A Comparison of the GCS and AVPU Scale Among Children Presenting to the Emergency Department

An assessment of level of consciousness is an essential component of the evaluation of children seeking emergency care. Guidelines from the National Pediatric Readiness Project emphasize assessment of level of consciousness for children in the emergency department (ED), considering this as a vital sign.¹ The Glasgow Coma Scale (GCS)² and the Alert, Verbal, Pain, Unresponsive (AVPU) scale³ are common tools for the assessment of level of consciousness in children with emergency conditions. Both are recommended by the 2020 Pediatric Advanced Life Support Guidelines.⁴ The GCS provides a detailed and granular assessment of 3 components of consciousness—eye response, verbal response, and motor response—giving a total score ranging from 3 to 15. It was specifically designed for the assessment of patients with trauma,² although it is widely applied in uninjured patients.^5,6 The AVPU scale is a simpler tool, particularly in prehospital and triage scenarios, with 4 levels based on the global level of responsiveness: alert, responsive to verbal stimuli, responsive to painful stimuli, or unresponsive.^7–9 Prior work has demonstrated the prognostic importance of the consciousness assessment for prediction of critical care admission,¹⁰ intracranial injury,^11,12 disability,^13–15 and mortality^15–17 among children.

The ability to translate information between consciousness assessment scales is important to ensure that clinicians using different scales have a clear understanding of a patient’s condition across settings and over time. Points of transition between scales occur during handoffs between various teams (including paramedics, nurses, ED physicians, and subspecialists) in the ED and guide treatment decisions. As such, establishing a crosswalk between these scores in the ED setting will be helpful in predictive modeling and machine learning applications to identify children at risk of clinical deterioration, when patients may have only 1 score available.

Prior research on mapping equivalent scores from the 2 scales (a crosswalk) are limited by small patient samples.^7,12,18 In recent work, we used paired, contemporaneous GCS and AVPU scale scores among children in the prehospital setting collected by prehospital clinicians to generate mutually exclusive categorizations of the GCS for each level of the AVPU scale.¹⁹ However, the relevance of this comparison in other clinical settings, such as the ED, requires further evaluation because of differences in how clinicians apply consciousness assessment scores based on setting and differences in case mix between the ED and prehospital setting (including having a higher proportion of children with medical complexity in the ED). We therefore sought to evaluate a previously developed prehospital-derived GCS-to-AVPU crosswalk among children presenting to the ED.

We performed a retrospective single-center study among children presenting to an urban tertiary care pediatric ED using a local version of an electronic health record–based dataset from a multicenter study (PED Screen with the Pediatric Emergency Care Applied Research Network) with established data extraction, harmonization, and deidentification pipelines, including data standardization and generation of clinical data files. Our ED has an annual volume of approximately 55 000 patients and serves as a referral center for patients with medical complexity and trauma. The present study was approved by our institutional review board as exempt human subjects research. For the present analysis, we included all children (aged <18 years) who presented to our ED between 2016 and 2022 and who had both a documented GCS and AVPU score in the ED. Among patients who had both scores available, we limited our analysis to encounters for which the GCS and AVPU documentation occurred within 15 minutes of each other.

At this institution, the GCS is part of vital sign documentation, and AVPU is documented in the nursing neurologic assessment. Both are charted within flowsheets within the hospital electronic health record. Except for the mandated requirement to chart GCS for trauma level A and level B activations, charting of these measures is practice dependent. Approximately half of the nurses in the ED are either a certified pediatric nurse or certified pediatric emergency nurse, with varying years of experience and prior practice locations.

We abstracted the following data elements from the PED Screen database for each included patient: age, sex, time of day of arrival (in 8-hour windows from 00:00 to 07:59, 08:00 to 15:59, and 16:00 to 23:59), diagnoses using International Classification of Diseases, 10th Revision Clinical Modification codes, triage acuity, and ED disposition. We categorized age into less than 5 years and at least 5 years based on previous research comparing the GCS and AVPU scale.¹² The principal diagnosis was categorized using the major group of the Diagnosis Grouping System (DGS), a consensus-derived classification scheme.^4,5 Triage acuity was documented using the Emergency Severity Index (ESI), a triage system based on the severity of patient conditions and the urgency of the need for medical care using categories of 1 to 5 (with lower numbers indicating higher acuity).²⁰ The presence of medical complexity was ascertained from all encounter-level diagnosis and procedure codes using version 3 of the Complex Chronic Condition algorithm described by Feinstein et al.²¹ 

We compared characteristics of patients with missing AVPU and/or GCS data with those who were included in the study sample. Our outcome of interest was the AVPU score. Our predictor was the GCS. We described sample demographic and treatment characteristics, stratified by AVPU. We evaluated a previously developed prehospital-derived GCS-to-AVPU crosswalk to estimate the AVPU score from the GCS.¹⁹ From this, a GCS of 14 to 15 corresponds to “Alert,” GCS 11 to 13 corresponds to “Verbal,” GCS 7 to 10 corresponds to “Pain,” and GCS 3 to 6 corresponds to “Unresponsive.”

We compared the observed AVPU score recorded by nursing staff with the predicted AVPU score (from the GCS as assessed by ED nursing staff and using the prehospital-derived GCS-to-AVPU crosswalk), reporting the overall accuracy, Cohen κ (with 95% CIs), and intraclass F1 score. We interpreted the κ statistic as follows: 0.8 to 1.0, almost perfect; 0.6 to 0.8, substantial; 0.4 to 0.6, moderate; 0.2 to 0.4, fair; and 0 to 0.2, slight.²² In imbalanced datasets, in which one class significantly outnumbers the other (eg, “Alert” occurs more frequently than any other category), accuracy can be misleading. The F1 score is useful in this type of scenario because it penalizes this lack of performance on the minority class. The intraclass F1 statistic is the harmonic mean of the precision (positive predictive value) and recall (sensitivity), performed individually for each class in a multiclass classification problem, and ranges between 0 and 1.²³ As a post hoc analysis, to account for the “kappa paradox” (in which a large degree of imbalance can result in a high raw agreement but low accuracy),²⁴ we additionally calculated Cohen κ by performing random under sampling. To do this, we randomly sampled the majority class (AVPU of “Alert”) to be equal to the sum of all the other classes and recalculated Cohen κ in this subgroup.

We performed the following additional analyses. First, we evaluated the agreement between the GCS-derived AVPU and the observed AVPU among (1) children with trauma (as identified by the DGS major grouping) and (2) children with medical complexity. Second, we derived a new crosswalk by identifying an optimal fitting categorization of GCS scores in this dataset compared with the previously established prehospital criteria. To do this, we identified all possible 4-way potential combinations in which the GCS could fit within the AVPU scale and determined which had the highest κ between the observed (AVPU recorded by nursing staff) and predicted values (determined from the documented GCS). We reported the same metrics of accuracy when using this internally derived crosswalk (from this ED population) among all patients and in subgroups of trauma, medical complexity, and age. Analyses were performed using the caret (v6.0-94)²⁵ package in R, version 4.3.2 (R Foundation for Statistical Computing).

We identified 408 162 encounters to the ED during the study period. After excluding adults (n = 10 832), encounters missing both GCS and AVPU data (n = 16 412), encounters missing AVPU data without missing GCS data (n = 548), and encounters missing GCS data without missing AVPU data (n = 41 742), there were 338 628 encounters with both GCS and AVPU data available. Patients with missing GCS data were younger, and patients with missing AVPU data were generally older than patients who had both data points available (Supplementary Table 1). Of encounters having both GCS and AVPU data, we included 295 247 (87.2%) encounters for which the initial GCS and AVPU scores were documented within 15 minutes of each other. Within this included sample, the GCS and AVPU had the same exact time stamp in 285 220 (96.6%). The median patient age was 4.6 years (IQR, 1.6–10.0 years), with a slight majority of boys (53.6%). Approximately half of encounters (153 543; 52.0%) were by patients aged younger than 5 years. Among patients with a GCS score of less than 15 (n = 7425; 0.3%), 781 (10.5%) had an abnormal motor, 7149 (96.3%) had an abnormal verbal, and 656 (8.8%) had an abnormal eye score (Figure 1).

Most patients in our sample were “Alert” (n = 294 470; 99.7%), with small proportions designated “Verbal” (n = 382; 0.1%), “Pain” (n = 287; 0.1%), and “Unresponsive” (n = 108, <0.1%). Children with an abnormal AVPU score, compared with those with a normal AVPU, more frequently were medically complex, were brought to the hospital by emergency medical services (EMS), had higher acuity based on triage classification, had a neurologic condition, and were admitted to the hospital (Table 1). GCS scores were progressively lower among patients with lower AVPU scores, with a clear stepwise association between the two (Figure 2).

When applying the prehospital GCS-derived AVPU crosswalk, 294 182 (99.6%) patients were classified as “Alert,” 615 (0.2%) as “Verbal,” 319 (0.1%) as “Pain,” and 131 (<0.1%) as “Unresponsive.” Comparing these with the nursing-documented AVPU, the overall accuracy of the prehospital GCS-to-AVPU crosswalk was greater than 99%, with a κ of 0.49 (95% CI, 0.47–0.52) indicating moderate interrater reliability (Table 2). Intraclass F1 statistics were higher for the AVPU of “Alert” (0.99) and “Unresponsive” (0.64) and were lower for “Verbal” (0.26) and “Pain” (0.56). When performing random undersampling in a post hoc analysis (in a dataset encompassing 777 encounters with an AVPU rating of “Alert” after random sampling and all 777 encounters with an AVPU not classified as “Alert”), there was a gain in κ (0.57; 95% CI, 0.54–0.61), with a decline in overall accuracy (74.1%).

Agreement between scales was lower among younger (κ = 0.39) compared with older children (κ = 0.56). Among injured children (n = 39 316), κ was slightly higher compared with the overall sample, although CIs overlapped (0.48; 95% CI, 0.39–0.58). F1 scores were similar between this subset and the overall sample. Among 1825 encounters for children with medical complexity, κ was higher, although again with overlapping CIs (0.60; 95% CI, 0.51–0.69) and similar F1 scores between the samples.

When internally deriving a GCS to AVPU crosswalk from ED data, the crosswalk with the highest κ corresponded to a GCS of 14 to 15 for “Alert,” 12 to 13 for “Verbal,” 6 to 11 for “Pain,” and 3 to 5 for “Unresponsive.” Cohen κ when using this crosswalk was similar when applied overall (0.51; 95% CI, 0.49–0.54) within subgroups of trauma (0.48; 95% CI, 0.39–0.58) and medical complexity (0.58; 95% CI, 0.49–0.67) to the prehospital GCS-derived AVPU crosswalk. When stratified into age groups, greater agreement was again noted among older (κ = 0.58) compared with younger children (κ = 0.40).

We performed a single-center retrospective study to evaluate a previously derived crosswalk from the GCS to the AVPU scales based on prehospital data. The performance of the prehospital-derived crosswalk on this ED-based population was comparable to the derivation study, with lower performance in the “Verbal” and “Pain” AVPU scale than in the prehospital study. These findings support the utility of the prehospital GCS-to-AVPU crosswalk. There was only a minimal gain in performance when internally deriving a crosswalk within this ED sample. In even optimally fit samples, agreement between the scales remained moderate, suggesting that even under ideal conditions, the AVPU scale cannot fully substitute for the GCS, particularly in patients with more nuanced clinical presentations. This underscores the need for careful consideration when using AVPU in settings and with patients requiring a more detailed neurological assessment.

Our findings inform the comparability of the AVPU and GCS scales and specifically evaluate these in the ED setting. This is an important extension of our prior research in the prehospital setting given differences in the type of clinician performing the level-of-consciousness assessment (nurses vs paramedics and emergency medical technicians) and in patient case mix. Our findings expand upon the relevance of the level-of-consciousness assessment and its implications for patient outcomes. Prior work on the topic has largely focused on children with neurotrauma. One single-center study by Nuttall et al demonstrated that among children aged younger than 15 years with suspected head injury, both an abnormal GCS score and AVPU score were associated with the presence of intracranial injury or a depressed skull fracture.¹² An evaluation of a statewide dataset in New York suggested that among injured children, the presence of an abnormal GCS motor score (of 1, indicating no motor response) or an “Unresponsive” AVPU was independently associated with mortality during their hospitalization.¹⁷ Our findings corroborate these previous studies and additionally expand the evaluation of these consciousness assessment scores to children without trauma.

Similar to prior work among both adults^9,26,27 and children,^7,12,28 we note a clear association between the GCS and AVPU scale. However, prior work has not attempted a definitive crosswalk between these scales and instead has reported medians with IQRs for the GCS at each level of the AVPU, resulting in overlapping scales.^7,12,28 As in prior work, we found that the crosswalk between the “Verbal” and “Pain” categories had lower fidelity (represented in our study by lower F1 scores), highlighting specific challenges in differentiating one category from the other in these intermediate levels. Importantly, children with a lower AVPU scale score were generally in need of more interventions compared with those who were classified as “Alert,” as reflected in arrival mode, ESI, and hospitalization. In addition, these children generally had greater medical complexity compared with those who were classified as “Alert” on the AVPU scale, potentially reflective of their baseline neurologic status. A challenge nevertheless arises when attempting to use mutually exclusive cut points to crosswalk these 2 scales. When doing so, there was only moderate agreement between the scales when measured by Cohen κ. We found the GCS and AVPU scale scores are not truly interchangeable: Although our results can facilitate an intelligible comparison between these scales, their crosswalking should not be understood as a 1:1 conversion. As such, caution is needed when translating between scales in specific clinical contexts, such as during patient handoffs. This is of greater importance when comparing these scales in younger children, among whom agreement was lower.

Notably, the κ scores identified in this study (0.49) are lower than those reported by the authors in a GCS-to-AVPU crosswalk derived from a prehospital dataset (0.61, meeting criteria for “substantial” agreement).¹⁹ However, the gain in performance in κ score when deriving an ED-based GCS-to-AVPU crosswalk was minimal (0.51). These findings may be reflective of the differing distribution of the AVPU scores within the dataset: Whereas the EMS study identified that 89% of children were “Alert,”¹⁹ this reached 99% in the present analysis. A lower Cohen κ despite the high overall agreement can be expected because Cohen κ adjusts for this expected agreement. As a result, even small disagreements in the less frequent categories can disproportionately lower the κ. This can lead to a situation wherein the raw agreement can appear near perfect, but the κ score is penalized because it accounts for the likelihood that this agreement might occur by chance alone.²⁴ The lower proportion of “Alert” patients in our prior study may reflect higher acuity of pediatric EMS transports compared with children presenting to the ED from all sources²⁹ and because of prior work suggesting that EMS level-of-consciousness assessments are frequently lower than in-ED assessments.³⁰ The difference in distribution may also derive from systemic differences in consciousness assessments by nurses (in this study) vs a sample of primarily paramedics and emergency medical technicians (as in the derivation study).

Our findings are subject to limitations. This was a single-center study that used data collected and charted for clinical purposes. As the study data were from a children’s hospital, findings may not be generalizable to other EDs or acute care settings. Not all patients had both an AVPU and GCS score documented, limiting their inclusion in this study. Prior research has suggested that both the AVPU and GCS have limited interrater reliability.^9,31 We used nursing-assigned AVPU and GCS scores, which may be different from scores assigned by other clinicians (including physicians and nurse practitioners). Based on our clinical experience, the individual practice of nurses within the ED can vary. One study comparing the interrater reliability of the GCS among adult patients in an Australian setting found that agreement was excellent between nurses and physicians (κ > 0.75) for verbal and total GCS scores and was intermediate (κ = 0.40–0.75) for motor and eye scores.³² Despite these limitations, this study provides a useful comparison of these commonly used consciousness assessment scales and how they are applied to children, with a proposed translation between the 2 scales.

In this retrospective single-center study, there was a clear stepwise association between the GCS and AVPU scale. There was a decline in performance when comparing our results to those from the derivation study, with only a mild gain in performance when internally deriving cut points based on the study data. Children with lower AVPU scores demonstrated greater resource need. Our findings demonstrate the use of the AVPU in identifying ill children and provide a crosswalk or schema to compare the GCS and AVPU scale among children in the ED.

AVPU

Alert, Verbal, Pain, Unresponsive

DGS

Diagnosis Grouping System

ED

emergency department

EMS

emergency medical services

ESI

Emergency Severity Index

GCS

Glasgow Coma Scale.