Injury, Int. J. Care Injured (2005) 36, 900—904

www.elsevier.com/locate/injury

Comparing logistic models based on modified GCSmotor component with other prognostic tools inprediction of mortality: Results of study in 7226trauma patients

Behzad Eftekhar *, Mohammad Reza Zarei, Mohammad Ghodsi,Koorosh MoezArdalan, Moosa Zargar, Ebrahim Ketabchi

Departments of Neurosurgery and Surgery, Sina Trauma Research Center, Sina Hospital,Tehran University, Iran

Accepted 20 December 2004

KEYWORDSMortality;Logistic Model;GCS;Motor component

Summary A simple reproducible and sensitive prognostic trauma tool is stillneeded. In this article we have introduced modified GCS motor response (MGMR)and evaluated the performance of logistic models based on this variable.

The records of 8452 trauma patients admitted to major hospitals of Tehran from1999 to 2000 were analysed. 7226 records with known outcome were included in ourstudy.

Logistic models based on outcome (death versus survival) as a dependent variableand Injury Severity Score (ISS), Revised Trauma Score (RTS), Glasgow Coma Scale(GCS), GCS motor component (GMR) and MGMR (following command [=2], movementbut not following [=1] command and without movement [=0]) were compared basedon their accuracy and area under the Receiver Operating Characteristic (ROC) curve.

The accuracy of the Trauma and Injury Severity Score (TRISS), RTS, GCS, GMR andMGMR models were almost the same. Considering both the area under the ROC curveand accuracy, the age included MGMR model was also comparable with other ageincluded models (RTS + age, GCS + age, GMR + age).

We concluded that although in some situations we need more sophisticatedmodels, should our results be reproducible in other populations, MGMR (with orwithout age added) model may be of considerable practical value.# 2005 Elsevier Ltd. All rights reserved.

* Corresponding author. Tel.: +98 21 6008504;fax: +98 21 8760967.

E-mail address: Eftekhar@sina.tums.ac.ir (B. Eftekhar).

0020–1383/$ — see front matter # 2005 Elsevier Ltd. All rights resedoi:10.1016/j.injury.2004.12.067

Introduction

Although a number of prognostic tools and scoreshave been proposed for differentiating trauma

rved.

Comparing logistic models based on modified GCS motor component with other prognostic tools 901

patients at risk of significant morbidity and mortal-ity, many of them take time and require medicalexpertise and training to be calculated in the field. Asimple reproducible and sensitive tool for theabove-mentioned goal still is needed.

The complexity of these scoring systems is one ofthe major factors that may contribute to the lowreporting rate and use of the prognostic tools andmay interfere with their use in actual practice.

There are reports that suggest that the motorcomponent of the GCS (GMR) is a powerful predictorof outcome in trauma patients, especially in thosewith head injury.2,9,11,13,17 As those who are directlyinvolved in both training and treatment, despiteprevious reports,5 we think that the number ofmotor response levels from 1 to 6 might contributeto some of the variations among observers duringthe assessment of the motor response of the GCS.6

Modification of the motor component of the GCSto three levels simplifies these calculations signifi-cantly.

The influence of age on the outcome of thetrauma patients has been studied in neurosurgicalliterature14 and others. Its role in measures of injuryseverity was reemphasised in the recent work ofStephenson et al.18

In this study we have compared the ability of thelogistic models based on the simplified variable,especially when age of the patients is included, inthe prediction of mortality with some of the knownprognostic tools.

Materials and methods

In a prospective study, the records of 8452 traumapatients admitted to the emergency departments ofsix major university hospitals in Tehran from 23August 1999 to 22 September 2000 were analysed.This study ran in continuity with the Trauma DataRegistry Programme (started in 1996 in Sina HospitalTrauma Research Center), which is affiliated to theTehran University of Medical Sciences.20,16 The widedistribution of these hospitals throughout Tehranprovided good coverage for trauma cases in differ-ent parts of this huge metropolitan area. The popu-lation for this study was all trauma victims who hadbeen admitted to one of the hospitals for more than24 h during the data-gathering period. The timelimit was not employed in the consideration of deadpatients and all dead patients who were referred tothese hospitals were included in our study. We haveexcluded those transferred to other hospitals orwith related missing values. 7226 records had aknown outcome (death/survival) and were includedin our study. Structured, closed-question data

checklists were used for the data gathering process.We collected three major categories of injury-related information: demographic data, prehospitaldata (if they were available) and inhospital data.Hospital related data included: vital signs, GlasgowComa Scale (GCS),19 Abbreviated Injury Scale (AIS-90),10 clinical findings (according to InternationalClassification of Diseases 10th revision (ICD-10)),hospital expenses and the outcome of the patients.Data collection was conducted by a group of trainedphysicians. The physicians had completed specialtraining courses to become familiar with the processof extracting Abbreviated Injury Score (AIS-90)codes and filling out questionnaires. For qualitycontrol (QC) purposes each hospital had a physicianwho was responsible for overseeing the data gather-ing process. Finally, a medical practitioner exam-ined all the checklists in order to evaluate andcorrect them (if required) based on fixed protocols.

The Revised Trauma Score (RTS),8 Glasgow ComaScale (GCS), Injury Severity Score1 (ISS, based onAIS-90 dictionary), Trauma and Injury SeverityScore Probability of Survival (TRISS Ps)4 and themotor response of the GCS (GMR) upon arrival atthe emergency department were calculated andused in our study. The TRISS Ps is derived from amultiple logistic regression model and uses thepatient’s age, mechanism of injury, RTS and ISS.Both ISS and TRISS have been included in this ana-lysis only as reference points, not as clinically usefulprognostic indices.

We defined a new variable named modified GMR(MGMR) to be equal to 2 when the patient followedcommands (GMR = 6), 1 when the patient had move-ment but did not follow commands and 0 whenwithout any movement (GMR = 0). Another variable(M) was defined to be equal to 1 when the patientobeyed the command (GMR = 6) and 0 when not.

Using logistic regression, RTS, ISS, GCS, GMR, Mand MGMR were compared as predictors of the riskof mortality. The dependent variable in thesemodels was outcome (death versus survival) andthe independent variables were each of theabove-mentioned scores. We repeated the analy-sis once again adding age as a covariate. Thecomparison was based upon calculation of thearea under the Receiver Operating Characteristic(ROC) curve,12 sensitivity, specificity and accuracyrate ([true positive + true negative]/total) of themodels. A ROC value of 1 corresponds to a testthat perfectly separates two populations, whereasa ROC value of 0.5 corresponds to a perfectlyuseless test that performs no better than chance.All the analyses were performed using IntercooledSTATA for Windows, version 6 (STATA Corp., Col-lege Station, TX).

902 B. Eftekhar et al.

Figure 1 Distribution of transportation to emergencycare.

Figure 2 Distribution of injuries by body system.

Figure 3 Frequency and distribution of ISS.

Results

After exclusion of incomplete data and outliers,there were 7226 patients’ records available for finalanalysis. The mean age of the study population was32.5 � 21 years. The age distribution showed a peak(23.7%) in 10—20 age range. Seventy-six percent ofour patients were male and 24% were female. 78.1%were transferred to the hospital by private vehicles(other than ambulances) (Fig. 1). The physiologicmeasures of study patients on arrival are shown inTable 1. The distribution of injuries by body systemfor all patients as a whole is shown in Fig. 2. Thelimbs were most commonly injured. The head,chest, and abdomen were the most severely injuredbody systems. 7.8% of the patients had penetratinginjuries. Figs. 3 and 4 show the frequency anddistribution of the Injury Severity Score and GCS

Table 1 The physiologicmeasures of study patients onarrival

Variable Mean � S.D.

Pulse 83.5 � 14.5Respiratory rate 17.8 � 3.7Blood pressure 117.70 � 21.7Revised Trauma Score 7.65 � 0.94Injury Severity Score 7.1 � 10.1GCS 14.5 � 2Motor response 5.8 � 0.79

motor response. The overall mortality rate for thestudy population was 3.8%. Figs. 3 and 4 and themortality rate show that most of the patients werenot severely injured. The results of comparison ofthe models built on logistic regression analyses areshown in the Table 2. As is seen, TRISS (where theage is automatically included) compares the mostfavourably (when only the area under the ROC curveis considered). GCS (with or without age added) has

Figure 4 Frequency and distribution of the GlasgowComa Scale motor response.

Comparing logistic models based on modified GCS motor component with other prognostic tools 903

Table 2 The results of comparison of the models

Variable Includedcases/total

Specificity/sensitivity (%)

Area underROC curve

Accuracy Area underROCa

Accuracya

ISS 7208/8452 99.8/49.2 0.944 97.9 0.948 97.9TRISS(Ps) 7186/8452 99.9/62.7 0.969 98.6RTS 7211/8452 99.8/64.5 0.911 98.5 0.956 98.6GCS 7225/8452 99.7/71.5 0.907 98.6 0.955 98.7GMR 7202/8452 99.7/67.8 0.899 98.5 0.946 98.6M(0,1) 7202/8452 97.6/80.5 0.891 97 0.942 97.2MGMR (0,1,2) 7202/8452 99.9/55 0.897 98.3 0.947 98.4a Age included.

been the most accurate identifier of death risk.Including the age significantly increases the indicesfor all of the models (except ISS). The accuracy ofthe TRISS, RTS, GCS, GMR and MGMR models arealmost the same. The area under the ROC curve ofthese models when age is added (the two rightmostcolumns) increases. Considering the age addedmod-els, the performance of the MGMR + age is similar tothe GMR + age and slightly weaker than RTS + ageand GCS + age models. Its accuracy, sensitivity andspecificity were 98.4%, 99.7% and 63.8%, respec-tively.

The M(1,0) model is the simplest and the mostsensitive in prediction of outcome of injury. Whenage is added, its performance improves significantly,but still is weaker than the others.

Discussion

There are differences between the rate of mortalityand the numbers of patients transferred to the hos-pital by trained staff between our study and previousones. The lower mortality rate does not truly repre-sent a real incidence in our community, where manydeaths from various causes occur outside of the hos-pital.20,16 Other indices such as age and sex distribu-tion and the physiologic measures of our patients arecomparable with other studies.16,15 Although predic-tion ofmortality is not the only index for evaluation ofthe performance of a prognostic model, in relation tothe above mentioned differences, prediction of out-come in the form of death or survival seems to be themost reliable index in our situation. In spite of somedifferences betweenour study population andothers,known tools suchasTRISS andRTShaveworkedwell inprediction of mortality in our study. The calculatedarea under the ROC curve for TRISS model (0.969) isalmost similar to the findings of the Cologne Valida-tion Study.3

On the whole, this study reemphasises the role ofmotor response in prediction of outcome in traumapatients. In accordance with Meredith et al.,15 our

study demonstrated that the accuracy of the GMR(98.5%) had been similar to the RTS (98.5%) andbetter than ISS (97.9%) in identifying the outcomeof the patients. In our study the area under the ROCcurve in GMR model was less than both the RTS andISS models.

Meredith et al. used Trauma Score (TS)7 instead ofRTS and the calculated accuracy and sensitivity oftheir models were slightly less than ours. They didnot use the area under the ROC curve as an index ofthe ability of the model to distinguish those who diefrom those who survive. These differences, consid-ering the above seem to be justifiable.

The performance of the MGMR + age model,though still weaker than TRISS, was comparablewith other age included models (RTS + age,GCS + age, GMR + age). This model requires muchless observer training and has less interobservervariability, but still needs some mathematical cal-culations that can be computed using a prepro-grammed calculator.

Although in some situations we need more sophis-ticated models, should our results be reproducible inother populations,MGMR (with orwithout ageadded)model may be of considerable practical value.

Acknowledgments

The authors would also like to express their grati-tude to Ms. Orla Dunne of Mater Misericordiae Uni-versity Hospital, Dublin for her assistance with theediting of this paper.

References

1. Baker SP, O’Neill B, Haddon W. The Injury Severity Score: amethod for describing patients with multiple injuries andevaluating emergency care. J Trauma 1974;14:187.

2. Baxt WG, Jones G, Fortlage D. The trauma triage rule: a new,resource-based approach to the prehospital identification ofthe major trauma victims. Ann Emerg Med 1990;19:1401.

904 B. Eftekhar et al.

3. Boullon B, Lefering R, Vorweg M, et al. Trauma score systems:cologne validation study. J Trauma 1997;42:652—8.

4. Boyd CR, Tolson MA, Copes WS. Evaluating trauma care: theTRISS method. J Trauma 1987;27:370.

5. Braakman R, Avezaat CJ, Maas AI, et al. Interobserver agree-ment in the assessment of the motor response of the Glasgow‘coma’ scale. Clin Neurol Neurosurg 1977;80(2):100—16.

6. Buechler CM, Blostein PA, Koestner A, et al. Variation amongtrauma centers’ calculation of Glasgow coma scale score;results of a national survey. J Trauma 1998;45:429—32.

7. Champion HR, Sacco WJ, Carnazzo AJ, et al. Trauma score.Crit Care Med 1981;9:672.

8. Champion HR, Sacco WJ, Copes WS. A revision of the traumascore. J Trauma 1989;29:623.

9. Combes P, Fauvage B, Colonna M, et al. Severe head injuries:an outcome prediction and survival analysis. Intensive CareMed 1996;22(12):1391—5.

10. Des Plaines IL. Association for the Advancement of Automo-tive Medicine. Abbreviated Injury Scale: 1990 Revision. Asso-ciation for the Advancement of Automotive Medicine; 1990.

11. Emmerman CL, Shade B, Kubincanek J. Comparative perfor-mance of the Baxt Trauma Triage Rule. Am J Emerg Med1992;10:294.

12. Hanley JA, Mc Neil BJ. Themeaning and use of the area undera receiver operating characteristic (ROC) curve. Radiology1982;29:143—9.

13. Lee EJ, Hung YC, Wang LC, et al. Factors influencing thefunctional outcome of patients with acute epidural hemato-mas: analysis of 200 patients undergoing surgery. J Trauma1998;45(5):946—52.

14. Luerssen TG, Klauber MR, Marshall LF. Outcome from headinjury related to patient’s age. A longitudinal prospectivestudy of adult and pediatric head injury. J Neurosurg1988;68(3):409—16.

15. Meredith W, Rutledge R, Hansen AR, et al. Field triage oftrauma patients based upon the ability to follow commands:a study in 29573 injured patients. J Trauma 1995;38:129—35.

16. Moini M, Rezaishiraz H, Zafarghandi MR. Characteristics andoutcome of injured patients treated in urban trauma centersin Iran. J Trauma 2000;48(3):503—7.

17. Ross SE, Leipold C, Terregino C, O’Malley KF. Efficacy of themotor component of the Glasgow Coma Scale in traumatriage. J Trauma 1998;45(1):42—4.

18. Stephenson SCR, Langley JD, Civil ID. Comparing measures ofinjury severity for use with large databases. J Trauma2002;53:326—32.

19. Teasdale G, Jennett B. Assessment of coma and impairedconsciousness. Lancet 1974;13:81.

20. Zargar M, Modaghegh MHS, Rezaishiraz H. Urban injuries inTehran: demography of trauma patients and evaluation oftrauma care. Injury 2001;32(8):613—7.