Development and Validation of a Functionally Relevant Comorbid Health Index in Adults Admitted to Inpatient Rehabilitation for Traumatic Brain Injury

Abstract

Several studies have characterized comorbidities among individuals with traumatic brain injury (TBI); however, there are few validated TBI comorbidity indices. Widely used indices (e.g., Elixhauser Comorbidity Index [ECI]) were developed in other patient populations and anchor to mortality or healthcare utilization, not functioning, and notably exclude conditions known to co-occur with TBI. The objectives of this study were to develop and validate a functionally relevant TBI comorbidity index (Fx-TBI-CI) and to compare prognostication of the Fx-TBI-CI with the ECI. We used data from the eRehabData database to divide the sample randomly into a training sample (N = 21,292) and an internal validation sample (N = 9166). We used data from the TBI Model Systems National Database as an external validation sample (N = 1925). We used least absolute shrinkage and selection operator (LASSO) regression to narrow the list of functionally relevant conditions from 39 to 12. In internal validation, the Fx-TBI-CI explained 14.1% incremental variance over an age and sex model predicting the Functional Independence Measure (FIM) Motor subscale at inpatient rehabilitation discharge, compared with 2.4% explained by the ECI. In external validation, the Fx-TBI-CI explained 4.9% incremental variance over age and sex and 3.8% over age, sex, and Glasgow Coma Scale score,compared with 2.1% and 1.6% incremental variance, respectively, explained by the ECI. An unweighted Sum Condition Score including the same conditions as the Fx-TBI-CI conferred similar prognostication. Although the Fx-TBI-CI had only modest incremental variance over demographics and injury severity in predicting functioning in external validation, the Fx-TBI-CI outperformed the ECI in predicting post-TBI function.

Introduction

Each year in the United States (US), more than 2.5 million adults sustain a traumatic brain injury (TBI), and 21,400 receive inpatient rehabilitation (IPR) for TBI.^1,2 Injury-related symptoms, such as cognitive impairment, behavioral change, and mood disorders are well-documented in this population.³ There is growing awareness that TBI-related symptoms interact with pre-existing and new-onset health conditions, further complicating recovery.^4,5 In a study of long-term TBI survivors, individuals commonly reported a wide range of conditions, including depression, anxiety, substance use disorders, sleep disorders, hypertension, fractures, back pain, and osteoarthritis.⁴

Functional impairment during IPR is a well-known predictor of poor long-term outcome.⁶ Current prognostic models predicting TBI functional outcome have largely relied on demographic and injury severity measures alone, leaving considerable unexplained variance in TBI outcomes.⁷ Recent TBI studies have provided evidence that select health conditions are associated with poorer functioning over time.^8,9 Previous studies of post-TBI medical comorbidity, however, have been limited by incomplete documentation of health conditions relevant to TBI, reliance on self-reporting, and under-representation of older adults who often have the greatest disease burden.^10,11

Several well-validated indices of disease burden exist (e.g., Charlson Comorbidity Index, Elixhauser Comorbidity Index [ECI]), but they were developed in different clinical populations and are designed to predict risk of death or healthcare utilization.^12
–14 Other studies of health and aging have used counts of conditions to index disease comorbidity, but these are limited by the tenuous assumption that all conditions (i.e., epilepsy and arthritis) have equal impact on outcomes.¹⁵ These widely used indices are therefore of limited utility for studying the role of disease comorbidity on functional outcomes after TBI.¹¹

One previous study evaluated utility of a functional comorbidity index developed in a general medical population¹⁶ to predict function in a non-TBI IPR population, but found it was not predictive of IPR discharge functioning or community discharge destination.¹⁷ In addition, a recent TBI Model Systems (TBIMS) National Database study aimed to develop one or more comorbid health indices in a sample of adults admitted to IPR for TBI, but ultimately did not recommend a singular index based on their validation analyses.¹⁸

There remains a need to develop a TBI comorbid health index with greater relevance to IPR functional outcome compared with currently available indices. To this end, the objectives of the present study were: (1) to develop and validate a TBI functionally relevant comorbid index, and (2) to compare its prognostication to the widely used ECI.¹⁴ We hypothesized the functionally relevant TBI comorbid index (Fx-TBI-CI) would be superior.

Methods

Data sources

The primary training and internal validation sample used in the present study came from the American Medical Rehabilitation Providers Association's subscription database, eRehabData. This database was created in 2002 as a standardized platform for participating IPR providers to fill out the Inpatient Rehabilitation Facility–Patient Assessment Instrument (IRF-PAI). The database is used for insurance payment reimbursement, inpatient rehabilitation facility-level and national benchmarking, and research among inpatient rehabilitation populations.

Our analytic sample was restricted to adults 16+ at IPR admission for whom TBI was the primary reason for rehabilitation. Cases of TBI were ascertained by the impairment group code of 2.21 or 2.22, and International Classification of Diseases (ICD) diagnoses codes for TBI from the Centers for Disease Control and Prevention case definition,¹⁹ which was the same criteria used by a previous TBI study using the eRehabData database.² Individuals who died during IPR were not included because of our primary interest in predicting function.

As an external validation sample, we used data from the TBIMS National Database. This multi-center longitudinal prospective cohort study enrolls participants who are 16+ years old at time of injury, received their acute care within 72 h of injury, and received IPR at a designated TBIMS facility. Eligible TBIMS participants sustained a moderate-to-severe TBI defined by post-traumatic amnesia >24 h, intracranial neuroimaging abnormalities on computed tomography, loss of consciousness >30 min, or Glasgow Coma Scale (GCS) score in the Emergency Department <13.

Measures

Comorbid conditions

We classified the presence of comorbidities using ICD diagnoses codes according to Clinical Classification Software (CCS) produced by the Healthcare Cost and Utilization Project (HCUP), a federally sponsored program by the Agency for Healthcare Research and Quality.²⁰ The CCS contains a uniform and standardized categorization schema for mutually exclusive case definitions that may be employed for administrative claims-based research.²¹ In the IRF-PAI, providers may list up to 25 ICD diagnoses codes for comorbid conditions present during inpatient rehabilitation.

Our list of selected comorbidities included 29 conditions included in the ECI¹⁴ and an additional 10 conditions we hypothesized a priori would be relevant to TBI functional outcome based on a review of previous literature.^{22

–32} One condition, “Epilepsy and seizure disorder,” was subsumed in the ECI category “Other neurological disorders”; however, because of its high incidence after TBI,³⁰ we classified this condition into its own independent category. The full list of 39 conditions we considered and the corresponding ICD codes are provided in Supplementary Table S1.

When we compared prevalence of each condition in the eRehabData database using ICD-9 and ICD-10 diagnosis codes, we determined there were non-trivial differences in prevalence rates of the same conditions using ICD-9 versus ICD-10 (Supplementary Table S2), a trend that has been documented previously in the literature.^33
–35 Therefore, we restricted the analysis to only persons with ICD-10 codes (IPR admissions between 2016–2020) in the eRehabData database (n = 30,458) and the external TBIMS National Database sample (n = 1925).

Primary outcome: functional outcome at inpatient rehabilitation discharge

Our primary outcome was the Functional Independence Measure^TM (FIM) Motor score at IPR discharge. The FIM Motor subscale measures functional performance on a 1 (total assist) to 7 (complete independence) scale for 13 modalities of motor function, where higher overall scores indicate better outcome. The FIM Motor is both functionally relevant and an important indicator of IPR outcome.³⁶ The FIM Cognitive subscale is also relevant to long-term functioning³⁷; however, given the centrality of the skills measured by FIM Motor to independence in activities of daily living (e.g., eating, grooming, toileting) and post-rehabilitation discharge care needs, we focused this initial study on development of comorbidity score anchored to FIM Motor. We used the Rasch-transformed FIM Motor score because of its linear scaling properties.³⁸

Covariates

Demographic and selected clinical characteristics were available in the eRehabData database as a part of the IRF-PAI and were used to characterize the sample and compare the training and validation samples. Only age (considered continuously) and sex were used as covariates in internal validation models. The TBIMS National Database contains additional data on TBI severity not available in eRehabData, thereby facilitating adjustment for brain injury severity in the external validation. We used the GCS total score³⁹ as an index of brain injury severity.

Statistical analysis

Using the eRehabData database, we calculated prevalence rates for all 39 comorbid health conditions. Next, we randomly divided our eRehabData sample into two independent subsamples: training (70%) and internal validation (30%). In the training subsample (n = 21,292), we calibrated comorbidity weights of the Fx-TBI-Health Index. To calibrate the weights, we used a penalized regression methodology, known as least absolute shrinkage and selection operator (LASSO) regression.

The LASSO regression uses a shrinkage parameter, λ, to select and retain only the most important variables for prediction by minimizing the sum of squared errors with a bound on the sum of the absolute value of beta coefficients.⁴⁰ This facilitates the selection of key covariates for retention in the model by identifying and excluding variables that do not contribute to model fit, resulting in parsimonious models. Traditional regression-based models are susceptible to multi-collinearity and to overfitting model parameters narrowly to study data. LASSO regression has been used in a previous study in the general medical population to create a comorbid health index.⁴¹

We performed the LASSO regression in the training sample using the glmnet package in R program.⁴² In the LASSO model, we included all 39 comorbidities, sex, and age at admission as predictors of Rasch-transformed FIM Motor at discharge. The variables included in the Fx-TBI-CI were selected using LASSO in the training sample, and the Fx-TBI-CI was then validated in independent internal and external validation samples. Because higher FIM Motor scores indicate more favorable outcomes, we multiplied all coefficients by -1, so higher values of our Fx-TBI-Health Index would correspond to poorer FIM Motor scores. We rounded the LASSO parameters to the next highest positive or negative integer for ease of interpretation.

We also created a Sum Condition Score, which was an unweighted score based on the same conditions as the Fx-TBI-CI, but instead of weighting the index based on model-based parameters, 1 point was assigned to conditions negatively associated with functional outcome, and -1 points were assigned to conditions positively associated with functional outcome. We created the Sum Condition Score to quantify prognostication of the weighted index versus an unweighted sum.

In the validation samples, we directly compared prognostication of the (1) Fx-TBI-CI, (2) Sum Condition Score, and the (3) ECI using incremental variability explained (ΔR²) between Model 1 including age and sex only, and Model 1 plus each index added in separately. For the external validation, we also compared the incremental variance of each index added to Model 2, which included age, sex, and GCS. We documented the Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC), wherein lower values indicated better model fit.

We visually inspected the linearity of age and our outcome in both samples, but found no evidence that a non-linear age term was required in the models. To evaluate the performance of newly derived Fx-TBI-CI across different patient populations, we repeated the same validation analyses within sex and age (<65 vs. ≥65 years) subgroups. We chose the age cutoff of 65 because that has been used as a threshold for older adulthood.⁴³

Results

Characterization of eRehabData and TBIMS samples

We compared characteristics between the training and internal validation of the eRehabData sample using standardized mean differences (Table 1 ).⁴⁴ We also compared characteristics of the internal validation and external validation samples. The average age of the training sample was 65.1 years, and 61.9% were males; 78.4% were white, 62.3% had Medicare coverage during an average IPR stay of 15.3 days. There were no major differences in any characteristics between individuals in the training and internal validation samples. Individuals in the external validation sample, however, were on average 13.1 years younger, more often male, non-white, and more often had non-Medicare payor sources than the internal validation sample. The prevalence rates of 39 health conditions are presented in Table 2 in the eRehabData sample.

Table 1.

Characteristics of the Training, Internal, and External Validation Samples

	Training (n = 21,292)	Internal validation (n = 9166)	SMD (Training vs. internal validation)	External validation (n = 1925)	SMD (Internal vs. external validation)
Age, mean (SD)	65.1 (20.4)	65.1 (20.6)	0.001	52.0 (21.1)	0.628^*
Age above or below 65, n (%)
Age <65	8125 (38.2%)	3519 (38.4%)	-0.004	1,452 (67.8%)	-0.617
Age ≥65	13,167 (61.8%)	5647 (61.6%)	0.004	689 (32.2%)	0.617
Sex, male (%)	13,177 (61.9%)	5724 (62.5%)	0.012	1,373 (71.3%)	0.187^*
Race/ethnicity, n (%)
White	15,776 (78.4%)	6812 (78.4%)	0	1,237 (64.5%)	0.311^*
Black	2360 (11.7%)	1048 (12.1%)	0.012	313 (16.3%)	-0.121^*
Hispanic/Latino	1240 (6.2%)	545 (6.3%)	0.004	274 (14.3%)	-0.266^*
Other	740 (3.7%)	280 (3.2%)	0.027	95 (5.0%)	-0.091
Rehabilitation payor source, n (%)
Medicare	13,346 (62.7%)	5,711 (62.3%)	0.008	378 (19.7%)	0.961^*
Other	7946 (37.3%)	3,455 (37.7%)	-0.008	1,545 (80.3%)	-0.961^*
Inpatient rehabilitation length of stay, mean (SD)	15.3 (11.6)	15.5 (11.8)	0.019	24.3 (25.0)	-0.450^*
Discharge disposition after inpatient rehabilitation, n (%)
Home	15,149 (71.4%)	6588 (72.1%)	-0.016	1520 (79.2%)	-0.166^*
Skilled nursing facility	3252 (15.3%)	1,368 (15.0%)	0.008	271 (14.1%)	0.026
Acute hospital	2471 (11.6%)	1045 (11.4%)	0.006	48 (2.5%)	0.355^*
Other	353 (1.7%)	136 (1.5%)	0.016	80 (4.2%)	-0.163^*
Rasch-transformed FIM Motor at admission, mean (SD)	35.8 (14.9)	35.7 (14.9)	0.007	34.1 (16.2)	0.103^*
Rasch-transformed FIM Motor at discharge, mean (SD)	52.9 (14.5)	52.9 (14.4)	0	56.6 (14.5)	-0.256^*
Comorbid health conditions included in Fx-TBI-Health Index, n (%)
Paralysis	3955 (18.9%)	1711 (18.7%)	0.005	259 (12.0%)	0.187
Congestive heart failure	2264 (10.6%)	974 (10.6%)	0	103 (4.8%)	0.219
Fluid and electrolyte disorders	5038 (23.7%)	2189 (23.9%)	-0.005	790 (36.7%)	-0.281
Renal disease	2611 (12.3%)	1191 (13.0%)	-0.021	107 (5.0%)	0.282
Diabetes without chronic complications	2661 (12.5%)	1118 (12.2%)	0.009	157 (7.3%)	0.166
Weight loss	1249 (5.9%)	583 (6.4%)	-0.021	209 (9.7%)	-0.122
Alcohol abuse	2419 (11.4%)	1043 (11.4%)	0	319 (14.8%)	-0.100
Dysphagia	6810 (32.0%)	3040 (33.2%)	-0.026	311 (16.2%)	0.402
Other neurological disorders (excluding epilepsy/seizure)	6145 (28.9%)	2683 (29.3%)	-0.009	519 (24.1%)	0.118
Epilepsy and seizure disorder	3134 (14.7%)	1289 (14.1%)	0.017	270 (12.5%)	0.047
Pneumonia	839 (3.9%)	366 (4.0%)	-0.005	281 (13.0%)	-0.327
Headache/migraine	2993 (14.1%)	1259 (13.7%)	0.012	83 (3.9%)	0.351
Fx-TBI-CI score, mean (SD)	n/a^¥	4.4 (5.1)	n/a^¥	3.6 (4.6)	0.164

SMD, standardized mean difference; SD, standard deviation; FIM, Functional Independence Measure; Fx-TBI-CI, functionally relevant TBI comorbid index.

FX-TBI-CI not derived in the testing sample, as the weights were calibrated in this sample.

Indicates a SMD >0.1, considered a meaningful difference.

Table 2.

Prevalence of Health Conditions among Patients with Traumatic Brain Injury Admitted to Inpatient Rehabilitation in eRehabData Overall Sample (N = 30,458)

Comorbidities	Prevalence (%)^§
Hypertension	16138 (53.0%)
Swallowing disorder/dysphagia	9850 (32.3%)
Other neurological disorders (excluding epilepsy/seizure)	8828 (29.0%)
Fluid and electrolyte disorders	7227 (23.7%)
Deficiency anemias	6973 (22.9%)
Sleep wake disorders	6898 (22.7%)
Malaise and Fatigue	6087 (20.0%)
Depression	6095 (20.0%)
Paralysis	5666 (18.6%)
Esophageal disorders	5073 (16.7%)
Hypothyroidism	4642 (15.2%)
Diabetes with chronic complications	4364 (14.3%)
Epilepsy and seizure disorder	4423 (14.5%)
Headache/migraine	4252 (14.0%)
Chronic pulmonary disease	4146 (13.6%)
Renal disease	3802 (12.5%)
Diabetes without chronic complications	3779 (12.4%)
Alcohol abuse	3462 (11.4%)
Congestive heart failure	3238 (10.6%)
Nervous system pain and pain syndromes	3086 (10.1%)
Hypotension	2403 (7.9%)
Obesity	2151 (7.1%)
Weight loss	1832 (6.0%)
Peripheral vascular disease	1699 (5.6%)
Valvular disease	1609 (5.3%)
Coagulation deficiency	1469 (4.8%)
Psychoses	1418 (4.7%)
Pneumonia (except that cause by tuberculosis)	1205 (4.0%)
Drug abuse	1073 (3.5%)
Liver disease	925 (3.0%)
Rheumatoid arthritis/collagen vascular diseases	843 (2.8%)
Solid tumor without metastasis	580 (1.9%)
Pulmonary circulation disorders	326 (1.1%)
Pulmonary embolism	282 (0.93%)
Blood loss anemia	244 (0.8%)
Metastatic cancer	220 (0.7%)
Chronic peptic ulcer disease	206 (0.7%)
Lymphoma	176 (0.6%)
HIV/AIDS	69 (0.2%)

HIV, human immunodeficiency virus; AIDS, acquired immunodeficiency syndrome.

Prevalence rates are ordered greatest to smallest

Calibration of Fx-TBI-CI weights

The results of the model conducted in the training sample (N = 21,292) are presented in Table 3. Twenty-seven of the 39 conditions were shrunk to zero in the LASSO model. Of the 12 retained conditions, we assigned point-based weight assignments from the LASSO coefficients. We provided a comparison of the overlapping and unique health conditions and corresponding weights between the three indices in Table 4. We showed the distribution of the Rasch-transformed FIM Motor outcome and three comorbidity indices in Supplementary Figure S1a–d.

Table 3.

LASSO-Selected^┼ Comorbid Health Conditions Negatively Associated with Discharge FIM Motor in Training Sample (N = 21,292)

Variable	LASSO^┼ Adjusted β	Point-based weight assignment^€ (rounded LASSO Adjusted [-β])
Swallowing disorder/dysphagia	-7.43	8
Paralysis	-4.78	5
Other neurological disorders (excluding epilepsy/seizure)	-1.78	2
Epilepsy and seizure disorder	-0.83	1
Congestive heart failure	-0.69	1
Fluid and electrolyte disorders	-0.51	1
Pneumonia	-0.39	1
Renal disease	-0.35	1
Weight loss	-0.32	1
Diabetes without chronic complications	-0.16	1
Headache/migraine	2.78	-3
Alcohol abuse	1.83	-2

LASSO, least absolute shrinkage and selection operator.

┼

The LASSO model was adjusted for age and sex; however, sex shrunk to 0 in the LASSO. The following comorbid health conditions were included in the initial model selection, and beta coefficients were shrunk to 0 during LASSO regression: HIV/AIDS, deficiency anemias, rheumatoid arthritis/collagen vascular diseases, blood loss anemia, chronic pulmonary disease, coagulation deficiency, depression, diabetes with chronic complications, drug abuse, hypertension, hypothyroidism, liver disease, lymphoma, metastatic cancer, obesity, peripheral vascular disease¸ psychoses, pulmonary circulation disorders¸ solid tumor without metastasis, chronic peptic ulcer disease, valvular disease¸ malaise and fatigue, esophageal disorders, nervous system pain and pain syndromes, pulmonary embolism, sleep wake disorders, hypotension.

€

The point-based weight assignment was derived by rounding –β up to the next highest positive or negative integer.

Table 4.

Comparison of Conditions and Weights Across the Three Health Indices

	Conditions	Fx-TBI-CI weight	Sum Condition Score weight	Elixhauser Comorbidity Index weight^¥
Overlapping conditions	Paralysis	5	1	5
	Congestive heart failure	1	1	9
	Fluid and electrolyte disorders	1	1	11
	Renal disease	3	1	6
	Diabetes without chronic complications	1	1	0
	Weight loss	1	1	9
	Alcohol abuse	-2	-1	-1
Unique to Fx-TBI-Health Index/Sum Condition Score	Dysphagia	8	1
	Other neurological disorders (excluding epilepsy/seizure)	1	1
	Epilepsy and seizure disorder^┼	1	1
	Pneumonia	1	1
	Headache/migraine	-3	-1
Unique to Elixhauser Comorbidity Index	Other neurological disorders (with epilepsy/seizure)			5
	HIV/AIDS			0
	Deficiency anemia			-2
	Rheumatic disease			0
	Blood loss anemia			-3
	Chronic pulmonary disease			3
	Coagulation deficiency			11
	Depression			-5
	Diabetes with complications			-3
	Drug abuse			-7
	Hypertension			-1
	Hypothyroidism			0
	Liver disease			4
	Lymphoma			6
	Metastatic cancer			14
	Obesity			-5
	Peripheral vascular disease			3
	Psychoses			-5
	Pulmonary circulatory disorders			6
	Solid tumor without metastasis			7
	Chronic peptic ulcer disease			0
	Valvular disease			0

Fx-TBI-CI, functionally relevant TBI comorbid index; HIV, human immunodeficiency virus; AIDS, acquired immunodeficiency syndrome.

┼

Epilepsy and seizure disorder codes were included in the Elixhauser Comorbidity Index, but were subsumed into the other neurological disorders category.

In-hospital mortality Elixhauser Comorbidity Index weights were used.

Internal validation of Fx-TBI-CI

In the internal validation (N = 9166), we compared prognostication Fx-TBI-CI, Sum Conditions Score, and ECI scores in Table 5. We determined that Model 1 (age and sex only) explained 2.9% variance in Rasch-transformed FIM Motor scores. The Fx-TBI-CI explained 14.1% incremental variance over Model 1, compared with 10.4% with the Sum Conditions Score, and 2.4% with the ECI. The model with the Fx-TBI-CI had the lowest AIC and BIC values, suggesting superior model fit. The subgroup analyses by sex and age are presented in Supplementary Tables S3 and S4, respectively. Similar results were documented in subgroup analyses among males and females. The Fx-TBI-CI had greater variance explained among adults <65 (17.6%) than adults ≥65 (11.5%). Similarly, the Sum Conditions Score and ECI had greater prognostication among adults <65 than adults ≥65 years old.

Table 5.

Internal (eRehabData) and External (Traumatic Brain Injury Model Systems) Validation of the Health Indices Predicting Functional Independence Measure Motor at Rehabilitation Discharge

		Model 1^§	Model 2^¥	Model 1^§ + Fx-TBI-CI	Model 2^¥ + Fx-TBI-CI	Model 1^§ + Sum Conditions Score	Model 2^¥ + Sum Conditions Score	Model 1^§ + Elixhauser Comorbidity Index	Model 2^¥ + Elixhauser Comorbidity Index
Full internal validation sample (n = 9166)	R²	0.029	—	0.170	—	0.133	—	0.052	—
	Δ R² from Model 1	—	—	0.141	—	0.104	—	0.024	—
	AIC^┼	74704	—	73278	—	73670	—	74488	—
	BIC^┼	74733	—	73313	—	73705	—	74523	—
Full external validation sample with non-missing covariate information (n = 1925)	R²	0.076	0.108	0.125	0.146	0.128	0.149	0.097	0.124
	Δ R² from Model 1	—	0.032	0.049	—	0.052	—	0.021	—
	Δ R² from Model 2	—	—	—	0.038	—	0.041	—	0.016
	AIC	15623	15559	15522	15476	15514	15471	15582	15526
	BIC	15646	15587	15549	15509	15542	15504	15609	15559

Fx-TBI-CI, functionally relevant TBI comorbid index; GCS, Glasgow Coma Scale; AIC, Akaike Information Criterion; BIC, Bayesian Information Criterion.

Model 1 adjusted for age (continuous) and sex only; in the external sample, Model 1 was conducted among persons with non-missing GCS data to make Δ R² from all models comparable.

Model 2 adjusted for age (continuous), sex, and injury severity (GCS); Model 2 was only conducted in the external sample, because GCS data was not available in the eRehabdata database.

┼

Lower AIC and BIC indicate better model fit; higher R² indicates better model greater variability explained.

External validation of Fx-TBI-CI

We used the TBIMS National Database for external validation (Table 5 ). We determined that Model 1 (age and sex only) explained 7.6% total variance in Rasch-transformed FIM Motor scores. Model 2 (age, sex, and GCS) accounted for 10.8% total variance. Overall, the Fx-TBI-CI explained 4.9% incremental variance over Model 1, and 3.8% incremental variance over Model 2. The Sum Conditions Score explained 5.2% and 4.1% incremental variance over Model 1 and Model 2, respectively. The ECI explained 2.1% and 1.6% incremental variance over Models 1 and 2, respectively. Subgroup analyses in the external sample indicated largely similar results among males and females, and adults age ≥65 and age <65.

Discussion

We used a national sample of adults admitted to IPR for TBI to develop the Fx-TBI-CI. Existing comorbid health indices are used widely in clinical research, but few were developed and validated in TBI populations with anchors to functional outcome. We determined the Fx-TBI-CI had superior prognostication relative to one widely used comorbidity index, the ECI; however, the absolute incremental variance of the Fx-TBI-CI over demographic and TBI severity models was modest on external validation.

The associations of individual health comorbidities on post-TBI recovery are well documented.^4,8,9,18 Another recent effort¹⁸ to develop a comorbid index using a traditional regression-based approach with the TBIMS National Database did not result in recommendation for future use because of a drop in the incremental variance explained in their validation sample, similar to what we found in the Fx-TBI-CI. They did, however, offer several suggestions for future studies in this area, including further external validation, using only ICD-10 codes, and comparing a weighted versus unweighted comorbid index.

Our index performed relatively well in the internal eRehabData validation subsample (14.1% incremental variance above age and sex). The prognostic performance of the index, however, decreased significantly when applied to an external sample, the TBIMS National Database (4.9% and 3.8% incremental variance over demographics and demographics and TBI severity, respectively).

Given the marked decrease in variance explained in the external sample, exploration of contributors to this difference may be highly informative. There are systematic differences between eRehabData, a national IPR claims database, and the TBIMS National Database, a prospective observational cohort study.^1,2 One key difference is the age distribution between samples; the eRehabData database heavily skews toward older adults (62% of the sample were over age 65), while only 32% of individuals were over 65 years in the TBIMS sample, and the average age was 52. Accordingly, the prevalence rates of selected chronic diseases included in the Fx-TBI-CI were disparate between the eRehabData and TBIMS external validation samples. For example, the rate of congestive heart failure and renal disease were 10.6% and 13.0% in the eRehabData sample, respectively, compared with 4.8% and 5.0% in the TBIMS sample, respectively.

Our subgroup analyses in the internal validation sample indicated strong model performance in both age groups, although the Fx-TBI-CI had better prognostication among adults <65. In the external validation sample, the Fx-TBI-CI had largely similar prognostication among age subgroups (4.7% incremental variance above demographics and injury severity among adults ≥65 and 4.1% among adults <65).

Other national TBI cohort studies, like the Chronic Effects of Neurotrauma Consortium⁴⁵ and the Transforming Research and Clinical Knowledge in Traumatic Brain Injury^46,47 have similar average age to the TBIMS National Database. As such, future studies investigating the impact of comorbidity indices would benefit from further comparing prognostication of indices (the Fx-TBI-CI or newly developed indices) among older versus younger adults. This detail is particularly important to reconcile because of the rising trend of incident TBI among older adults,⁴⁸ a demographic subgroup with higher comorbidity burden⁴ in which standard injury severity-based prognostic models have performed poorly.¹⁰

We compared the prognostic performance of a weighted and unweighted index. We found in the internal validation that the Fx-TBI-CI (incremental variance = 14.1%) outperformed the Sum Conditions Score (incremental variance = 10.4%) in prognostication, and both scores outperformed the ECI (incremental variance = 2.4%). In the external validation, the Sum Conditions Score and Fx-TBI-CI had largely similar prognostication, and both provided consistently greater prognostication than the ECI. In totality, our results do support the utility of an empirically informed count of selected conditions that are found relevant to discharge TBI functional outcome; however, an index with greater prognostication across multiple samples would be desirable.

Of note, we chose not to adjust for admission motor functioning in our models. The rationale for this decision was that the eRehabData database does not contain time stamps on comorbidities, therefore, prohibiting us from distinguishing among comorbidities that were pre-existing conditions, co-occurring conditions, or comorbidities that first occurred during rehabilitation. All comorbidities occurring before rehabilitation admission (presumably the majority) logically would have influenced motor function at the point of rehabilitation admission. Therefore, adjusting for rehabilitation admission motor function in the models would have removed the variance in discharge function accounted for by admission function, and by doing so, it would have masked all the shared variance accounted for by all the comorbidities occurring before rehabilitation admission in the Fx-TBI-CI.

With that said, future studies with an ability to definitively distinguish between pre-existing and co-occurring conditions may benefit from quantifying the mediating (e.g., intermediary) role of admission motor functioning on the relationship between pre-existing comorbidities and discharge motor function.

Our list of candidate conditions were decided a priori in the design phase of this study, and we narrowed the list based on an empirically driven methodology using LASSO. This did result in certain conditions (e.g., headache, alcohol) with a seemingly counterintuitive positive association with outcome. Past work developing comorbidity scores in TBI populations, however, similarly included selected comorbidity with positive associations to outcome.¹⁸ We do not suggest presence of these conditions are necessarily protective; instead, there may be unmeasured or residual confounding accounting for these positive associations.

There are limitations to the current study. We chose the outcome of physical function (FIM Motor); it is unknown whether our index is relevant to functional cognition. Duration of follow-up is limited to inpatient rehabilitation in the eRehabData database; however, this limitation is offset by a very large sample size, little missing data, and systematic data collection across inpatient rehabilitation facilities.

Using ICD-10 codes, we were not able to ascertain whether conditions were present before TBI or diagnosed after TBI, and do not have information on severity, chronicity, or treatment of conditions. Administrative claims are in place primarily for billing, not research, purposes; thus it is possible that codes tied to payment may be prioritized. The eRehabData database only collects up to 25 comorbidity diagnosis codes as part of IRF-PAI. Roughly 12.1% of the eRehabData sample had as many as 25 codes, and it is possible among this subsample we underestimated their comorbidities from diagnoses not documented in the database. There was no limit on number of possible ICD-10 codes in the TBIMS National Database.

The timing of collection of ICD-10 codes between the eRehabData (internal) and TBIMS National Database (external) differed; the former were coded during IPR, and the latter were coded at the time of acute hospitalization discharge. Although all participants in the eRehabData sample had TBI impairment codes (e.g., TBI was the reason necessitating admission to rehabilitation), we cannot confirm that rehabilitation care received was TBI-specific, as would be observed in the TBIMS centers. The aforementioned factors may have contributed to observed differences in index prognostication across samples.

We adjusted for GCS score as a measure of injury severity in the TBIMS National Database, and persons who were intubated or on paralytics had missing data; therefore, the external validation findings may not generalize to this group. It is possible that some individuals had extracerebral injuries in addition to TBI. Therefore, to empirically evaluate potential confounding, we conducted post hoc analyses adding Injury Severity Scale (ISS) score as a covariate to our models using a published algorithm⁴⁹ translating ICD-10 codes to an ISS score (data not shown). This had a very marginal difference on the ΔR² (≤1%) of the three comorbidity scores; therefore, it did not appear that our results were confounded to a large degree by overall injury burden, after adjustment for age, sex, and GCS. In addition, all individuals in the internal and external validation received inpatient rehabilitation, and results of this study may not generalize to persons with TBI not receiving inpatient rehabilitation.

Strengths of the present study included large, national internal and external samples to develop and validate our indices. We reduced potential measurement bias by using only ICD-10 codes (not mixed with ICD-9) and calculated incremental variance compared with both demographics and TBI severity.

In the present study, we used a contemporary regression shrinkage methodology to narrow a list of 39 conditions to 12 that are specifically relevant to functional motor outcome at IPR discharge after TBI. We created a weighted Fx-TBI-CI that added significant incremental variance in an internal validation sample. The Fx-TBI-CI, however, had only modest incremental value in an external sample, although it did outperform the widely used ECI. Because our index had only modest prognostication in the external sample, the generalizability of this index to broad, external TBI samples requires further refinement. We recommend future studies to seek better prognostication and further evaluate prognostication of indices among older versus younger adults.

The relatively stronger prognostication of Fx-TBI-CI in older adults suggests that TBI samples that focus on enrollment of older adults with TBI (e.g., TRACK-GERI) may benefit from developing a health index score separately in older adults. Future work would also benefit from exploring development of indices in TBI datasets with available patient-reported outcomes related to cognitive, emotional, and behavioral outcomes.

Footnotes

Funding Information

Dr. Kumar was supported on this project through funding from the Brain Injury Association of America's Brain Injury Research Fund through their Seed Grant Award for Young Investigators. Dr. Dams-O'Connor and Dr. Kumar's effort were also support in part by grants from NIDLRR to the Icahn School of Medicine at Mount Sinai (90DP0038 and 90DPTB0009). Dr. Mazumdar receives grant funding paid to her institution for grants unrelated to this work from NCI (P30CA196521, CA220491, U24CA224319-01, DK124165), NCATS (TR002997), and NIA (AG028741, AG066605, P30AG028741, R01AG054540). Dr. Whiteneck was partially supported by grants from MINDSOURCE, Colorado Department of Human Services, State of Colorado (Contract Number: IHEA 101422) and NIDILRR (grant number 90DP0084) to Craig Hospital. Dr. Hammond was also partially supported by grants from NIDILRR to Indiana University School of Medicine (grant number 90DRTB0002 and 90DPHF0006-01-00). The contents of this publication do not necessarily represent the policy of BIAA, MINDSOURCE, NIDILRR, ACL, or HHS, and you should not assume endorsement by BIAA or the Colorado or Federal Government.

Author Disclosure Statement

No competing financial interests exist.

Supplementary Material

Supplementary Table S1

Supplementary Table S2

Supplementary Table S3

Supplementary Table S4

References

Corrigan

J.D.

, Cuthbert

J.P.

, Whiteneck

G.G.

, Dijkers

M.P.

, Coronado

, Heinemann

A.W.

, Harrison-Felix

, and Graham

J.E.

(2012). Representativeness of the traumatic brain injury model systems national database. J. Head Trauma Rehabil. 27, 391–403.

Cuthbert

J.P.

, Corrigan

J.D.

, Whiteneck

G.G.

, Harrison-Felix

, Graham

J.E.

, Bell

J.M.

, and Coronado

V.G.

(2012). Extension of the representativeness of the Traumatic Brain Injury Model Systems National Database: 2001 to 2010. J. Head Trauma Rehabil. 27, E15–E27.

Masel

B.E.

, and DeWitt

D.S.

(2010). Traumatic brain injury: a disease process, not an event. J. Neurotrauma, 27, 1529–1540.

Hammond

F.M.

, Corrigan

J.D.

, Ketchum

J.M.

, Malec

J.F.

, Dams-O'Connor

, Hart

, Novack

T.A.

, Bogner

, Dahdah

M.N.

, and Whiteneck

G.G.

(2019). Prevalence of medical and psychiatric comorbidities following traumatic brain injury. J. Head Trauma Rehabil. 34, E1–E10.

Kumar

R.G.

, Juengst

S.B.

, Wang

, Dams-O'Connor

, Dikmen

S.S.

, O'Neil-Pirozzi

T.M.

, Dahdah

M.N.

, Hammond

F.M.

, Felix

E.R.

, Arenth

P.M.

, and Wagner

AK.

(2018). Epidemiology of comorbid conditions among adults 50 years and older with traumatic brain injury. J. Head Trauma Rehabil. 33, 15–24.

Dams-O'Connor

, Pretz

, Billah

, Hammond

F.M.

, and Harrison-Felix

(2015). Global outcome trajectories after TBI among survivors and nonsurvivors: a National Institute on Disability and Rehabilitation Research Traumatic Brain Injury Model Systems study. J. Head Trauma Rehabil. 30, E1–E10.

Maas

A.I.

, Lingsma

H.F.

, and Roozenbeek

(2015). Predicting outcome after traumatic brain injury. Handb Clin Neurol, 128, 455–474.

Kumar

R.G.

, Olsen

, Juengst

S.B.

, Dams-O'Connor

, O'Neil-Pirozzi

T.M.

, Hammond

F.M.

, and Wagner

A.K.

(2019). Comorbid conditions among adults 50 years and older with traumatic brain injury: examining associations with demographics, healthcare utilization, institutionalization, and 1-year outcomes. J. Head Trauma Rehabil. 34, 224–232.

Malec

J.F.

, Ketchum

J.M.

, Hammond

F.M.

, Corrigan

J.D.

, Dams-O'Connor

, Hart

, Novack

, Dahdah

, Whiteneck

G.G.

, and Bogner

(2019). Longitudinal effects of medical comorbidities on functional outcome and life satisfaction after traumatic brain injury: an individual growth curve analysis of NIDILRR traumatic brain injury model system data. J. Head Trauma Rehabil. 34, E24–E35.

10.

Peters

M.E.

, and Gardner

R.C.

(2018). Traumatic brain injury in older adults: do we need a different approach?. Concussion 3, CNC56.

11.

Thompson

H.J.

, McCormick

W.C.

, and Kagan

S.H.

(2006). Traumatic brain injury in older adults: epidemiology, outcomes, and future implications. J. Am. Geriatr. Soc. 54, 1590–1595.

12.

Charlson

M.E.

, Pompei

, Ales

K.L.

, and MacKenzie

C.R.

(1987). A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J. Chron. Dis. 40, 373–383.

13.

Deyo

R.A.

, Cherkin

D.C.

, and Ciol

M.A.

(1992). Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. J. Clin. Epidemiol. 45, 613–619.

14.

Elixhauser

, Steiner

, Harris

D.R.

, and Coffey

R.M.

(1998). Comorbidity measures for use with administrative data. Med. Care, 36, 8–27.

15.

Piazza

J.R.

, Charles

S.T.

, and Almeida

D.M.

(2007). Living with chronic health conditions: age differences in affective well-being. J. Gerontol. B. Psychol. Sci. Soc. Sci. 62, P313–P321.

16.

Groll

D.L.

, To

, Bombardier

, and Wright

J.G.

(2005). The development of a comorbidity index with physical function as the outcome. J. Clin. Epidemiol. 58, 595–602.

17.

Kumar

, Graham

J.E.

, Resnik

, Karmarkar

A.M.

, Tan

, Deutsch

, and Ottenbacher

K.J.

(2016). Comparing comorbidity indices to predict post-acute rehabilitation outcomes in older adults. Am. J. Phys. Med. Rehabil. 95, 889–898.

18.

Whiteneck

, Ketchum

, Almeida

E.J.

, Goldstein

, Brown

A.W.

, Corrigan

, Hammond

F.M.

, Weintraub

, and Tefertiller

(2021). Developing an index of medical conditions associated with outcomes after moderate to severe traumatic brain injury. J. Neurotrauma, 38, 593–603.

19.

Marr

A.L.

, and Coronado

V.G.

(2004). Central nervous system injury surveillance data submission standards—2002. Atlanta, GA: US Department of Health and Human Services, CDC.

20.

Steiner

, Elixhauser

, and Schnaier

(2002). The healthcare cost and utilization project: an overview. Eff. Clin Pract. 5, 143–151.

21.

Healthcare Cost and Utilization Project. (2017). HCUP Clinical Classifications Software (CCS). Agency for Healthcare Research and Quality, Rockville,MD. Available at: www.hcup-us.ahrq.gov/toolssoftware/ccs/ccs.jsp. Last accessed November 14, 2021 .

22.

Berry

, Ley

E.J.

, Bukur

, Malinoski

, Margulies

D.R.

, Mirocha

, and Salim

(2012). Redefining hypotension in traumatic brain injury. Injury, 43, 1833–1837.

23.

Cantor

J.B.

, Ashman

, Gordon

, Ginsberg

, Engmann

, Egan

, Spielman

, Dijkers

, and Flanagan

(2008). Fatigue after traumatic brain injury and its impact on participation and quality of life. J. Head Trauma Rehabil. 23, 41–51.

24.

Kumar

R.G.

, Kesinger

M.R.

, Juengst

S.B.

, Brooks

M.M.

, Fabio

, Dams-O'Connor

, Pugh

M.J.

, Sperry

J.L.

, and Wagner

A.K.

(2020). Effects of hospital-acquired pneumonia on long-term recovery and hospital resource utilization following moderate to severe traumatic brain injury. J. Trauma Acute Care Surg. 88, 491–500.

25.

Lucas

, Hoffman

J.M.

, Bell

K.R.

, Walker

, and Dikmen

(2012). Characterization of headache after traumatic brain injury. Cephalalgia, 32, 600–606.

26.

Mackay

L.E.

, Morgan

A.S.

, and Bernstein

B.A.

(1999). Swallowing disorders in severe brain injury: risk factors affecting return to oral intake. Arch. Phys. Med. Rehabil. 80, 365–371.

27.

Mathias

, and Alvaro

(2012). Prevalence of sleep disturbances, disorders, and problems following traumatic brain injury: a meta-analysis. Sleep Med. 13, 898–905.

28.

Morgan

A.S.

, and Mackay

L.E.

(1999). Causes and complications associated with swallowing disorders in traumatic brain injury. J. Head Trauma Rehabil. 14, 454-461.

29.

Nampiaparampil

D.E.

(2008). Prevalence of chronic pain after traumatic brain injury: a systematic review. JAMA, 300, 711–719.

30.

Ritter

A.C.

, Wagner

A.K.

, Fabio

, Pugh

M.J.

, Walker

W.C.

, Szaflarski

J.P.

, Zafonte

R.D.

, Brown

A.W.

, Hammond

F.M.

, Bushnik

, Johnson-Greene

, Shea

, Krellman

J.W.

, Rosenthal

J.A.

, and Dreer

L.E.

(2016). Incidence and risk factors of posttraumatic seizures following traumatic brain injury: A Traumatic Brain Injury Model Systems Study. Epilepsia, 57, 1968–1977.

31.

Seifi

, Dengler

, Martinez

, and Godoy

D.A.

(2018). Pulmonary embolism in severe traumatic brain injury. J. Clin. Neurosci. 57, 46–50.

32.

Vane

D.W.

, Shiffler

, Grosfeld

J.L.

, Hall

, Angelides

, Weber

T.R.

, and Fitzgerald

J.F.

(1982). Reduced lower esophageal sphincter (LES) pressure after acute and chronic brain injury. J. Pediatr. Surg. 17, 960–964.

33.

Elixhauser

, Heslin

K.C.

, and Owens

P.L.

(2017). Healthcare Cost and Utilization Project (HCUP) Recommendations for Reporting Trends Using ICD-9-CM and ICD-10-CM/PCS Data. US Agency for Healthcare Research and Quality. Available at: www.hcup-us.ahrq.gov/datainnovations/icd10_resources.jsp. Last accessed November 14, 2021 .

34.

Hamedani

A.G.

, Blank

, Thibault

D.P.

, and Willis

A.W.

(2021). Impact of ICD-9 to ICD-10 coding transition on prevalence trends in neurology. Neurology: Clinical Practice, 11, e612–e619.

35.

Salemi

J.L.

, Tanner

J.P.

, Kirby

R.S.

, and Cragan

J.D.

(2019). The impact of the ICD-9-CM to ICD-10-CM transition on the prevalence of birth defects among infant hospitalizations in the United States. Birth Defects Res. 111, 1365–1379.

36.

Corrigan

J.D.

, Smith-Knapp

, and Granger

C.V.

(1997). Validity of the functional independence measure for persons with traumatic brain injury. Arch. Phys. Med. Rehabil. 78, 828–834.

37.

Zarshenas

, Colantonio

, Horn

S.D.

, Jaglal

, and Cullen

(2019). Cognitive and motor recovery and predictors of long-term outcome in patients with traumatic brain injury. Arch. Phys. Med. Rehabil. 100, 1274–1282.

38.

Granger

C.V.

, Deutsch

, and Linn

R.T.

(1998). Rasch analysis of the Functional Independence Measure (FIM) mastery test. Arch. Phys. Med. Rehabil. 79, 52–57.

39.

Teasdale

, and Jennett

(1974). Assessment of coma and impaired consciousness: a practical scale. Lancet, 2, 81–84.

40.

Tibshirani

(1996). Regression shrinkage and selection via the lasso. J. Royal Statistical Soc.: Series B (Methodological) 58, 267–288.

41.

Mukherjee

, Ou

H.T.

, Wang

, and Erickson

S.R.

(2011). A new comorbidity index: the health-related quality of life comorbidity index. J. Clin. Epidemiol. 64, 309–319.

42.

R Core Team. (2017). R: A Language and Environment for Statistical Computing. 887 R Foundation for Statistical Computing, Vienna, Austria.

43.

Harvey

L.A.

, and Close

J.C.

(2012). Traumatic brain injury in older adults: characteristics, causes and consequences. Injury, 43, 1821–1826.

44.

Flury

B.K.

, and Riedwyl

(1986). Standard distance in univariate and multivariate analysis. American Statistician, 40, 249–251.

45.

Cifu

D.X.

, Williams

, Hinds

S.R.

, and Agyemang

A.A.

(2018). Chronic effects of neurotrauma consortium: a combined comparative analysis of six studies. Introduction to Special edition of Brain Injury. Brain Inj. 32, 1149–1155.

46.

Agtarap

S.D.

, Campbell-Sills

, Jain

, Sun

, Dikmen

, Levin

, McCrea

M.A.

, Mukherjee

, Nelson

L.D.

, Temkin

, Yuh

E.L.

, Giacino

J.T.

, Manley

G.T.

, Stein

M.B.

, and TRACK-TBI

Investigtors.

(2021). Satisfaction with life after mild traumatic brain injury: a TRACK-TBI study. J. Neurotrauma, 38, 546–554.

47.

Stein

M.B.

, Jain

, Giacino

J.T.

, Levin

, Dikmen

, Nelson

L.D.

, Vassar

M.J.

, Okonkwo

D.O.

, Diaz-Arrastia

, Robertson

C.S.

, Mukherjee

, McCrea

, Mac Donald

C.L.

, Yue

J.K.

, Yuh

, Sun

, Campbell-Sills

, Temkin

, Manley

G.T.

, and TRACK-TBI

Investigators.

(2019). Risk of posttraumatic stress disorder and major depression in civilian patients after mild traumatic brain injury: a TRACK-TBI study. JAMA Psychiatry, 76, 249–258.

48.

Dams-O'Connor

, Cuthbert

J.P.

, Whyte

, Corrigan

J.D.

, Faul

, and Harrison-Felix

(2013). Traumatic brain injury among older adults at level I and II trauma centers. J. Neurotrauma, 30, 2001–2013.

49.

Clark

D.E.

, Black

A.W.

, Skavdahl

D.H.

, and Hallagan

L.D.

(2018). Open-access programs for injury categorization using ICD-9 or ICD-10. Inj. Epidemiol. 5, 11.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.02 MB

0.01 MB

0.12 MB