BACKGROUND: To validate the predictive value of the Glasgow Coma Score (GCS) and find the best way to model the score in a logistic regression model predicting mortality. METHODS: Analyses were based on 20,494 patients from the trauma registries of three urban Level I trauma centers in the province of Quebec, Canada. The predictive value of the GCS and its components was evaluated in logistic regression models predicting in-hospital mortality with measures of discrimination and calibration. The performance of the GCS with no transformation and as an ordered categorical variable was compared with two transformation techniques: fractional polynomials and spline regression. RESULTS: The GCS had excellent discrimination (area under Receiving Operator Characteristic Curve=0.833 95% confidence interval=0.820-0.846) but fairly poor calibration (Pearson's Chi-squared statistic=122 on 11 df). The eye component added no predictive information to the verbal and motor components in the whole sample but was important in certain sub-populations. Using the three components separately, rather than the sum, did not improve the predictive model. Fractional polynomial transformation of the GCS improved calibration and spline regression performed even better. GCS modeled as an ordered categorical variable performed badly both in terms of discrimination and calibration. CONCLUSIONS: The GCS in its present form is an efficient predictor of in-hospital mortality, which could benefit from statistical transformation in logistic regression models when the accuracy of estimated probabilities of mortality is important. The common use of GCS categories for modeling mortality leads to loss of information and should be discarded.
BACKGROUND: To validate the predictive value of the Glasgow Coma Score (GCS) and find the best way to model the score in a logistic regression model predicting mortality. METHODS: Analyses were based on 20,494 patients from the trauma registries of three urban Level I trauma centers in the province of Quebec, Canada. The predictive value of the GCS and its components was evaluated in logistic regression models predicting in-hospital mortality with measures of discrimination and calibration. The performance of the GCS with no transformation and as an ordered categorical variable was compared with two transformation techniques: fractional polynomials and spline regression. RESULTS: The GCS had excellent discrimination (area under Receiving Operator Characteristic Curve=0.833 95% confidence interval=0.820-0.846) but fairly poor calibration (Pearson's Chi-squared statistic=122 on 11 df). The eye component added no predictive information to the verbal and motor components in the whole sample but was important in certain sub-populations. Using the three components separately, rather than the sum, did not improve the predictive model. Fractional polynomial transformation of the GCS improved calibration and spline regression performed even better. GCS modeled as an ordered categorical variable performed badly both in terms of discrimination and calibration. CONCLUSIONS: The GCS in its present form is an efficient predictor of in-hospital mortality, which could benefit from statistical transformation in logistic regression models when the accuracy of estimated probabilities of mortality is important. The common use of GCS categories for modeling mortality leads to loss of information and should be discarded.
Authors: Randall S Scheibel; Mary R Newsome; Maya Troyanskaya; Joel L Steinberg; Felicia C Goldstein; Hui Mao; Harvey S Levin Journal: J Neurotrauma Date: 2009-09 Impact factor: 5.269
Authors: Robson Luis Amorim; Louise Makarem Oliveira; Luis Marcelo Malbouisson; Marcia Mitie Nagumo; Marcela Simoes; Leandro Miranda; Edson Bor-Seng-Shu; Andre Beer-Furlan; Almir Ferreira De Andrade; Andres M Rubiano; Manoel Jacobsen Teixeira; Angelos G Kolias; Wellingson Silva Paiva Journal: Front Neurol Date: 2020-01-24 Impact factor: 4.003