Ethics statement

Using a multicampus review mechanism, we obtained institutional review board approval from the University of California, Los Angeles (UCLA) Committee on Human Research for the 2 University of California sites. We obtained separate approvals from the Antelope Valley Hospital and UCSF Fresno, Community Regional Medical Center institutional review boards for their respective participating institutions. Because the study did not alter patient care, or present more than minimal risk, and because many patients would be critically ill, intoxicated, or have neurological impairments rendering them incapable of providing consent, we obtained a waiver of informed consent at all sites. We controlled all aspects of the study design, implementation, analysis, and manuscript preparation without influence from the grant funding agencies.

Participating centers

We conducted a prospective observational study of consecutive blunt head injury patients presenting to 4 general EDs (Antelope Valley Hospital, Lancaster, California; San Francisco General Hospital, San Francisco, California; UCLA Ronald Reagan Medical Center, Los Angeles, California; UCSF Fresno Community Regional Medical Center, Fresno, California). Participating institutions were specifically chosen to provide experience from university and community hospitals, centers with and without residency programs, public and private facilities, and exposure to a broad range of patients (children, adults, and the elderly) and environments, including urban, suburban, and rural communities. We conducted the study between April 2006 and December 2015, and the full protocol is included as a supplementary file to this article (S1 Protocol).

The study was designed to assess the validity of the following 8 criteria (the NEXUS Head CT DI) to exclude intracranial injuries in blunt head injury patients: no evidence of skull fracture, no scalp hematoma, no neurosurgical deficits, normal level of alertness, normal behavior, no persistent vomiting, no coagulopathy, and age less than 65 years.

These criteria were identified in our prior derivation study as indicators for potential candidates for our NEXUS head imaging DIs [8]. Box 1 presents the detailed description of each criterion we provided to the treating clinicians. We considered patients to be at low risk of intracranial injuries and safe for omission of CT head imaging if all 8 criteria were absent. We regarded patients who exhibited 1 or more of the criteria, and those with 1 or more unassessed criteria, to be at high risk for intracranial injury and in need of CT imaging.

An emergency physician at each center served as a study liaison and was charged with training participating clinicians in the criteria definitions and conduction of the study. Liaisons ensured that criterion assessments, radiographic results, and final outcomes were collected on all enrollees.

Patients

The study population consisted of all acute blunt head trauma patients undergoing CT head imaging at the participating centers. We enrolled patients when the treating provider ordered CT head imaging. We excluded patients with penetrating trauma, those with delayed presentations (greater then 24 hours after injury), and patients undergoing imaging for reasons unrelated to trauma. We also excluded patients who were transferred into a participating center with known intracranial injuries. To maximize compliance, participating centers adopted a protocol whereby CT imaging would not be performed until decision criteria had been assessed and recorded [19]. We cautioned clinicians against using the DIs as the sole determinant in making imaging decisions. The ultimate decision to obtain or omit imaging was made at the discretion of the treating provider and was not dictated by the study protocol.

Data collection

At the time of enrollment, the treating clinician collected and recorded limited demographic information (date of birth, sex, race, and ethnicity) and documented whether DI criteria were present, absent, or could not be assessed (for example coagulopathy in a comatose patient). In implementing the DI, we considered criteria that could not be assessed to be abnormal and excluded the patient from low-risk classification. This approach maximized safety, ensuring that low-risk assignments were based on actual measured assessments rather than missing information.

Physicians had the option to bypass data collection and obtain immediate imaging, prior to criterion assessment, on any patient they felt might be harmed by even minimal delay. Such patients were labeled as “unstable,” and clinicians were instructed to complete assessments of the criteria as soon as possible—and before imaging results were available. We considered “unstable” patients to be excluded from low-risk classification.

We also collected assessments for the Canadian Head CT rule, including inclusion criteria (initial Glasgow Coma Scale [GCS] score of 13 or greater, and injury within the last 24 hours), exclusion criteria (age less than 16 years, minimal head injury, no clear history of trauma as the primary event, obvious penetrating skull injury or depressed skull fracture, an acute focal neurological deficit, seizure prior to assessment, bleeding disorder or using anticoagulants, returned for reassessment of the same head injury), and risk criteria (High-risk criteria [for neurosurgical intervention]: GCS score <15 two hours after injury, suspected open or depressed skull fracture, any sign of basal skull fracture, 2 or more episodes of vomiting, and age ≥65 years. Medium-risk criteria [for significant brain injury]: Amnesia before impact >30 minutes, and dangerous mechanism).

Outcome measures

We defined our primary outcome a priori as the need for neurosurgical intervention, defined specifically as 1) death due to head injury, 2) need for craniotomy, 3) elevation of skull fracture, 4) intubation related to head injury, or 5) intracranial pressure monitoring, within 7 days of head injury [14].

Our secondary outcome was clinically significant head injury evident on CT imaging, as defined by Stiell, et al [14]. This included all injuries evident on CT head imaging except for the following in neurologically intact individuals: solitary small contusions, localized subarachnoid hemorrhage less than 1 mm thick, thin subdural hematomas less than 4 mm thick, isolated pneumocephaly, and closed depressed skull fractures that did not violate the inner table [14].

Formal radiographic interpretations and outcome assignments were all completed without knowledge of the criteria assessments recorded for each patient. We assigned each patient to the following 3 final outcome classes: 1) no significant injury, 2) significant injury (including injuries requiring neurosurgical intervention), and 3) injury requiring neurosurgical intervention. Two separate reviewers completed outcome assessments for each patient, with a third reviewer assigning outcomes in instances where the 2 initial assignments were discordant.

Assessing the potential for verification bias

Because we only enrolled patients undergoing head imaging, it is possible that important injuries may have gone unrecognized among the unimaged patients. To address the potential for verification bias that might arise from this approach, we conducted 3-month follow-up interviews on 368 consecutive blunt head injury patients who presented between July 2011 and March 2015. Patients were eligible for inclusion in this study if they had been evaluated in the UCLA ED for blunt head injury, did not receive head imaging as part of their evaluation, and agreed to participate in a 3-month follow-up interview and medical record review, and provide written informed consent. Patients were excluded from the study if they were not evaluated for blunt head injury during the enrollment period, underwent head CT imaging as part of their initial evaluation, or refused to participate in a 3-month follow-up interview and medical record review, or provide written informed consent. The size of this cohort reflects our desire to estimate the potential for verification bias to within 1.0%. In follow-up interviews, each patient was asked whether they had received radiographic head imaging elsewhere during a subsequent visit, and if so, the type of imaging (CT, MRI, skull films). Each patient was also assessed to determine whether they were diagnosed with intracranial injuries or required neurosurgical intervention. We also conducted a review of case logs and trauma logs to identify any instances of significant intracranial injuries or injuries requiring neurosurgical intervention that occurred among head injury patients who were seen on their initial presentation but not imaged.

Statistical analysis

Performance of the NEXUS instrument among the enrolled cohort

The NEXUS Head CT DI correctly assigned high-risk status to 420 of the 420 patients requiring neurosurgical intervention, yielding a sensitivity of 100.0% (95% CI: 99.1%–100.0%). The instrument correctly assigned low-risk status to 2,823 of 11,350 patients who did not require neurosurgical intervention to yield a specificity of 24.9% (95% CI: 24.1%–25.7%). None of the 2,823 patients assigned low-risk status required intervention, yielding a NPV of 100.0% (95% CI; 99.9%–100.0%).

The DI correctly assigned high-risk status to 759 of the 767 patients with significant intracranial injuries, yielding a sensitivity of 99.0% (95% CI: 98.0%–99.6%). The instrument assigned low-risk status to 2,815 of 11,003 patients who did not have significant injuries to yield a specificity of 25.6% (95% CI: 24.8%–26.4%). Significant injuries were absent in 2,815 of the 2,823 patients assigned low-risk status to yield a NPV of 99.7% (95% CI: 99.4%–99.9%). Box 2 describes the eight patients who sustained significant injuries and who were incorrectly assigned to low-risk status by the NEXUS Head CT DI.

Among the 420 patients requiring neurosurgical intervention, 1 or more of the high-risk criteria were documented as being definitely present in 416 patients. Four individuals, including 3 judged as unstable, were classified as high risk due to inability to assess all criteria. In no patients requiring intervention were all of the high-risk criteria documented as being absent. Table 2 presents the prevalence of individual criterion assignments for the total enrolled population, for patients with significant injuries, and those needing neurosurgical intervention. Abnormal level of alertness was present in 508 of the 767 (66.2%) patients with significant injuries, and 342 of the 420 (81.4%) patients requiring intervention, making it the most prevalent predictive criterion.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Prevalence of predictor variables among the enrolled population.

https://doi.org/10.1371/journal.pmed.1002313.t002

Performance of the NEXUS and Canadian instruments

Our population included 7,759 (65.9%) patients who satisfied the inclusion and exclusion criteria for classification by the Canadian Head CT rule. This subpopulation included 111 patients (1.43%) who required neurosurgical intervention and 306 patients (3.94%) with significant intracranial injuries. Fig 2 presents the flow chart detailing the subclassification of the original cohort, while Table 3 presents the classification summaries, operator characteristics, and associated CIs for both the Canadian and NEXUS Head CT rules, as measured on this subcohort. The Canadian “high-risk” criteria for neurosurgical intervention identified 108 of 111 patients requiring neurosurgical intervention and 252 of 306 patients with significant injuries to yield a sensitivity of 97.3% (95% CI: 92.3%–99.4%) in detecting injuries requiring intervention and a sensitivity of 82.4% (95% CI: 75.2%–86.5%) in detecting significant injuries. Box 3 describes the 3 patients requiring neurosurgical intervention who were incorrectly assigned to low-risk status by the Canadian “high-risk” criteria. The NEXUS rule, when applied to this same cohort, identified all 111 patients requiring neurosurgical intervention to yield a sensitivity of 100% (95% CI: 96.7%–100.0%). The Canadian “medium-risk” criteria identified 301 of 306 patients with significant injuries (sensitivity = 98.4%; 95% CI: 96.2%–99.5%), while the NEXUS rule identified 299 of these patients (sensitivity = 97.7%; 95% CI: 95.3%–99.1%). The Canadian rule exhibited a specificity of 12.3% (95% CI: 11.6%–13.1%). In comparison, the NEXUS rule exhibited a specificity of 33.3% (95% CI: 32.3%–34.4%).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Identification of patients suitable for evaluation by the Canadian Head computed tomographic (CT) rule.

https://doi.org/10.1371/journal.pmed.1002313.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Classification summary and operator characteristics of the Canadian and NEXUS decision instruments.^*

https://doi.org/10.1371/journal.pmed.1002313.t003

Assessment of the potential for verification bias

Our substudy examining the potential for verification bias enrolled 368 patients, 356 of whom received no further imaging or treatment for their injuries. Twelve patients ultimately underwent cranial imaging during their recovery, including six who received head CT imaging, and six who received magnetic resonance head imaging. No injuries were found on these imaging studies, and none of the 368 patients required neurosurgical intervention, yielding a measured potential for verification bias of 0.0% (95% CI: 0.0%–1.0%). Our review of case and trauma logs failed to reveal any missed instances of significant intracranial injuries or injuries requiring neurosurgical intervention among head injury patients who were discharged without imaging on their initial presentation.

Validation of the NEXUS Head CT DI

Our study demonstrates that the NEXUS Head CT DI is highly sensitive in detecting blunt trauma patients who harbor injuries that require neurosurgical intervention. Similarly, the high NPV provides reliable assurance that patients designated as “low risk” will not harbor these same injuries. The DI also exhibits extremely high sensitivity and NPV in detecting clinically significant injuries. Taken together, the precise measurements of sensitivity and NPV provide the validation needed to support the use of the tool in evaluating blunt head injury patients [15].

While the ability to detect and exclude injury is essential for the DI, its practical value lies in its ability to decrease imaging. Physicians in our study ordered imaging using whatever criteria they deemed appropriate and were specifically cautioned against using an unvalidated rule. The 11,770 patients selected for imaging included 2,815 (23.9%) patients who were ultimately classified as “low risk” by the tool. Application of the NEXUS Head CT DI would have allowed clinicians to safely omit imaging in these patients and suggests that the rule could significantly reduce CT imaging of blunt head injury patients. Taken together, these observations imply that the NEXUS Head CT DI is a reliable tool for identifying blunt head trauma patients at risk of important intracranial injuries while simultaneously offering the potential to reduce CT imaging.

Implementing the NEXUS Head CT rule

In general, DIs that exhibit high sensitivity are best used in conjunction with clinical judgment. This approach is particularly valuable in implementing the NEXUS Head CT DI. Numerous studies and our assessment of verification bias all reveal that injury prevalence among unimaged patients is negligible [8,14,16,17]. Clinicians, using clinical judgment alone, are already very sensitive in identifying and imaging patients with important injuries. DIs are not needed to improve sensitivity, and applying DIs to patients that are not currently selected for imaging offers no advantage. Instead this practice is likely to unnecessarily increase imaging. The greatest value of the highly sensitive NEXUS DI arises from its ability to safely reduce imaging by improving the specificity of imaging decisions. Consequently, the NEXUS DI should only be applied after a clinical evaluation suggests that imaging is indicated; patients exhibiting all of the low-risk criteria may safely forego imaging.

Assessment of the Canadian Head CT rule

A secondary goal of our study was to perform a more precise assessment of the Canadian Head CT rule. The initial development of the rule included only 44 patients requiring neurosurgical intervention. While the rule correctly classified all 44 patients as high risk, the corresponding lower CI for sensitivity was 92%, leaving questions about its validation [14,15]. This uncertainty has yet to be corrected by other studies [16,17].

In our study, among the subset of patients who met the inclusion and exclusion requirements, the Canadian Head CT “high-risk” criteria exhibited a sensitivity of 98.7% (with a lower confidence bound of 96.8%) in identifying patients requiring neurosurgical intervention. While these numbers are consistent with the original study, they raise concern regarding whether the rule has sufficient sensitivity in detecting injuries that require intervention. Review of the patients with missed injuries reveals that 2 of the 3 misses occurred in patients with epidural hematomas who experienced prolonged lucid intervals. In assessing the construct validity of the rule, it is apparent that the rule is vulnerable to such incidents. It is also important to note that 1 of the patients with missed injuries was not identified by the medium-risk factors of amnesia or dangerous mechanism and would have been classified as “low risk” by both the “high-risk” and “medium-risk” versions of the Canadian rule.

In our study, the Canadian “high-risk” criteria exhibited a specificity of 58.8% in identifying patients who do not have intracranial injuries requiring intervention. While this high specificity is advantageous, it comes at the cost of poor sensitivity in detecting significant injuries (sensitivity of 82.4%). Thus, while the high specificity is compelling, clinicians are unlikely to embrace the “high-risk” criteria in isolation because they are likely to miss nearly one-sixth of the patients having significant injuries. From a practical perspective, the reliable detection of injuries requires use of the full Canadian rule, including application of both the “high-risk” and “medium-risk” criteria.

With this in mind, it is worth noting that both the NEXUS and full Canadian rules exhibited similar high sensitivity and NPV in identifying patients with significant injuries. The primary difference between the rules lies in their potential utility. Because the NEXUS rule has no inclusion or exclusion criteria, it can be applied to all blunt head injury patients. In contrast, the Canadian rule has numerous restrictions and could not be applied to approximately one-third of our cohort (4,011 patients). Clinicians who employ the Canadian rule will have to find alternative means of assessing the need for imaging among such patients.

It is worth noting that we found lower specificity than that observed in the original Canadian study [14]. This difference reflects the large number of patients who were never imaged in the original Canadian cohort (1,043 of 3,121 patients), the majority of whom were classified as “low risk” by the instrument. The inclusion of such patients acts to increase the observed specificity of the Canadian rule, but limits an assessment of the ability of the rule to decrease imaging. We found that the full Canadian rule assigned “low-risk” status to 917 of the 7,987 patients who satisfied the inclusion and exclusion criteria. Assuming imaging could be omitted in all of these “low-risk” patients, the Canadian rule could potentially reduce imaging by 11.5%. In the same cohort, the NEXUS rule classified 2,571 patients as “low risk,” with the corresponding potential reduction in imaging of 32.2%, a nearly 3-fold reduction in imaging.

Using DIs to guide CT head imaging

Clinical DIs such as the NEXUS and Canadian Head CT instruments are designed to serve as screening tools in the initial assessment of blunt head injury patients. Imaging may be safely omitted for patients who are classified as low risk, while imaging or another form of evaluation, such as prolonged observation, is appropriate for patients that do not meet low-risk classification [12].

Other tools, such as biomarkers, may also play a role in injury assessment, but the exact role these tools play in injury detection has yet to be determined [22,23]. For example, there are currently no reliable direct comparisons examining the performance of DIs and biomarkers to determine whether one means of injury assessment is superior to another. Further research will also be needed to examine the combined use of these tools and whether they are best used to enhance sensitivity, or whether they might play an important role in enhancing specificity and reducing imaging.

Limitations

While we employed a naturalistic design in our study, screening performance results are likely to differ when applied in different environments. This is particularly true for our measured specificity and potential to reduce imaging. For example, centers that encounter a larger proportion of low-risk patients will likely find a higher proportion of patients that meet “low-risk” classification and are likely to find a higher specificity for the instrument. Hospitals that encounter a higher proportion of seriously injured patients may find fewer “low-risk” patients and observe a lower specificity. Similarly, clinicians differ from center to center, and while several hundred clinicians participated in our trial, their combined performance will likely differ from that of any individual or single center.

The criteria in our rule are also subject to interpretation. We tried to provide a pragmatic, “real world” exposure of these criteria to the clinicians participating in the study by providing them with definitions of the criteria but refraining from intense training in the assessment of each criterion. Most of our criteria are very straightforward, but interpretation of a few of the criteria and application of the rule may change as exposure to the rule changes. This may be particularly relevant to centers with cultural and language differences [24].

Because we employed neurosurgical intervention as a primary outcome in our study, our results are vulnerable to the decision process used in implementing these interventions. Other centers, with different neurosurgical support services, may find differing rates and levels of intervention. These differences could impact all of the measured operator characteristics.

Tracking outcomes for all blunt head injury patients for an extended period is not feasible. This implies that our assessment of the potential for verification bias is not exhaustive, and it is possible that a patient with serious intracranial lesions may have been discharged with unrecognized injuries. However, UCLA is the regional trauma and neurosurgical center, and it is likely that any patient with injuries missed on an initial evaluation would ultimately return for evaluation and intervention. Thus, while instances of missed injury may have occurred during the study, our failure to identify any such injuries indicates that they are extremely rare and well below significant thresholds.