Elsevier

Injury

Volume 42, Issue 1, January 2011, Pages 83-89
Injury

Trauma and Injury Severity Score (TRISS): Is it time for variable re-categorisations and re-characterisations?

https://doi.org/10.1016/j.injury.2010.08.036Get rights and content

Abstract

Background

Despite its limitations, the Trauma and Injury Severity Score (TRISS) continues to be the most commonly used tool for benchmarking trauma outcome. Since its inception, considerable energy has been devoted to improving TRISS. However, there has been no investigation into the classification or characterisation of the TRISS variables. Using a major nationally representative database, this study aims to explore the adequacy of the existing TRISS model by investigating variable re-categorisations and alternative characterisations in a logistic model used to predict survival in adults after traumatic injury.

Materials and methods

Data were obtained from the National Trauma Data Bank National Sample Project (NSP). Each variable in the TRISS model was related to discharge status and various categorisations considered using weighted logistic regression. Categorisations were treated nominally, using a series of indicator variables. For each variable and classification level, the best category combination was ascertained using the Bayesian Information Criterion (BIC). All best 5-category classified TRISS variables were combined, as were all best 10-category classified TRISS variables, and their predictive performance assessed against two conventionally defined TRISS models on the unweighted NSP sample using area under the Receiver Operating Characteristic curve (AUC) and BIC statistics.

Results

Overall, the weighted sample included 1,124,001 adults with injury events and known discharge status, of whom 1,061,709 (94.5%) were alive at discharge. When separately related to discharge status, each re-classified TRISS variable yielded a superior BIC statistic to its original specification. When investigating predictive performance, complete information was available for 167,239 (79.9%) adults with blunt and 20,643 (82.3%) adults with penetrating injury mechanisms. AUC and BIC estimates for the re-classified TRISS models were superior to the conventionally defined TRISS models. While having better predictive precision, the complexity associated with the best 10-category model resulted in the best 5-category model being preferred for penetrating mechanism injuries and being negligibly inferior for blunt mechanism injuries.

Discussion

Substantial improvements in the predictive power of TRISS were demonstrated by re-classifying the component variables and treating the variable categories nominally. However, before a new TRISS model with updated coefficients can be published, variable interactions and the effect of missing data needs thorough statistical evaluation.

Introduction

Injury remains a public health problem of vast proportions, leading to a massive loss of health and life,20 and is estimated to cause 12–13% of the entire global burden of disease.19 While reducing preventable injury is a health priority for most agencies worldwide,17 it remains relatively under-funded and many avoidable traumatic injuries occur.19, 26 Governments, health departments, health administrators and clinicians constantly seek to improve the delivery of medical health care to reduce the mortality and morbidity associated with injuries, once sustained.27 Trauma scoring systems provide one vehicle for the benchmarking and monitoring of trauma system performance over time, between hospitals and over jurisdications.25, 27, 31

Trauma scoring systems have traditionally focused on reducing preventable deaths,3, 25 and performance monitoring has primarily involved comparing observed survival outcomes with expected norms.4 Multiple scoring systems exist.25, 30 However, despite its extensively documented limitations,6, 12, 14, 15 the Trauma and Injury Severity Score (TRISS) continues to be the most commonly used tool for benchmarking trauma outcome.12, 13 Originally proposed in 1983 to predict a patient's probability of survival,7 TRISS is a weighted combination of patient age (AGE), Injury Severity Score (ISS) and Revised Trauma Score (RTS). The TRISS coefficients (which give the variable weights) were estimated from ordinary logistic regression models in 1987,3 and then revised in 1995,6 using the American College of Surgeons Committee on Trauma coordinated Major Trauma Outcome Study (MTOS) database. More recently, in 2010, these TRISS coefficients have been further revised using data obtained from the American College of Surgeons Committee on Trauma National Trauma Data Bank (NTDB) and the NTDB National Sample Project (NSP).24

Considerable energy has been devoted to improving or refining TRISS; for example, through recalibration of the coefficients,6, 24 careful consideration of the effect of missing data,16, 24 or through specification of new or modified variables, such as the New Injury Severity Score (NISS)22. These investigations are often18, 21 but not always1, 28 fruitful, usually yielding models with incrementally superior predictive performance. However, what has not been challenged in the literature is the classification of the TRISS variables themselves or the way in which they are characterised within the ordinary logistic regression models. In the TRISS model, patient age is treated as a binary variable, ISS and RTS are both treated as continuous variables, each assumed to have a linear relationship with survival over their value range, and the RTS is a weighted linear combination of three variables (respiratory rate [RR], systolic blood pressure [SBP], Glasgow Coma Score [GCS]), each categorised into 5 groups, assigned a value from 0 to 4, and then treated as a linear continuous variable. The rationale for the age and RTS component variable categorisations in predicting survival does not appear to have been tested in the literature nor have the relatively strong linear assumptions. Since the derivation of TRISS over 30 years ago, major contemporary datasets have been established, more sophisticated statistical techniques developed, and computational capacity and power have dramatically increased. Revisiting and improving the existing TRISS variable specifications and relaxing the regression model assumptions may improve the predictive power of the TRISS model; perhaps substantially.

Using a major nationally representative database, this study aims to explore the adequacy of the existing TRISS variable categorisations and assumptions by investigating variable re-classifications and alternative variable characterisations in the logistic model used to predict survival after traumatic injury. Predictive performances of two conventionally defined TRISS models and the newly specified TRISS models will be presented and assessed.

Section snippets

Data sources: the National Trauma Data Bank (NTDB) National Sample Project (NSP)

The American College of Surgeons Committee on Trauma established the NTDB in 1997.30 Currently, the NTDB contains detailed data on over 3 million cases from over 900 United States trauma centres.30 However, like the MTOS database, the NTDB is not population based and consists solely of data submitted by participating trauma centres. It includes a disproportionate number of larger hospitals with younger and more severely injured patients,30 which may, in turn, affect the generalisability of

Results

For the NSP, 280,129 patients had valid eligible trauma codes. In these, the mechanism of injury was burns for 4854 (1.7%) or unknown for 5194 (1.9%), leaving 270,081 (96.4%) patients who provide a weighted estimate number of 1,278,563. Table 2 includes the socio-demographic and injury profile of this full weighted NSP sample (for adults and children) so that it is directly comparable with the full MTOS sample of 80,544 trauma patients from 139 United States and Canadian trauma centres,

Discussion

Substantial and important improvements in the predictive power of TRISS were demonstrated in this paper simply by re-classifying the component variables and treating the variable categories nominally, with a series of indicator variables each having their own coefficient. This adopted analytic approach employs standard contemporary epidemiological and statistical techniques, and relaxes the strong assumption associated with the continuous linear relationships inherent within the conventionally

Conclusion

While substantial improvements to TRISS have been demonstrated in this paper, further research is needed to derive an appropriate re-classified TRISS model and specify coefficients. Even if additional new potential component variables are ignored (such as: co-morbidities, body region, injury intent, NISS), a thorough statistical investigation is required studying existing variable relationships and the effect of missing data.16, 24 Currently, TRISS is essentially a main-effects binary logistic

Conflict of interest statement

The author does not have any financial or personal relationships with other people or organisations that could inappropriately influence this work.

Funding source

None.

Acknowledgements

Grateful thanks is extended to the NTDB and NSP personnel who approved and supplied the data and assisted with data related questions, including Dr. Avery Nathens, St. Michael's Hospital, Toronto, Canada; and Ms. Melanie Neal and Ms. Sandra Goble, NTDB, American College of Surgeons, Chicago, US. Thanks also is extended to Dr. Cate Cameron, Griffith University, School of Medicine, Logan, Australia; Ms. Tamzyn Davey, National Trauma Registry Consortium, Royal Australasian College of Surgeons,

References (31)

  • H.R. Champion et al.

    Trauma severity scoring to predict mortality

    World J Surg

    (1983)
  • D.E. Clark et al.

    Risk adjustment for injured patients using administrative data

    J Trauma

    (2004)
  • P. Congdon

    Bayesian statistical modelling

    (2002)
  • W.D. Dupont

    Statistical modeling for biomedical researchers: a simple introduction to the analysis of complex data

    (2002)
  • B.J. Gabbe et al.

    TRISS: does it get better than this?

    Acad Emerg Med

    (2004)
  • Cited by (32)

    • Predicting in-hospital death among patients injured in traffic crashes in Saudi Arabia

      2014, Injury
      Citation Excerpt :

      An H–L statistic <15.5 (8 degrees of freedom; p > 0.05) indicates no evidence of lack of fit [32]. Because the value of prognostic models in general tends to be overestimated, it is widely recommended that they be further validated [18,19,25,26,32,34,36,40]. We used the bootstrapping method to calculate a corrected C-statistic for the derived model because it has been shown in simulation studies to reflect all sources of model uncertainty [41,42].

    • Comparison of two comorbidity scoring systems for older adults with traumatic injuries

      2014, Journal of the American College of Surgeons
      Citation Excerpt :

      Chart reviews were conducted on all included patients via the electronic medical record and trauma registry. Data collection variables included demographic information (ie, age, sex, height, and weight), Injury Severity Score (ISS),8,9 home medications, medical history (comorbidities), blood pressure in the emergency department, presence of brain injury, living situation before trauma, transfer from an outside hospital, traumatic brain injury, and disposition after discharge. The confounding variables were chosen based on our previous data on the best model to predict mortality in trauma patients.10

    View all citing articles on Scopus
    View full text