Mortality After Discharge From Acute Care Hospitalization With Traumatic Brain Injury: A Population-Based Study

Mortality and causes of death 

Vital status during the study period was determined by linking cohort cases to death certificate data from the Colorado Death Master File, which captures all deaths occurring in Colorado and out-of-state deaths of Colorado residents. For persons identified as dead, death certificate data provided the date, manner, and up to 11 causes of death (ICD-9 codes for deaths occurring in 1998; ICD-10 codes for death after 1998), 1 of which was designated as the underlying cause. Prior to data analysis, ICD-9 codes were converted to ICD-10 codes using a crosswalk file provided by the National Center for Health Statistics.15 A comprehensive list of ICD-10 codes from the World Health Organization is available elsewhere.16 To avoid overestimation of deaths caused by external causes attributed to the TBI itself, cases with TBI as the only cause of death indicated were recoded to “unknown” cause of death. Where TBI was indicated as the primary cause of death among a list of several causes of death, the next available non-TBI cause of death was recoded to become the primary cause of death for the purposes of this analysis.

Demographics 

The CTBISS data included age at injury (in years), sex, and race/ethnicity. The original dataset contained 5898 cases (27%) with unknown race. For these, race/ethnicity data were supplemented with linkage to demographic information from the Colorado Trauma Registry, which provided detailed data from Colorado's Level I, II, and III trauma centers, as reported to CDPHE. This linkage enabled the recoding of 66% of the unknown race/ethnicity cases to a known race or ethnicity and reduced the proportion of the “other/unknown” cases to 10% overall. Expected source of hospitalization payment was collapsed into 4 categories: private health insurance, liability insurance (automotive), government pay (ie, Medicare, Medicaid, Civilian Health and Medical Program of the Uniformed Services, Colorado Medically Indigent Program or other government payment source), and self-pay/no charge. Only the primary payor was included in the dataset. Also included was residence at the time of hospitalization, which was assigned an MSA classification per the U.S. Office of Management and Budget designations17 and collapsed to metropolitan (ie, “Denver-Boulder 6-county consolidated MSA” and “other MSA”) versus nonmetropolitan (ie, “non-MSA”) for analysis. Additional community-level sociodemographic variables from the 2000 U.S. Census (including median income and population) were linked to individual cases by zip code of residence at the time of hospitalization.

Injury severity and external cause of injury 

CTBISS data included up to 15 ICD-9-CM diagnoses associated with the acute hospitalization. External cause of injury was extracted from ICD-9-CM external cause of injury codes (E-codes) provided among the available diagnoses. AIS scores18 were computed using ICDMAP 90 softwarea and added to the CTBISS by CDPHE personnel. The AIS has the following values and ratings: 1 (minor), 2 (moderate), 3 (severe), 4 (serious), 5 (critical), and 6 (unsurvivable). The maximum AIS score assigned to the head region served as the indicator of TBI severity. For risk factor modeling, maximum head AIS was collapsed into 3 categories according to the distribution of cases in the dataset: minor/moderate (AIS=1 or 2), severe/serious (AIS=3 or 4), and critical (AIS=5). Cases coded as unsurvivable (AIS=6; n=8) were not included in severity analyses. There were also 270 cases with AIS rating of 9 (unknown) and 132 cases not assigned an AIS because of insufficient data, all of which were excluded from the severity analyses.

Acute care measures 

Also included in CTBISS data were the length of acute hospital stay (in days) and discharge destination, which was collapsed into 4 levels: home with no follow-up care required; home with outpatient care; inpatient nursing/rehabilitation; and an “other” category consisting of swing beds, home intravenous care service, and other miscellaneous types of care.

Comorbid health conditions and associated injuries 

Non-TBI diagnoses (per ICD-9-CM) were classified using an adaptation of the Elixhauser comorbidity framework, which delineated 31 types of comorbid health conditions. The coding algorithm, developed by researchers at University of Manitoba, has produced estimates of comorbidity prevalence in administrative data similar to other algorithms, but potentially outperforms the other algorithms in predicting mortality.19, 20 Variables were created to indicate overall comorbidity (ie, the number of comorbid health conditions present during hospitalization) and associated injuries to other body regions per AIS scores (ie, the number of injured regions other than the head). For analysis, comorbid health conditions and associated injuries were each collapsed into 4 levels: 0, 1, 2, or ≥3 conditions or injuries.

Standardized mortality ratios and life expectancy 

Relative mortality among TBI survivors compared with the general population was estimated using the ratio of observed to expected deaths: the SMR.21 CDPHE provided year 2000 rates for all-cause and cause-specific mortality for the Colorado population, stratified by age, race/ethnicity, and sex. Expected mortality was calculated by applying the Colorado mortality rate to each year of follow-up for each person in the study matched by age, race/ethnicity, and sex, and then summing the result. For estimates of mortality prior to 1 year, the monthly expected mortality rate computed as one twelfth of the yearly expected rate. All-cause SMR was calculated for the entire cohort and for separate subgroups based on stratifications of postdischarge period, age, sex, race/ethnicity, injury severity, and underlying cause of death, respectively. Underlying causes selected for this analysis were those that accounted for at least 5% of deaths in the cohort, as well as those identified in previous literature as elevated among persons with TBI.10 A 95% CI based on the chi-square statistic was calculated for each SMR.

Life expectancy reduction among the TBI cohort, stratified by age group, race/ethnicity, and sex, was estimated with a method developed for the spinal cord injury population, used in previous research.1, 22 In brief, this method uses the most recently reported population mortality rates (for this study, Colorado residents in 2005) and the estimated all-cause SMR to compute cumulative probabilities of survival for selected demographic strata (eg, probability of 20-year-old Hispanic men surviving 1 year, 2 years, 3 years, and so on up to age 100 years). The individual cumulative probabilities of survival are then summed to obtain life expectancy for the selected stratum.

Supplemental analyses (early deaths, associated injuries) 

To characterize early deaths after hospitalization with TBI, supplemental chi-square analyses were performed to examine differences in injury severity before and after 1 month postdischarge. Risk factor modeling prompted additional chi-square analyses to characterize external causes of injury in relation to the presence of non-TBI associated injuries.

Risk factor modeling 

Cox PH models23, 24 were used to identify risk factors for mortality among the TBI cohort. The outcome of interest was number of days from hospital discharge to death. Those still alive as of December 31, 2005, were censored for data analysis. The PH models were stratified by age group to minimize violations of proportionality caused by differences in cause and severity of TBI across age3, 4 as well as disease comorbidity. Six age groups were delineated by examination of Kaplan-Meier hazard curves: infants (0–3y); children/adolescents (4-19y); and young (20-39y), middle-age (40-59y), older (60-79y), and elderly (≥80y) adults. Modeling was not performed for infants with TBI because mortality was very low: of the 791 infants in the cohort, only 9 (1.1%) died during the study period. Hence, models for the remaining 5 age groups were developed.

To reduce confounding and create a more parsimonious model, each covariate was grouped into 1 of 3 domains for preliminary model building: (1) demographics, (2) injury-related/acute care measures, or (3) comorbidities. Within each domain, a Cox PH model was estimated for each of the covariates, controlling for age. Age was included in construction of all models because of its importance as a predictor of mortality.7, 9, 10 Covariates with a P value of .20 or smaller in the above univariate models were then included in a domain-specific multivariable Cox PH model. Thus, at this stage of analysis, there were 3 Cox PH models of interest, 1 for each domain. To arrive at a more parsimonious model for each domain, backward elimination was performed where the covariate with the highest P value greater than .10 was removed until only covariates with P values less than .10 remained in the domain-specific models.21, 24 Finally, to assess the influence of the domains on each other, the reduced multivariable domain-specific models were then combined into a full, adjusted model for each age group. All statistical analyses were performed using SAS 9.1 for Windows.b HRs and 95% CIs are reported for the reduced domain-specific models (demographics, injury-related/acute care measures, comorbidities) and for full, adjusted models incorporating all significant predictive factors from each of the 3 individual domains. Interpretations of the results are based on the full models.

All-cause standardized mortality ratio 

Over the entire study period, the observed risk of death among patients with TBI was almost 2.5 times the expected risk of those in the general population of similar age, sex, and race/ethnicity (SMR=2.47; 95% CI, 2.31–2.65) (table 2). The risk of dying within the first month after discharge was much higher than expected (SMR=25.20; 95% CI, 16.13–39.38). As survival time increased after the first postdischarge month (3, 6, 12mo), the excess in mortality dropped markedly and continued to diminish to less than twice the expected risk (SMR=1.71; 95% CI, 1.57–1.86) for those surviving at least 1 year.

Table 2. SMRs⁎ by Postdischarge Period, Patient Characteristics, and Injury Severity

	Characteristic	Observed Deaths	Expected Deaths	SMR	95% CI
	Overall	2494	1008	2.47	2.31–2.65
	Postdischarge period
	0–1mo	504	20	25.20	16.13–39.38
	1–3mo	225	40	5.63	4.02–7.87
	3–6mo	209	59	3.54	2.66–4.73
	6–12mo	241	119	2.03	1.63–2.52
	After 12mo	1315	770	1.71	1.57–1.86
	Age at admission (y)
	<1	6	3	2.00	0.50–7.94
	1–19	41	11	3.73	1.92–7.24
	20–39	135	35	3.86	2.66–5.58
	40–59	392	83	4.72	3.74–5.96
	60–79	777	299	2.60	2.30–2.93
	≥80	1143	576	1.98	1.84–2.14
	Sex
	Male	1345	509	2.64	2.39–2.92
	Female	1149	499	2.30	2.09–2.54
	Race/ethnicity
	White	1865	769	2.43	2.24–2.63
	Black	67	24	2.79	1.78–4.39
	Asian	16	6	2.67	1.06–6.69
	Native American	3	1	3.00	0.33–27.72
	Hispanic	179	72	2.49	1.90–3.25
	Injury severity (maximum AIS of the head region)
	1=Minor	113	52	2.17	1.59–2.97
	2=Moderate	470	262	1.79	1.55–2.08
	3=Serious	336	140	2.40	1.98–2.90
	4=Severe	1360	504	2.70	2.45–2.97
	5=Critical	161	29	5.55	3.78–8.15

⁎ Boldface data reflect SMRs that are significantly greater than 1.0.

Age-stratified, all-cause SMR was largest among those 40 to 59 years (SMR=4.72; 95% CI, 3.74–5.96) and decreased with successively older age groups. Whether stratified by sex, race/ethnicity, or injury severity, SMRs were similar in magnitude to the unstratified all-cause cohort SMR, with the exception of patients with critical injury (SMR=5.55; 95% CI, 3.78–8.15) (see table 2).

Causes of death and cause-specific standardized mortality ratio 

Deaths among the TBI cohort were significantly greater than expected for almost all specific causes of death examined (table 3). The data suggested a substantially elevated risk of death caused by seizures relative to the expected rate, but the estimated SMR of 15.0 is based on only a few observed events and only 1 expected death. There were notable elevations in death caused by mental/behavioral disorders (SMR=4.70; 95% CI, 3.30–6.70), which include dementia as well as death caused by nervous system diseases (SMR=3.33; 95% CI, 2.48–4.46), which include some types of seizures and Alzheimer disease. Of similar magnitude to the estimated all-cause SMR of 2.5 were deaths caused by sepsis, digestive system diseases, and stroke. Smaller but significant SMRs were estimated for deaths caused by circulatory system diseases, respiratory diseases, malignant neoplasms, and external causes, including suicide, which had 2.4 times the expected rate (see table 3).

Table 3. SMRs⁎ for ICD-10 Underlying Causes of Death

	Cause of Death	Observed Deaths	Expected Deaths	SMR	95% CI
	All causes	2494	1008	2.47	2.31–2.65
	Mental/behavioral disorders (F00–F79)	174	37	4.70	3.30–6.70
	Dementia (F00–F03)	95	30	3.17	2.10–4.77
	Nervous system diseases (G00-G99, R56.8)	193	58	3.33	2.48–4.46
	Alzheimer disease (G30.9)	106	37	2.87	1.97–4.16
	Seizures (G40-G41, R56.8)	15	1	15.00	1.98–113.54
	Sepsis (A40-A41)	30	10	3.00	1.47–6.14
	Digestive system diseases (K00-K93)	119	42	2.83	1.99–4.03
	Stroke (I64)	120	52	2.31	1.67–3.19
	Circulatory system diseases (I00-I99)	864	377	2.29	2.03–2.58
	External causes (V01–Y98)	174	75	2.32	1.77–3.04
	Unintentional injury (V01–X59)	113	51	2.22	1.59–3.08
	Assault (X85.0–Y09.9)	10	4	2.50	0.78–7.97
	Suicide (X60–X84)	38	16	2.38	1.33–4.26
	Respiratory system diseases (J00–J99)	242	118	2.05	1.65–2.55
	Pneumonia (J12–J18)	91	28	3.25	2.13–4.96
	Aspiration pneumonia (J69)	37	11	3.36	1.72–6.59
	Malignant neoplasms (C00–D48)	259	184	1.41	1.17–1.70

⁎ Boldface data reflect SMRs that are significantly greater than 1.0.

For the 504 deaths occurring within the first postdischarge month, supplementary analyses identified circulatory system diseases as the most prevalent underlying cause of death category (n=182; 36%). After this was the category for “unknown” underlying cause of death (n=105; 21%), among which 86% were accompanied by TBI diagnosis codes on the death certificate. Additionally, early deaths were associated with a higher prevalence of severe/serious TBIs (77% vs 42%; P<.001) and a lower prevalence of minor/moderate TBIs (11% vs 41%; P<.001) compared with later deaths.

Demographics 

Older age was a consistent predictor of death for all of the adult age groups (see table 5). Increasing age was the only significant demographic predictor of death among those 4 to 19 years old. Among the other specific age groups, younger adults (20–39y) were also more likely to die if they lived in a lower income neighborhood. The same was true for those 40 to 59 years old, for whom metropolitan residence and hospitalization payment source also predicted mortality, with government payor (versus private health insurance) associated with greater risk of death. For older adults (age 60–79y) and elderly adults (age 80y and older), the significant demographic risk factors for death were older age, male sex, and metropolitan residence (see table 5).

Injury-related/acute care measures 

For all age groups except the oldest (80y and older), injury severity was the strongest injury-related predictor of mortality (see table 5). Critical injuries (vs minor/moderate) conferred significant elevations in mortality. For all adult age groups, external cause of injury predicted mortality, with falls (versus vehicular causes) conferring significant elevations in risk. Discharge destination predicted mortality for the oldest 3 age groups (ie, 40y and older), with destinations other than home without follow-up care associated with greater risk of mortality.

Comorbidities 

The presence of comorbid health conditions at the time of injury predicted death among adults over 20 years old (see table 5), with 3 or more health conditions (vs none) associated with a significantly higher risk of death. Number of associated injuries was generally not predictive of mortality. However, the direction of the parameter estimates suggests that the greatest risk of death is associated with having no associated injuries (ie, no injuries other than TBI), notably among older and elderly adults. Supplemental chi-square analysis demonstrated that among those with no associated injuries, there was a higher proportion of injuries with fall-related TBI and a lower prevalence of vehicular-related TBIs.

Age effects 

Although not formally tested, modification of effects by age were demonstrated across the risk factor modeling strata. Among the 3 oldest age groups, the strength of the association between death and government payor demonstrated an overall decrease with older age and was significant only for the youngest of the 3 age groups (age 40–59y). A more striking pattern over the age strata was that the importance of comorbid conditions on the risk of death decreased with increasing age, notably for those with 3 or more comorbid health conditions, versus none.