can reflect glial, neuronal, or axonal injury and may improve population-level prediction when added to clinical models. Their concentrations are affected by sampling time, extracranial injury, renal clearance, assay platform, and the mixture of focal and diffuse injury. No biomarker met the evidence requirements for inclusion as a reliable individual predictor in the 2024 guideline. A biomarker can improve the area under a receiver-operating curve without achieving the near-zero false-positive rate required for a decision that will lead to death.
CRASH and IMPACT are therefore best used as calibration tools against intuition. IMPACT begins with age, motor response, and pupils, then adds hypotension, hypoxemia, CT characteristics, glucose, and hemoglobin in its expanded versions. CRASH similarly combines clinical severity with CT and extracranial injury. They can expose an attending whose subjective estimate is implausibly optimistic or pessimistic, but they do not incorporate many aspects of the subsequent trajectory, modern rehabilitation, patient-specific reserve, or the exact treatment-withdrawal practices embedded in their derivation cohorts. Their output should be expressed as an approximate population probability with a confidence range, never as “the patient has an 80% chance of a bad outcome.” Even a well-calibrated 80% estimate means that approximately one in five comparable patients does not experience the predicted outcome, and the model cannot identify which individual is that one.
Advanced detection of consciousness demonstrates why absence of behavioral output is especially dangerous to overinterpret. In the 2024 multinational study of 353 patients with disorders of consciousness, task-based EEG or functional MRI detected cognitive motor dissociation in 60 of 241 patients—25%—who showed no observable response to commands. That finding does not mean that one quarter of all acutely comatose TBI patients have covert consciousness. The cohort was a convenience sample from specialized centers, the median time since injury was 7.9 months, only one quarter were assessed within 28 days, methods varied by site, and agreement between EEG and fMRI was low. More than 60% of patients who could follow commands behaviorally also failed to demonstrate command-following on the advanced tests, underscoring their substantial false-negative problem. A positive result can reveal preserved cognition that the motor examination missed; a negative result cannot establish absence of consciousness. The study changes the conceptual ceiling of what may be hidden, not the current standard for acute WLST decisions.
More recent acute-TBI data add an important counterweight to both enthusiasm and nihilism. A 2026 prospective study enrolled 55 patients with acute severe TBI in a multimodal protocol using CRS-R, task-based EEG and fMRI, passive-language paradigms, and default-mode-network imaging. Among the 45 patients with six-month outcomes, behavioral evidence of consciousness or command-following on the ICU CRS-R was associated with better Disability Rating Scale outcomes. Advanced EEG or fMRI command-following did not improve that association, and cognitive motor dissociation detected in non-command-followers was not independently associated with outcome in this small cohort. Ten patients died, all after withdrawal of life support, which both limits the analysis and illustrates the self-fulfilling-prophecy problem. For now, careful standardized behavior remains the foundation, while advanced testing remains an adjunct whose incremental prognostic value requires larger validation.
The natural-history data are the strongest antidote to premature certainty. In prospective TRACK-TBI data, only 12.4% of patients with severe TBI had reached GOSE 4 or better at two weeks, but 45% had done so by three months and 52.4% by twelve months. Good recovery, defined as GOSE 7 or 8, rose from less than 1% at two weeks to 22.9% at one year. Among 79 patients classified as being in a vegetative state at two weeks, 62 survived to one year; all of those survivors recovered consciousness during that year, and approximately one quarter of those with complete follow-up regained orientation and reached a favorable functional category. These numbers cannot be applied indiscriminately to an older patient with bilateral brainstem destruction, and conditioning on survival introduces important selection. They do demonstrate that the two-week phenotype is not the twelve-month outcome. The steepest observable recovery occurs after many acute withdrawal decisions have already been made.
The 2024 TRACK-TBI withdrawal analysis approached the same problem counterfactually. Ninety patients died after withdrawal of life-sustaining treatment, at a median of 5.4 days, and 80 were matched to patients with similar measured characteristics whose support was not withdrawn. Among matched patients followed without withdrawal, death and severe disability remained common, especially in the highest-risk tier, but more than 30% of survivors in the lower and intermediate tiers achieved at least partial independence. This does not prove that the patients who died would have recovered; propensity matching cannot recreate unmeasured examination findings, values, clinician judgments, or injury details. It does show that routinely measured acute variables could not reliably distinguish all patients capable of recovery from those who died after withdrawal. The appropriate conclusion is caution, not a claim that continued treatment usually produces independence.
A 2026 single-center propensity-matched study extended follow-up to 24 months. Among comparable patients whose support was continued, 38.3% of survivors who were fully dependent at three months reached at least partial independence by two years. The corresponding transitions occurred in 24.5% of those still dependent at six months and 13.1% of those still dependent at twelve months. Again, these are counterfactual estimates from matched survivors, not proof of what would have happened to a specific patient who underwent withdrawal. The study nevertheless refutes the idea that functional outcome has stabilized by six months. Recovery probability diminishes with time, but it does not abruptly become zero at an arbitrary milestone.
Long-term outcome also has to be described honestly in both directions. TRACK-TBI survivors with moderate-to-severe injury increased from 72% functionally independent at one year to 80% at five years, but complete recovery remained uncommon, and limitations in employment, relationships, leisure, cognition, and symptom burden persisted even among patients living independently. Mortality also remained higher than in mild TBI or orthopedic-trauma controls. Independence is therefore neither the same as restoration nor the same as unacceptable quality of life. The five-year data support continued rehabilitation while warning against presenting survival as uncomplicated recovery.
A defensible bedside process now becomes clear.
- First, state exactly what outcome is being estimated and at what time.
- Second, identify every reversible confounder of the examination and calculate whether drug clearance is biologically plausible rather than assuming an infusion has “worn off.”
- Third, reconstruct the injury phenotype from serial CT, MRI when safe and useful, extracranial injuries, and the presence or absence of hypoxic–ischemic damage.
- Fourth, examine serially with standardized behavioral assessment, quantitative pupils, and EEG rather than privileging one worst value.
- Fifth, use CRASH or IMPACT to calibrate—but not replace—clinical judgment.
- Sixth, separate evidence of catastrophic irreversibility, such as death by neurologic criteria or anatomically destructive brainstem injury, from evidence that merely increases the probability of disability.
- Finally, communicate the synthesis as a distribution: the best plausible outcome, the worst plausible outcome, the most likely range, the degree of confidence, and what new information over the next several days would move that estimate.
A time-limited trial of continued support is most useful when it is a genuine plan rather than a euphemism for indecision. The team should specify what will be treated, when sedation can be reduced, what repeat imaging or electrophysiology will answer, when standardized examinations will be repeated, and when the family will reconvene. Milestones such as resolving ICP burden, return of pupillary reactivity, emergence of visual pursuit, localization, or command-following can update prognosis positively. Failure to meet one early milestone should not automatically trigger withdrawal unless that failure has a validated relationship to the outcome the patient would reject. The duration of the trial should reflect injury biology, patient values, and medical complications rather than an institutional habit.
This framework is not therapeutic vitalism. Withdrawal of life-sustaining treatment can be ethically appropriate even when prognosis is uncertain if the range of plausible outcomes is incompatible with the patient’s known values. The corrective is neither “never withdraw” nor “always wait two weeks.” It is to make certain that the decision belongs to the patient’s values rather than to an overstated prediction, an unexamined disability bias, a single dramatic MRI image, or the discomfort of clinicians and families with uncertainty. The attending’s task is not to promise recovery. It is to prevent uncertainty from being silently converted into hopelessness.
That completes the severe-TBI sequence: early neuroresuscitation, lesion evolution, multimodality monitoring, tiered intracranial-hypertension management, rescue therapy, and longitudinal neuroprognostication. The next installment will transition into acute traumatic spinal cord injury, beginning with cord perfusion physiology, neurogenic versus hemorrhagic shock, respiratory failure by lesion level, imaging, and the timing of decompression.
Leave a Reply