Development and Validation of the Mortality Risk for Trauma Comorbidity Index

METHODS

The current study is a cohort study using data from the National Study on the Costs and Outcomes of Trauma (NSCOT), a multicenter, prospective cohort study of outcomes following traumatic injury conducted in 14 states in the United States.²⁶ The study protocol was approved by the institutional review boards at all participating centers. Patients were eligible for NSCOT if they were between 18 and 84 years of age, arrived alive at a participating hospital, and had at least 1 injury of 3 or higher on the Abbreviated Injury Scale (N = 18,198). A detailed description of the enrollment and data collection procedures has been published previously.^26,27 In brief, using a quota sampling strategy, a group of patients who were discharged alive (n = 8021) were selected for postinjury follow-up interviews and all patients who died in hospital (n = 1438) were included in the sample. This was used to enroll adequate numbers of patients across age groups and to ensure even distribution across hospital type (trauma/nontrauma center) and by severity and principal body region injured. To correctly represent the population, this sampling strategy required data to be weighted according to the population of eligible patients^28,29 because (a) only a sample of patients who were discharged alive was selected whereas all patients who died in-hospital were included and (b) not all selected patients were enrolled. The weights consisted of the reciprocal product of the probability of being selected for the study and the probability of being enrolled and having complete data given selection. The actual number of subjects (n) and the weighted number (wn) are presented for all analyses.

Subjects for the current study were identified from the NSCOT dataset and defined as patients who survived >24 hours postinjury to allow for the most complete information on comorbid health conditions (n = 4644/wn = 14069). The subject pool was randomly divided in half; the first half of subjects were used to derive the MoRT (n = 2322/wn = 7205), the second to validate the instrument (n = 2322/wn = 6864).

Injury Severity Score (ISS) was derived from individual anatomic injury severity scores and is a measure of the overall severity of multiple trauma and is scored on a scale of 1 (least severe) to 75 (most severe).^30,31

Comorbid health conditions within the NSCOT were collected using 3 methods in the parent study. (1) Data were abstracted from the medical record by trained research nurses on pre-existing conditions recorded in the medical record of the index hospitalization; (2) the CCI¹² for preinjury conditions was included in the 3-month interview, which was expanded to include comorbidities not included in the original CCI but that have been shown to correlate with mortality in trauma patients (eg, coagulopathy, obesity),^32–35 and (3) for adults aged 65 years and older, inpatient and outpatient Medicare claims were obtained for at least 6 months prior to injury to identify pre-existing conditions.²⁷ Scores on the CCI range from 0 to 31, with a score greater than 5 being severe.^10,12 For the purposes of this study, we began with a list of 41 pre-existing conditions that were derived from 3 commonly used instruments in the trauma and comorbidity literature: the CCI, the Elixhauser score, and the APACHE II (Tabular listings and descriptors of the conditions found in the 3 instruments can be found in Ho et al³⁶).

The primary outcome used to develop the mortality risk score was death from any cause at time of hospital discharge. No treatment-related variables were included in the analyses. Death at 1-year post injury was a secondary outcome of interest. Deaths from any cause, after discharge were documented either by proxy interview or through the National Death Index.³⁷

RESULTS

The demographic characteristics of the subjects in the derivation and validation samples were similar on all characteristics except maximum Face anatomic injury severity (Table 1). The mean age of the sample was 43 years old, were predominantly male, and had a mean ISS of about 17, indicating a moderate injury. The primary mechanism of injury for subjects was motor vehicle collision (>45%). In-hospital mortality was similar in the derivation and validation samples (4.5% vs. 5.5%).

About 28 conditions had an overall prevalence of ≥1% (Table 2). The most prevalent conditions in the sample were hypertension (20.1%), alcohol abuse (17.3%), chronic pulmonary disease (8.9%), diabetes (6.1%), and depression (5.2%). In the sample, only 2.6% of subjects had 2 or more comorbid conditions. We ran crosstabs on the 6 parameters to evaluate frequency of comorbid condition coexistence and in no case did the prevalence of exceed 1% (range, 0.02%–0.52%) therefore it was deemed noncontributory.

Nine factors were individually significantly associated with the risk of mortality at hospital discharge postinjury and included several cardiac related diseases: prior myocardial infarction, cardiac arrhythmias, hypertension, and coronary artery disease (Table 3). Nervous system and psychiatric diseases included cerebrovascular disease, dementia, and depression.

Six factors were independently associated with the risk of mortality and formed the basis for the MoRT (Table 4). Of these 6, a single factor, depression, was protective whereas the remaining 5 were associated with increased risk of mortality. The performance of the MoRT with 6 conditions in comparison to the CCI with 17 conditions is shown in Table 5. The range of the CCI scores in the validation sample was 0 to 11, whereas the MoRT ranged −4 to 8. Both scales demonstrated a right skew in the median score. The MoRT had a similar overall discrimination to the CCI for in-hospital mortality (Table 5). Adding age and gender improved the ability of both models to discriminate although this was not statistically significant (P ≥ 0.05). At 1-year postinjury, the CCI was better at detecting risk of mortality than the MoRT (0.64 vs. 0.58), but this discriminatory advantage was eliminated once gender and age were added to the model (Table 5). The further addition of ISS signifi-cantly improved the predictive ability of the MoRT (0.77, 95% CI: 0.74, 0.79) and the CCI (0.77, 95% CI: 0.75, 0.80) for in-hospital mortality (Table 5). A similar predictive ability was seen with this model (injury, age, gender, and comorbidity) for both in-hospital and 1-year postinjury in contrast to other models which did not include injury.

In an exploratory analysis, the effect of being on treatment for depression was examined. About 66% of those with comorbid depression in the sample were receiving antidepressant therapy at the time of injury. The odds of death in-hospital for depressed persons on an antidepressant were 0.29 (0.09, 0.90; P = 0.03) whereas the odds of death for those not on antidepressant were 0.25 (0.06, 1.06).

DISCUSSION

We compared the predictive ability of a tailored comorbidity index for trauma, the MoRT to the CCI on mortality at 2 different time points. We found that although individual comorbidities are prevalent and related to in-hospital mortality, comorbidity alone had a limited ability to distinguish between being alive or dead at discharge; ie, use of the indices for prediction of mortality at discharge was not much better than chance alone (0.5). Only the models that included injury severity (ISS) reached an acceptable level for predictive ability (0.7–0.8 per Hosmer and Lemeshow⁴⁰). The present finding for the predictive ability of the CCI at hospital discharge in a US population is similar to those previously reported from Australia (0.57, 95% CI: 0.54–0.61).¹⁸ At 1-year postinjury, the predictive ability of the models improved, suggesting that comorbidity alone contributes more substantively to mortality in the postacute phase and this time period offers an opportunity for improvement in care in these patients.

The MoRT was able to discriminate as well as the CCI for both hospital discharge and 1-year postinjury mortality and with only 6 factors compared with the CCI’s 17. The main advantage of the MoRT is its parsimony. Chart review is expensive and time consuming,⁴¹ and in populations such as trauma where it is difficult to obtain history prospectively, the use of only 6 conditions could be of particular benefit for both research and clinical applications.

Future predictive models should include gender, age (in decades), and ISS, as these significantly improved the performance of both the CCI and the MoRT, and these variables are easily discernable at admission and do not substantively increase data collection burden.

The weights in the MoRT are different for those conditions appearing in the CCI, likely reflecting the differences in the populations that each was developed in (medical inpatient vs. trauma). In comparing the CCI and MoRT, differential weighting is seen for severe liver disease and myocardial infarction. In the CCI, the lowest weight is given for MI (a base score of 1) whereas in the MoRT, a weight of 3 is given (or 1.5 × the base weight score). This difference is opposite in the case of severe liver disease, where the CCI provides a weight of 3 (or 3 times the base weight, whereas the MoRT provides a weight of 5 or 2.5 times the base weight. These prognostic differences are likely because of either differences in prevalence of the conditions in trauma patients versus medical patients or differential effects on hospital versus 1-year mortality.^25,39

Both comorbidity indices had better predictive ability for mortality at 1-year. Previous studies in patients post-coronary artery bypass graft surgery have also found that the effect of comorbidity on postdischarge survival is also greater than in-hospital and may reflect the low prevalence of some conditions.²⁵

In the present study, the comorbid conditions found to be predictive of mortality had some overlap with those found in the CCI, but there were 2 conditions identified which are novel predictors. Cardiac arrhythmias were found to be associated with increased risk of death, and depression was found to be protective. Although the MoRT includes cardiac arrhythmia, it does not include use of blood thinners/coagulopathy. As warfarin is usual treatment for patients with atrial fibrillation, which is a frequently occurring cardiac arrhythmia, it would have been expected to be correlated. However, we did not distinguish between cardiac arrhythmias and this is an area for further work. Additionally, in persons with acquired coagulopathy, the odds of death at hospital discharge were 2.4 times higher than those persons without the condition at time of injury although this was not significant. We do not have specific information on the measured anticoagulation status of the individuals, which has been shown to be predictive of mortality whereas self-report or administrative data may not.^9,42,43 Other reasons for acquired coagulopathy not entering the final model include a relatively low prevalence of this condition (1.1%) in the sample and our conservative method of modeling, which included entering conditions only when there was a significant univariate relationship. Recent revisions to the Centers for Disease Control field guidelines for injury triage have maintained decision criteria for patients with anticoagulation and bleeding disorders because of their increased risk of mortality.³³ Another reason in the present study for finding no relationship between acquired coagulopathy and in-hospital mortality may have been our decision to exclude early deaths from the current study. Lastly, contrary to previous studies which found hypertension to be protective,^44,45 we found pre-existing hypertension to be associated with a slightly increased odds (1.4) of death at hospital discharge. Previous studies have suggested that the protective effect of hypertension seen may be related to the effect of beta-blocker use within their samples; however as those studies did not directly address that issue, our study suggests that incomplete modeling of the involved variables (residual confounding) may actually be responsible for the results seen and further work is required to clarify this issue.

The relationship between preexisting depression and disability following injury has been previously established.^46–48 However, the finding in the present study that comorbid depression was protective was not expected. Although a protective effect cannot be ruled out, another possible explanation is bias against documenting in the medical record the diagnosis of depression in patients who died in-hospital. We further note that the present findings were not related to current treatment for depression, so this condition is not appearing to serve as a proxy for variables such as insurance status or preinjury health care not entered into the model. No data on adherence to or efficacy of prescribed antidepressant treatment are available, nor did we have data on DSM supported diagnosis or the severity of depression which may also be of importance in further understanding this relationship. In a population-based study in Canada, injured persons had a mental health claim rate in the year prior to injury 3.5 times higher than that of noninjured persons, with the majority of these claims being for conditions such as depression or anxiety.¹¹ Of those injured in that sample, more than 85% suffered only minor injury (ISS, 1–8), although no data were specifically provided on the relationship between preexisting depression and severity of injury. Further work is needed to understand the relationship between preexisting depression, level of injury, and mortality to understand whether injury-related or treatment-related variables are mediating this effect.

A previously identified issue with the CCI has been right skew in the distribution with relatively low median scores.¹⁰ This also occurred with the MoRT, despite restricting to conditions with >1% prevalence in the sample. In addition, the conditions that the MoRT identifies as predictive of mortality are also those that are increasingly prevalent with age eg, dementia, myocardial infarction, cardiac arrhythmias, and cerebrovascular disease. Trauma remains a young person’s disease, so the right skew in the comorbidity scores seen in this study is not wholly unexpected. Although the parent NSCOT study did not include persons over 85 in whom conditions like dementia are more prevalent, the present study restricted prediction of mortality to those persons <85 so this was not a factor in the predictive ability of the comorbidity indices alone. In fact, the prevalence rates reported for comorbid conditions in the CCI in the present report are similar to those reported in a recent study of injured persons 65 and older in Italy.¹⁹ Our a priori decision to evaluate the combined association of dementia with other comorbid conditions was primarily based on our prior traumatic brain injury work and was exploratory within the context of all-cause trauma. However, no additional interaction effects were seen between any comorbidity combinations tested and were noncontributory in the final model.

The use of split halves allowed for improved validity by reduced bias, and reducing model potential over-fitting. In addition, the NSCOT is a large, multicenter study which improves the generalizability of the findings to the larger US population. Although we would have preferred to have validated the MoRT in a separate dataset, this was not possible. We attempted to use the National Trauma Data Bank (NTDB) research dataset for validation purposes, but neither cardiac arrhythmia nor depression are specifically measured within the NTDB currently. Ultimate verification of the MoRT and its predictive utility will require testing it with an independent dataset, preferably one such as the NTDB with multiple data sources and all ages, including the “oldest old” represented.

Other limitations of the current study include the exclusion of other case-mix variables from our prediction models. Recent robust predictive models have been based on both injury-related factors and socioeconomic and demographic factors.^18,49,50 However, the point of the MoRT was to develop an injury-specific comorbidity index and to compare it to the CCI. Future case-mix studies using the MoRT could build additional predictive models to include variables such as socioeconomic status and preinjury functioning.

The inclusion of subjects who survived greater than 24 hours postinjury may have resulted in some loss of patients who die as a result of an interaction between comorbid health conditions and injury severity. However, information on comorbidity in these patients is likely to have had incomplete or missing documentation, therefore any analysis that included these patients would likely have underestimated this effect.

CONCLUSIONS

The MoRTs primary advantage over current instruments is its parsimony, containing only 6 items compared with the CCI. In the present study, the comorbid conditions found to be predictive of mortality had some overlap with the CCI, but this study identified 2 novel predictors: cardiac arrhythmias and depression. Inclusion and reporting of these items within trauma registries would therefore be an important step to allow further validation and use of the MoRT.