The Predictive Performance of Glasgow Coma Scale in Relation to Time

Background

Trauma is a critical global public health concern and the number of fatalities as a result of trauma continues to increase globally. In 2016 more than 4,6 million deaths were the result of trauma making it the 8th leading cause of death. In the Global Burden and Disease study (GBD), traumatic injuries account for nearly one-tenth of all deaths, more than malaria, tuberculosis, HIV/AIDS and maternal conditions combined.

TBI is defined as brain injury caused by trauma, trauma being defined as damage inflicted on the body as the direct or indirect result of an external force, with or without disruption of structural continuity. Traumatic brain injury (TBI) is a leading cause of morbidity and mortality globally. It is the largest contributor to trauma deaths in the world, having three times higher mortality rate than trauma without accompanied TBI. It is estimated that TBI affects more than 10 million people annually.

To guide physicians in diagnostics and resuscitation regimes, prognostic models predicting TBI severity and outcome are of great importance. The Glasgow Coma Scale (GCS) is one of the most widely used prognostic model. It is a neurological scale developed to assess the response to stimuli in patients with craniocerebral injuries by using three parameters; eye opening, verbal response and motor response. It has been used to assess level of consciousness in both clinical practice and neurotrauma research. The GCS Score, the sum of the three parameters used in GCS is used to assess severity, ranging from mild, moderate to severe.

Several studies have researched the prognostic value of GCS and the prognostic value of GCS individual components, but little is known about how the relationship of time between injury and first GCS affects the predictive value of GCS. Studies have shown that GCS measured on admission to trauma care is more predictive than GCS measured at the trauma site, and that the median GCS on admission is higher than the median GCS at trauma site but if and how time affects GCS scores has not been studied.

Aim

The aim of this study is to assess how the timing of measuring GCS affects its predictive value after traumatic brain injury (TBI) in adult patients. The first objective is to assess if the predictive performance of GCS is improved if adjusted for time from injury to when it was recorded. The second objective is to assess if the predictive performance of GCS varies depending on when it was recorded.

Study Design

This is a retrospective analysis of the cohort study Towards Improved Trauma Care Outcomes in India (TITCO).

Setting

The de-identified TITCO cohort will be used. This cohort includes 16,000 patients enrolled between July 2013 and December 2015 from four university hospitals in India. Project officers, holding a health science master degree, collected data prospectively on admission at each site by direct observation of the emergency room and filling out a standardized form. The project officers worked in rotating eight-hour shifts (morning, evening, night). Data was also retrieved retrospectively from patient records for patients admitted outside the observed shifts.

Source and method of participant selection

The one-site project officer included patients from participating hospitals, either by prospective observation or by retrospective data retrieval from patient records.

Explanatory variables The two explanatory variables of interest will be GCS and time between injury and GCS recording in hours, henceforth referred to as time to GCS. GCS was extracted from patient records as was the date and time of first GCS recorded. If date and time of first GCS recording are missing date then time of arrival to the participating centre will be used instead. Date and time of injury were extracted from patient records or directly reported by participants. Data and time of first GCS recording or arrival to participating centre were extracted from patient records.

Covariates The variables age, sex, mechanism of injury, whether the patient was transferred from another health facility, and anatomical injury severity quantified using the injury severity score (ISS) will be reported to characterise the study sample. Age, sex, mechanism of injury and transfer status were either extracted from patient records or reported by participants. ISS was calculated by a single accredited coder based on injury text descriptions.

Bias

To account for human errors in recording GCS, Aal data collector observers were holders of health science master degrees and were continually trained and supervised throughout the data collection period.

Quantitative variables

GCS will be treated both as a linear term and as an ordinal variable with 12 levels. The non-testable levels of the verbal and eye components will be treated as 1. Time between injury and recorded GCS will be treated both as a continuous variable and a categorical variable. When treated as continuous time between injury and GCS will be allowed to have a non-linear association with mortality by modelling it using restricted cubic splines with three knots placed at equally spaced percentiles. When treated as categorical it will be divided into blocks of two hours.

Statistical methods

All analyses will be conducted in the statistical language and programming environment R. The sample will first be temporally split into training and validation samples as outlined in the study size section below. Each sample will then be characterised using medians and inter-quartile ranges (IQR) to present quantitative variables and counts and percentages to present qualitative variables.

In the training sample four simple prediction models will be fit using logistic regression with mortality as the outcome. The first model will include only GCS as a linear term, the second GCS as a linear term and time to GCS modelled using restricted cubic splines, the third only GCS as an ordinal variable, and the fourth model GCS as an ordinal variable and time to GCS modelled using restricted cubic splines. To avoid overfitting a shrinkage factor will be estimated using a bootstrapping procedure which will then be applied to the model coefficients.

The four models will then be applied in the validation sample and their predictive performance estimated and compared. Predictive performance of each model will be evaluated using the area under the receiver operating characteristics curve (AUROCC), positive and negative predictive values. Differences in predictive performance between models and associated 95% confidence intervals will be estimated using bootstrapping.

Each of the training and validation samples will then be divided into subsamples based on time to GCS, so that the first subsample includes patients with time to GCS < 2 hours and the second subsample includes patients with time to GCS between two and four hours and so on in blocks of two hours. In each of the training subsamples a simple logistic model will be fit including GCS as a linear term as the only independent variable. The coefficient of GCS in each model will be shrunk.

The model developed in the first training subsample will then be applied to the first validation subsample and so on. Model performance in each validation subsamples will be evaluated using AUROCC and root mean square error. The trend in these measures across validation subsamples will then be quantified using a simple generalised linear model with performance measure data as the outcome variable and a nonlinear transformation of block index number using restricted cubic splines with three knots as the only independent variable.

Strategy to handle missing data

If the required sample size is reached if only patients with complete data on the outcome, explanatory variable, and covariates are included then a complete case analysis will be conducted. If not then missing data will be handled with multiple imputation using chained equations. The number of imputed datasets will be equal to the percentage of incomplete observations. The analysis will be conducted separately in each imputed dataset and the main results presented as medians with IQR across imputations. For confidence intervals the most extreme values of pooled upper and lower bounds will be reported.

Study size

The most data intensive analysis is likely to be fitting the model with GCS as an ordinal variable and time to GCS as restricted cubic splines and therefore the study size is estimated to accommodate this analysis. Simulation studies indicate that logistic regression models need at least ten events, or observations with the outcome, per included parameter to generate reliable coefficient estimates. Modelling GCS as an ordinal variable will involve estimating coefficients for eleven parameters and including time to GCS adds two additional parameters. The total number of parameters is then 13, indicating a need for at least 130 events. Assuming an outcome prevalence of 20% based on previous research the training sample needs to include at least 650 observations. If this number is less than half of the complete sample then the training and validation samples will be generated by splitting the complete sample in two parts of equal size. If the complete sample includes less than 1300 observations the first 650 observations will be included in the training sample and the remaining observations will be included in the validation sample. Regardless, the samples will be created in such a way that the relative contribution of each centre is approximately the same in both samples.