Beyond mild, moderate, and severe traumatic brain injury: modelling severity from clinical, neuroimaging, and blood-based indicators

Abstract

Background: The conventional clinical approach to characterising traumatic brain injuries (TBIs) as mild, moderate, or severe using the Glasgow Coma Scale (GCS) total score has well-known limitations, prompting calls for more sophisticated strategies.

Methods: We used item response theory (IRT) to develop a new method for quantifying TBI severity using 24 clinical, head computed tomography, and blood-based biomarker variables familiar to clinicians and researchers. IRT uses individuals' response patterns across indicators to estimate relationships between the indicators and a latent continuum of TBI severity. Model parameters were used to assign severity scores in two large cohorts, and associations with traditional GCS categories and 6-month functional outcomes (Glasgow Outcome Scale-Extended [GOSE]) were tested with correlational and logistic regression analyses.

Findings: In the prospective Transforming Research and Clinical Knowledge in TBI (TRACK-TBI) cohort (N = 2545), modelling showed the 24 indicators index a common latent continuum of TBI severity. IRT enabled us to identify the relative contribution of these features to estimate an individual's TBI severity. Finally, within both the TRACK-TBI derivation sample and an external validation sample (Collaborative European NeuroTrauma Effectiveness Research in TBI [CENTER-TBI]), TBI severity scores generated using this novel IRT-based method incrementally predicted functional (GOSE) outcome better than classic clinical (mild, moderate, severe) or International Mission for Prognosis and Analysis of Clinical Trials in TBI (IMPACT) classification methods.

Interpretation: Our findings directly inform ongoing international efforts to refine and deploy new pragmatic, empirically-supported strategies for characterising TBI, while illustrating a strategy that may be useful to improve staging systems for other diseases.

Funding: This secondary analysis project was funded by the U.S. National Institute of Neurological Disorders and Stroke (Grant No. R01 NS110856).

Keywords: Blood-based biomarkers; Classification; Item response theory; Neuroimaging; Severity; Traumatic brain injury.

Fig. 1

Item information curves from a single item response theory (IRT) model of 24 TBI severity indicators, stratified by measurement domain for readability(a: brain imaging; b: clinical assessments; c: blood biomarkers). The model reflects all 2545 individuals with TBI age 17 or older in the TRACK-TBI study. The x-axis reflects the latent TBI severity spectrum modelled from the associations between the indicators using IRT. The y-axis reflects IRT information, which reflects the precision with which each variable can be used to measure individuals on the severity dimension, which can vary at different levels of severity. Higher information reflects lower standard errors to estimate individuals at a given level of severity. Abbreviations: CT, computed tomography; EDH, epidural haematoma; GCS, Glasgow Coma Scale; GFAP, glial fibrillary acidic protein; hsCRP, high-sensitivity C-reactive protein; IVH, intraventricular haemorrhage; NSE, neuron-specific enolase; LOC, loss of consciousness; PTA, posttraumatic amnesia; S100B, S100 calcium binding protein B; SAH, subarachnoid haemorrhage; SDH, subdural haematoma; TBI, traumatic brain injury; UCH-L1, ubiquitin C-terminal hydrolase.

Fig. 2

Test information curves for TBI severity indicators grouped by measurement domains. The model reflects all 2545 individuals with TBI age 17 or older in the TRACK-TBI study. (a) Test information curves for each measurement domain. Test information reflects the sum of item-level information (see Fig. 1) across all indicators (items) in each measurement domain. Higher information reflects greater measurement precision (i.e., lower standard error) in characterising and distinguishing persons at a given level of TBI severity. (b) Test information (solid lines) and associated standard errors (dashed lines) for increasingly complex sets of indicators, starting with GCS domain scores and adding, in order: pupil reactivity, LOC duration, and PTA duration; CT findings; and blood-based biomarkers. Abbreviations: EDH, epidural haematoma; GCS, Glasgow Coma Scale; GFAP, glial fibrillary acidic protein; hsCRP, high-sensitivity C-reactive protein; IVH, intraventricular haemorrhage; NSE, neuron-specific enolase; LOC, loss of consciousness; PTA, posttraumatic amnesia; S100B, S100 calcium binding protein B; SAH, subarachnoid haemorrhage; SDH, subdural haematoma; TBI, traumatic brain injury; UCH-L1, ubiquitin C-terminal hydrolase.

Fig. 3

Distribution and prognostic value of item response theory (IRT)-based traumatic brain injury (TBI) severity scores. The model was developed from profiles of 24 indicators of acute TBI severity within the TRACK-TBI sample; those parameters were used to score individuals on the severity spectrum in both the TRACK-TBI (derivation) and CENTER-TBI (external validation) sample. Due to differences between studies, the CENTER-TBI subjects were scored using 17 of the original 24 variables. (a) Histogram depicting the distribution of TBI Severity IRT scores, which provides substantially more precision in estimating individual differences in TBI severity than traditional mild, moderate, and severe TBI classification (2446/2545 scores in the TRACK-TBI sample and 3438/4500 scores in the CENTER-TBI sample were unique). (b) Scatterplot of TBI Severity IRT scores versus traditional GCS-based classification of TBI; the robust, expected associations supports an interpretation that TBI Severity IRT scores reflect the same construct (TBI severity) as GCS-based classification, with more precision afforded by the IRT scores. Scatterplot points (individual subjects) are lagged in the direction of the y-axis to facilitate visualisation of the number of points along the x-axis. (c) Model Nagelkerke R² for models predicting death, unfavourable outcome, and incomplete recovery at 6 months post-injury from traditional 3-level GCS-based TBI severity categories (mild, moderate, severe) and IMPACT model scores, as well as the increase in R² observed after adding TBI Severity IRT scores to the models (all likelihood ratio test p ≤ 0.001). The figure illustrates that TBI Severity IRT scores, while not developed specifically to predict functional outcome, they nevertheless incrementally predict outcome beyond these other more traditional prognostic scores (GCS and IMPACT). Abbreviations: CENTER-TBI, Collaborative European NeuroTrauma Effectiveness Research in Traumatic Brain Injury study; GCS, Glasgow Coma Scale score; GOSE, Glasgow Outcome Scale-Extended; IMPACT, International Mission for Prognosis and Analysis of Clinical Trials in TBI; TRACK-TBI, Transforming Research and Clinical knowledge in TBI study.