TRISS Model Research Articles

Prediction of mortality among severely injured trauma patients A comparison between TRISS and machine learning-based predictive models

BackgroundGiven the huge impact of trauma on hospital systems around the world, several attempts have been made to develop predictive models for the outcomes of trauma victims. The most used, and in many studies most accurate predictive model, is the “Trauma Score and Injury Severity Score” (TRISS). Although it has proven to be fairly accurate and is widely used, it has faced criticism for its inability to classify more complex cases. In this study, we aimed to develop machine learning models that better than TRISS could predict mortality among severely injured trauma patients, something that has not been studied using data from a nationwide register before. MethodsPatient data was collected from the national trauma register in Sweden, SweTrau. The studied period was from the 1st of January 2015 to 31st of December 2019. After feature selection and multiple imputation of missing data three machine learning (ML) methods (Random Forest, eXtreme Gradient Boosting, and a Generalized Linear Model) were used to create predictive models. The ML models and TRISS were then tested on predictive ability for 30-day mortality. ResultsThe ML models were well-calibrated and outperformed TRISS in all the tested measurements. Among the ML models, the eXtreme Gradient Boosting model performed best with an AUC of 0.91 (0.88–0.93). ConclusionThis study showed that all the developed ML-based prediction models were superior to TRISS for the prediction of trauma mortality.

Validation of the Conventional Trauma and Injury Severity Score and a Newly Developed Survival Predictive Model in Pediatric Patients with Blunt Trauma: A Nationwide Observation Study.

To date, there is no clinically useful prediction model that is suitable for Japanese pediatric trauma patients. Herein, this study aimed to developed a model for predicting the survival of Japanese pediatric patients with blunt trauma and compare its validity with that of the conventional TRISS model. Patients registered in the Japan Trauma Data Bank were grouped into a derivation cohort (2009-2013) and validation cohort (2014-2018). Logistic regression analysis was performed using the derivation dataset to establish prediction models using age, injury severity, and physiology. The validity of the modified model was evaluated by the area under the receiver operating characteristic curve (AUC). Among 11 predictor models, Model 1 and Model 11 had the best performance (AUC = 0.980). The AUC of all models was lower in patients with survival probability Ps < 0.5 than in patients with Ps ≥ 0.5. The AUC of all models was lower in neonates/infants than in other age categories. Model 11 also had the best performance (AUC = 0.762 and 0.909, respectively) in patients with Ps < 0.5 and neonates/infants. The predictive ability of the newly modified models was not superior to that of the current TRISS model. Our results may be useful to develop a highly accurate prediction model based on the new predictive variables and cutoff values associated with the survival mortality of injured Japanese pediatric patients who are younger and more severely injured by using a nationwide dataset with fewer missing data and added valuables, which can be used to evaluate the age-related physiological and anatomical severity of injured patients.

Evaluation of quality of care in trauma patients using international scoring systems

ObjectiveTo evaluate the ability of the TRISS and PS14 models to predict mortality rates in our medical system and population. DesignA retrospective observational study was carried out over a 66-month period. BackgroundThe study was conducted in the Trauma Intensive Care Unit (ICU) of a third level hospital. PatientsAll severe trauma patients (Injury Severity Score≥16 and/or Revised Trauma Score<12) aged >14 years were included. Variables of interestMedical care data were prospectively recorded. The “W” statistic (difference between expected and observed mortality for every 100 patients) and its significance were calculated for each model. Discrimination and calibration were evaluated by means of receiver operating characteristic (ROC) curves, and the Hosmer–Lemeshow test and GiViTI calibration belt, respectively. ResultsA total of 1240 patients were included. Survival at hospital discharge was 81.9%. The “W” scores for the TRISS, TRISS 2010 and PS14 models were +6.72 (P<.01), +1.48 (P=.08) and +2.74 (P<.01) respectively. Subgroup analysis revealed significant favorable results for some populations. The areas under the ROC curve for the TRISS, TRISS 2010 and PS14 models were 0.915, 0.919 and 0.914, respectively. There were no significant differences among them (P>.05). Both the Hosmer–Lemeshow test and GiViTI calibration belt demonstrated poor calibration for the three models. ConclusionsThese models are suitable tools for assessing quality of care in a Trauma ICU, affording excellent discrimination but poor calibration. In our institution, survival rates higher than expected were observed.

Evaluation of quality of care in trauma patients using international scoring systems.

To evaluate the ability of the TRISS and PS14 models to predict mortality rates in our medical system and population. A retrospective observational study was carried out over a 66-month period. The study was conducted in the Trauma Intensive Care Unit (ICU) of a third level hospital. All severe trauma patients (Injury Severity Score≥16 and/or Revised Trauma Score <12) aged> 14 years were included. Medical care data were prospectively recorded. The "W" statistic (difference between expected and observed mortality for every 100 patients) and its significance were calculated for each model. Discrimination and calibration were evaluated by means of receiver operating characteristic (ROC) curves, and the Hosmer-Lemeshow test and GiViTI calibration belt, respectively. A total of 1240 patients were included. Survival at hospital discharge was 81.9%. The "W" scores for the TRISS, TRISS 2010 and PS14 models were+6.72 (P<.01), +1.48 (P=.08) and +2.74 (P<.01) respectively. Subgroup analysis revealed significant favorable results for some populations. The areas under the ROC curve for the TRISS, TRISS 2010 and PS14 models were 0.915, 0.919 and 0.914, respectively. There were no significant differences among them (P>.05). Both the Hosmer-Lemeshow test and GiViTI calibration belt demonstrated poor calibration for the three models. These models are suitable tools for assessing quality of care in a Trauma ICU, affording excellent discrimination but poor calibration. In our institution, survival rates higher than expected were observed.

Validating performance of TRISS, TARN and NORMIT survival prediction models in a Norwegian trauma population.

IntroductionAnatomic injury, physiological derangement, age, injury mechanism and pre‐injury comorbidity are well‐founded predictors of trauma outcome. Statistical prediction models may have poorer discrimination, calibration and accuracy when applied in new locations. We aimed to compare the TRISS, TARN and NORMIT survival prediction models in a Norwegian trauma population.MethodsConsecutive patients admitted to Oslo University Hospital Ullevål within 24 h after injury, with Injury Severity Score ≥ 10, proximal penetrating injuries, or received by trauma team, were studied. Original NORMIT coefficients were updated in a derivation dataset (NORMIT 2; n = 5923; 2005–2009). TRISS, TARN and NORMIT prediction models were evaluated in the validation dataset (n = 6348; 2010–2013) using two different AIS editions for injury coding. Exclusion due to missing data was 0.26%. Outcome was 30‐day mortality. Validation included AUROC, scaled Brier statistics, and calibration plots.ResultsThe NORMIT models had significantly better discrimination, calibration, and overall fit than the TRISS 09, TARN 09 and TARN 12 models. The updated NORMIT 2 had higher numerical values of AUROC and scaled Brier than the original NORMIT, but with overlapping 95%CI. Overlapping 95%CI for AUROCs and Discrimination slopes indicated that the TARN and TRISS models performed similarly. Calibration plots showed tight and consistent predictions over all Ps strata for NORMIT 2 run on AIS'98 coded data, and only little deterioration when AIS'08 data was substituted.ConclusionsIn a Norwegian trauma population, the updated Norwegian survival prediction model in trauma (NORMIT 2) performed better than well‐established British and US alternatives. External validation of these three models in other Nordic populations is warranted.

Validation of the probability of survival using the triss methodology in the Spanish Trauma ICU registry (RETRAUCI)

Data obtained from large databases in severe trauma patients generated the opportunity to develop different models of probability of survival. These models are supported by mathematics-based severity scales. Using these models in regional areas different from those where they were developed makes its validation mandatory. The TRISS model, which combines variables of RTS, ISS, age and type of injury allows obtaining the probability of survival.

Sudden survival improvement in critical neurotrauma: An exploratory analysis using a stratified statistical process control technique

BackgroundOutcome after trauma depends on patient characteristics, quality of care, and random events. The TRISS model predicts probability of survival (Ps) adjusted for Injury Severity Score (ISS), Revised Trauma Score (RTS), mechanism of injury, and age. Quality of care is often evaluated by calculating the number of “excess” survivors, year by year. In contrast, the Variable Life-Adjusted Display (VLAD) technique allows rapid detection of altered survival. VLAD adjusts each death or survival by the patient's risk status and graphically displays accumulated number of unexpected survivors over time. We evaluated outcome changes and their time relation to trauma service improvements. MethodsObservational, retrospective study of the total 2001–2011 trauma population from a Level I trauma centre. Outcome was 30-day survival. Ps was calculated with the TRISS model, 2005 coefficients. VLAD graphs were created for the entire population and for subpopulations stratified by ISS level, ISS body region (Head/Neck, Face, Chest, Abdomen/Pelvic contents, Extremities/Pelvic girdle, External), and maximum Abbreviated Injury Scale (maxAIS) score in each region. Piecewise linear regression identified VLAD graph breakpoints. Results12,191 consecutive trauma patients (median age 35 years, 72% males, 91% blunt injury, 41% ISS≥16) formed the dataset. Their VLAD graph indicated performance equal to TRISS predicted survival until a sudden improvement in late 2004. From then survival remained improved but unchanged through 2011. Total number of excess survivors was 141. Inspection of subgroup VLAD graphs showed that the increased survival mainly occurred in patients having at least one Head/Neck AIS 5 injury. The effect was present in both isolated and multitraumatised maxAIS 5 Head/Neck trauma. The remaining trauma population showed unchanged survival, superior to TRISS predicted, throughout the study period.Important general and neurotrauma-targeted improvements in our trauma service could underlie our findings: A formalised trauma service, damage control resuscitation protocols, structured training, increased helicopter transfer capacity, consultant-based neurosurgical assessment, a doubling of emergency neurosurgical procedures, and improved neurointensive care. ConclusionsStratified VLAD enables continuous, high-resolution system analysis. We encourage trauma centres to explore their data and to monitor future system changes.

Burden of traumatic injuries in Saudi Arabia: lessons from a major trauma registry in Riyadh, Saudi Arabia.

BACKGROUND AND OBJECTIVESIn Saudi Arabia (SA), injuries are the second leading cause of death; however, little is known about their frequencies and outcomes. Trauma registries play a major role in measuring the burden on population health. This study aims to describe the population of the only hospital-based trauma registry in the country and highlight challenges and potential opportunities to improve trauma data collection and research in SA.DESIGN AND SETTINGSUsing data between 2001 and 2010, this retrospective study included patients from a large trauma center in Riyadh, SA.PATIENTS AND METHODSA staff nurse utilized a structured checklist to gather information on patients’ demographic, physiologic, anatomic, and outcome variables. Basic descriptive statistics by age group (≤14 vs >14 years) were calculated, and differences were assessed using student t and chi-square tests. In addition, the mechanism of injury and the frequency of missing data were evaluated.RESULTS10 847 patients from the trauma registry were included. Over 9% of all patients died either before or after being treated at the hospital. Patients who were older than 14 years of age (more likely to be male) sustained traffic-related injuries and died in the hospital as compared to patients who were younger than or equal to years of age. Deceased patients were severely injured as measured by injury severity score and Glasgow Coma Scale (P<.001). Overall, the most frequent type of injury was related to traffic (52.0%), followed by falls (23.4%). Missing values were mostly prevalent in traffic-related variables, such as seatbelt use (70.2%).CONCLUSIONThis registry is a key step toward addressing the burden of injuries in SA. Improved injury classification using the International Classification of Disease-external cause codes may improve the quality of the registry and allow comparison with other populations. Most importantly, injury prevention in SA requires further investment in data collection and research to improve outcomes.

Outcomes of traumatic brain injury in Hong Kong: Validation with the TRISS, CRASH, and IMPACT models

We aimed to test prognostic models (the Trauma Injury Severity Score, International Mission for Prognosis and Analysis of Clinical Trials in Traumatic Brain Injury, and Corticosteroid Randomisation After Significant Head Injury models) for 14-day mortality, 6-month mortality, and 6-month unfavorable outcome in a cohort of trauma patients with traumatic brain injury (TBI) in Hong Kong. We analyzed 661 patients with significant TBI treated in a regional trauma centre in Hong Kong over a 3-year period. The discriminatory power of the models was assessed as the area under the receiver operating characteristic curve. One-sample t-tests were used to compare actual outcomes in the cohort against predicted outcomes. All three prognostic models were shown to have good discriminatory power and no significant systemic over-estimation or under-estimation. In conclusion, all three predictive models are applicable to eligible TBI patients in Hong Kong. These predictive models can be utilized to audit TBI management outcomes for trauma service development in the future.

Outcome prediction in trauma

In the Trauma Audit and Research Network (TARN), currently the largest trauma network in Europe, outcome prediction is performed using the TRISS methodology since 1989. Its database contains 200,000 hospital admissions from 110 hospitals over the country, but a large amount of data is lost for the modelling because of missing data. To improve some of the shortcomings of TRISS a new model was developed. The data for modelling consisted of 100,399 hospital trauma admissions over the period 1996 to 2001. Using the Glasgow Coma Score (GCS) instead of RTS has dramatically reduced the number of missing cases. Gender and its interaction with age have also been included in the model. The model was tested on different subsets of cases traditionally excluded, such as children, those with penetrating injuries, and ventilated and transferred patients. The new model included all those subsets using age, a transformation of ISS, GCS, gender and gender by age interaction as predictors. The model has shown a good discriminant ability tested by the area under the receiver operating characteristic (AROC) curve. The values of the AROC for the new model were 0.947 (95% CI: 0.943-0.951) on the prediction set and 0.952 (95% CI: 0.946-0.957) on the validation set compared respectively with 0.937 (95% CI: 0.932-0.943) and 0.941 (95% CI: 0.936-0.952) for TRISS. The new model has enabled us to include most of the cases that were excluded under the TRISS's inclusion criteria, less missing data are incurred and the predictive performance was significantly better than that of the TRISS model as shown by the AROC curves.

A New Approach to Outcome Prediction in Trauma: A Comparison With the TRISS Model

The Trauma Audit & Research Network (TARN) has been using the TRISS methodology since 1989. Its database contains 200,000 hospital admissions from 110 hospitals over the country. To improve outcome prediction, a revision of the current model became necessary. Our model tried to overcome some of the concerns of the trauma community, namely missing data, functional form of the predictors, inclusion criteria and patient's death within 30 days. The data for modeling consisted of 100,399 anonymized hospital trauma admissions during the period 1996 to 2001. Cross validation was performed on this data set, and a multiple logistic regression model was derived using the prediction set and then its prediction ability was tested on the validation set. Fractional polynomials modeling showed that the linear functional form of the Injury Severity Score (ISS) in the model was not satisfactory. Using the Glasgow Coma Score (GCS) instead of the revised trauma score (RTS) has dramatically reduced the number of missing cases. Sex and its interaction with age have also been included in the model. The model was tested on different subsets of cases, traditionally excluded, such as children, those with penetrating injuries, and ventilated and transferred patients. The new model included all those subsets using age, a transformation of ISS, GCS, sex, and sex by age interaction as predictors. The model has shown a good discriminant ability tested by the Area under the Receiver Operating Characteristic (AROC) curve. The values of the AROC for the new model were 0.947 (95% confidence interval [CI]: 0.943-0.951) on the prediction set and 0.952 (95% CI: 0.946-0.957) on the validation set compared, respectively, with 0.937 (95% CI: 0.932-0.943) and 0.941 (95% CI: 0.936-0.952) for TRISS. The new model has enabled us to include most of the cases that were excluded under the TRISSs inclusion criteria, less missing data are incurred and the predictive performance was significantly better than that of the TRISS model as shown by the AROC curves.

Pediatric Trauma Deaths: A Three-Part Analysis from a Nonacademic Trauma Center

A three-part analysis was undertaken to assess pediatric trauma mortality in a nonacademic Level II trauma center at Parkview Hospital in Fort Wayne, Indiana. Part I was a comparison of Parkview trauma registry data collected from 1999 through 2003 with those of pediatric and adult trauma centers in Pennsylvania. The same methodology used in Pennsylvania was used for the initial evaluation of pediatric deaths from trauma in our trauma center. Part II was a formal in-depth analysis of all individual pediatric deaths as well as surgical cases with head, spleen, and liver injuries from the same time frame. Part III proposes a new methodology to calculate a risk-adjusted mortality rate based on the TRISS model for the evaluation of a trauma system. The use of specific mortality and surgical intervention rates was not an accurate reflection of trauma center outcome. The proposed risk-adjusted mortality rate calculation is perhaps an effective outcome measure to assess patient care in a trauma system.

A new "TRISS-like" probability of survival model for intubated trauma patients.

Prehospital trauma patient field intubations and paralyzations, using neuromuscular blocking agents before emergency department respiratory and neurologic assessments are made, bias assessments and outcome evaluations using probability-of-survival models, such as TRISS and A Severity Characterization of Trauma (ASCOT). We present a newly developed "TRISS-like" probability-of-survival model for intubated blunt- and penetrating-injured patient assessment. From a population of 51397 consecutively admitted trauma patients, this study used all 5740 (11.2% of the total injured population) intubated patients with complete data from a statewide trauma registry from October 1, 1993, to September 30, 1996. Model performance was evaluated using standard calibration and discrimination measures and z and W statistics of significance. The new model accurately predicted survival for blunt- and penetrating-injured intubated patients and is applicable to 11 etiologic patient populations. Study findings suggest that the new TRISS-like model should be used to assess both blunt- and penetrating-injured intubated patients. Use of this new model provides an analytical method for addressing a significant limitation of both the standard TRISS and ASCOT models, which are not applicable to intubated injured patient assessment. In addition, use of this model will complement TRISS/ASCOT assessments of nonintubated trauma patients and thus permit appropriate assessments for both intubated and nonintubated injured patient study populations.

A simple mathematical modification of TRISS markedly improves calibration.

TRISS has reigned as the preeminent trauma outcome prediction model for 20 years. Despite this endorsement, the calibration of TRISS has been poor in most data sets where it has been examined. We hypothesized that the lack of calibration of TRISS was because of the inappropriate mathematical specification of the model that TRISS is based on, rather than the predictors in the model. In particular, we hypothesized that the nonlinearity of the Injury Severity Score (ISS) in the log odds of death was responsible for the poor calibration of TRISS, and further, that this nonlinearity could be corrected by the simple addition of an ISS squared term to the TRISS model. We examined ISS in the log odds of mortality for linearity in one large trauma data set, the National Pediatric Trauma Registry (NPTR) (n = 53,113 from 1985-1996; mortality, 1.3%); and two small data sets, the University of New Mexico (UNM) (n = 3,142 from 1991-1995; mortality, 8.6%) and Portland, Oregon (PORT) (n = 2,916 from 1990-1994; mortality, 1.75%). In addition, in the NPTR we compared the calibration of TRISS models with and without linearity in the log odds of death. In the NPTR, ISS was profoundly nonlinear in the log odds of death for both blunt and penetrating trauma (p < 0.001). Moreover, the overall calibration of the TRISS model for the NPTR data was significantly improved when the nonlinearity of ISS was corrected by the addition of a quadratic ISS term as demonstrated by a 70% reduction (improvement) in the Hosmer-Lemeshow statistic. Interestingly, the addition of the ISS squared term did not affect the discrimination of the model. The log odds of survival in the UNM and PORT data sets were also better modeled when an ISS squared term was added (UNM, p = 0 0.052; PORT, p = 0.014), but improvements in the Hosmer-Lemeshow statistic were smaller, possibly because of the small size of these data sets. The TRISS model for outcome prediction currently uses ISS in a mathematically inappropriate way that impairs the calibration, but not the discrimination, of its predictions. If TRISS is to continue as the prediction standard for trauma, a quadratic ISS term must be added to the model. In the future, outcome prediction models should undergo thorough statistical modeling and evaluation before being released. Injury severity descriptors other than ISS (such as ASCOT, ICISS, or NISS) may require other modeling techniques to optimize the calibration of survival models that use these injury scores.

Predicting Survival after Trauma: a Comparison of TRISS and ASCOT in the Netherlands

Background: Evaluating the performance of a trauma system may be attempted by comparing outcome in different trauma populations. Controlling for injury severity is a necessity for such evaluations. We compare two current models for doing so: the “Trauma and Injury Severity Score” (TRISS) and “A Severity Characterization Of Trauma” (ASCOT). Material and Methods: This study of high-energy trauma victims took place in Leiden, the Netherlands, between 1993 and 1998. Using the Hosmer-Lemeshow (HL) test and receiver operator characteristic (ROC) analysis, the TRISS and ASCOT models were compared for calibration and discrimination. Results: 1,024 patients, with an average Injury Severity Score (ISS) of 13.5, were eligible for inclusion. Blunt trauma was the predominant cause of injuries. Both models gave accurate, though pessimistic, results in predicting the actual number of fatalities (n = 71). The HL test indicated a sufficient fit for the ASCOT model (p = 0.28) and an insufficient fit (p = 0.02) for TRISS. The ROC curves were nearly identical (0.97). Including age as a linear variable, instead of using the current age groups, resulted in an improved discriminative power of the models. Conclusions: The ASCOT model proved superior over TRISS in its accuracy to estimate of survival chances. This difference was most evident for victims with an estimated survival chance of 60–90%. Future national trauma researchers should therefore collect ASCOT data. Improved ASCOT models could be developed, with age as a linear variable.

Low-impact falls: demands on a system of trauma management, prediction of outcome, and influence of comorbidities.

Falls from a low height are an extremely common source of injury, the severity of which is often underestimated. As a result, low fall patients are usually not transferred to Level I trauma centers. There are surprisingly few systematic data relating to the demands made on systems of trauma care by patients with low falls. This study addresses this issue using information from a comprehensive national trauma database. The performance of TRISS methodology, and the factors associated with prolonged hospital stay, in low fall patients is also examined. The study included 31,419 patients. Patients with low falls (< 2 m) were compared with those suffering high falls (> or = 2 m), motor vehicle crashes, assault, sports injuries, and a group with unclassified injuries. Probability of survival was estimated using TRISS, and its performance in different types of injury was assessed using measures of discrimination and calibration. The influence of coexistent medical conditions on mortality and length of stay was investigated using logistic regression. Low falls accounted for 45.5% of all admissions, and 43.9% of the total bed days. The low fall group was older (mean age, 61.6 years), and predominantly female (62.5%) in contrast to the other groups (both p < 0.001). There were fewer severely injured patients than in all of the other groups except sports injuries. The area under the receiver operating characteristic curve for TRISS applied to low falls (0.874) was less than that for high falls (0.969), motor vehicle crashes (0.973), assaults (0.960), sports (1.000), and unclassified injuries (0.965). Also, the calibration of the TRISS model was poor for patients with low falls. A logistic regression model derived from a training set of 5,000 patients gave slightly improved discrimination and markedly improved calibration when compared with TRISS. Although there was a strong relationship between the number of coexistent medical conditions and the risk of dying after a low fall, including data on comorbidities in a predictive model did not improve performance. Prolonged stay (defined as greater than the 90th centile, 23 days) was more likely in women (p < 0.005), or with advanced age (p < 0.001) or low initial calculated probability of survival (p < 0.001). Cardiovascular and central nervous system diseases and diabetes were associated with longer hospital stay (all p < 0.001). A logistic regression model using TRISS variables and comorbidity data gave poor prediction of prolonged stay. There was considerable variation in the length of stay between institutions. Patients with low falls make considerable demands on a system of trauma care. TRISS methodology performs less well in this group than with other types of injury. Chronic medical conditions are associated with increased mortality and more prolonged stay after a low fall. Between-institutional variation in length of stay was considerable and this, along with the poor performance of predictive models derived from routinely collected clinical data, make it unlikely that length of stay could be used as a measure of institutional performance. More robust audit measures for patients with low falls are required.

Identification of patients to include in trauma audit: a modification of the pre-chart

Trauma audit is commonly focused using a Pre-Chart to illustrate calculations made using the TRISS model. A line is drawn at Ps=0.5 to divide expected survivors and nonsurvivors. The use of this cut-off in a severely injured population was examined. The ‘M statistic’ for a group of injured patients selectively triaged to a Trauma Centre was calculated. The ideal cut-off point between predicted outcomes when using the TRISS model to focus trauma audit in this population was determined using a Receiver Operating Characteristic (ROC) curve. For this population the TRISS ‘M statistic’ was 0.71 (indicating a significantly different case mix from the reference database) and the best cut-off point was at Ps=0.76. Trauma audit in populations with a case mix of injury severity different from the reference database should use a different Ps line to define unexpected outcomes.

Differential performance of TRISS-like in early and late blunt trauma deaths.

(1) To independently validate the Trauma and Injury Severity Score-Like (TRISS-Like) model derived by Offner et al. (Revision of TRISS for intubated patients. J Trauma. 1992;32:32-35) in a population of Canadian blunt trauma victims, and (2) to compare the ability of this model to predict mortality in early and late trauma deaths. Prospective cohort of blunt trauma cases with Injury Severity Score > 12 identified from the Ontario Trauma Registry over a 5-year period. The TRISS-Like model consisting of age, Injury Severity Score, systolic blood pressure, and best motor response of the Glasgow Coma Scale was evaluated as to its ability to predict mortality by determining the sensitivity, specificity, and the area under the receiver operating characteristic curve. The sample was then divided into early (< or = 7 days) and late mortality subgroups in which model performance was evaluated with respect to time of death. A total of 7,703 patients were included in this analysis. The overall mortality was 12.3%. The TRISS-Like model allowed for assessment of an additional 23% of patients than would standard TRISS and performed with a sensitivity of 97.1%, specificity of 39.8% and an area under the receiver operating characteristic curve of 0.873. Analysis of mortality with respect to time demonstrated that 75% of deaths occurred by day 7. The specificity and receiver operating characteristic area increased in the early (< or = 7 days) subgroup, 46.5% and 0.935, respectively, compared with 20.8% and 0.778 in the late mortality group. TRISS-Like demonstrated similar performance to that reported with the standard TRISS model but with the additional advantage that it is more generalizable because it can be applied to intubated patients. TRISS-Like demonstrated substantially superior performance in early trauma deaths compared with those that occurred late. This differential performance may be because the model does not include risk factors for late mortality.