Multi-dimensional pre-injury health and 6-month outcome after traumatic brain injury in older adults presenting with mainly GCS 13–15

Abstract

Objective:

To test the accuracy of existing traumatic brain injury (TBI) outcome prediction models in older adults with TBI (oaTBI), to determine the independent association of five different measures of pre-injury health with 6-month outcome, and to investigate whether these pre-injury health measures improve outcome prediction.

Setting:

Two Level-1 Trauma Centers.

Participants:

112 adults aged ≥65 years who underwent head computed tomography after acute TBI (95% with Glasgow Coma Scale (GCS) score 13–15, 66% with acute intracranial trauma on head CT.

Design:

Prospective cohort study.

Main Measures:

Pre-injury health was characterized using Activities of Daily Living (ADL), Functional Activities Questionnaire (FAQ), Clinical Dementia Rating (CDR), Charlson Comorbidity Index (CCI), and Groningen Frailty Index (GFI). The primary outcome was poor recovery (Glasgow Outcome Scale–Extended <5: severe disability or death) at 6 months. We compared performance of CRASH and IMPACT models in predicting 6-month outcome. Each pre-injury health measure was tested for independent association with outcome and incremental prognostic value when added to the best-performing model. Model accuracy was assessed using area under the curve (AUC).

Results:

Participants’ mean±SD age was 77.8±7.9 years; 42% were women. The CRASH-basic model had the highest accuracy (AUC = 0.74, 95% CI = 0.64–0.83). Poor 6-month outcome was independently associated with pre-injury cognitive impairment (CDR OR = 1.38, 95% CI = 1.07–1.76), impaired ADLs (ADL OR = 2.45, 95% CI = 1.16–5.15), decreased cognitively mediated function (FAQ OR = 1.17, 95% CI = 1.06–1.29), and higher frailty (GFI OR = 1.30, 95% CI = 1.06–1.59), but not with comorbidities (CCI OR 1.00, 95% CI 0.76–1.32). No measure significantly improved CRASH-basic AUC (all p > 0.05).

Conclusions:

Among older adults with mainly mild TBI, pre-injury frailty and cognitive-mediated function were independently associated with poor 6-month outcome, yet these pre-injury measures did not significantly improve prediction accuracy. Instruments such as the FAQ and GFI capture clinically meaningful pre-injury heterogeneity and may inform the development of more precise oaTBI prognostic models.

Introduction

Traumatic brain injury (TBI) is a leading cause of morbidity and mortality, with over 1.5 million sustained every year in the United States (US) alone.¹ Older adult TBI (oaTBI), defined as TBI occurring at age 65 years or older, accounts for 46% of all TBI-related hospitalizations in the US.¹ oaTBI patients usually have worse outcomes and slower recovery compared to younger patients,² however, this is not always the case.^2,3 The field lacks reliable tools to predict functional recovery following oaTBI: available prediction models, comprised mainly of metrics of injury severity (e.g., Glasgow Coma Scale [GCS], head computed tomography [CT] results),⁴ tend to over- or under-estimate functional recovery in oaTBI,^2,5 leaving providers and families with inadequate evidence to guide decision-making and potentially leading to excessive nihilism or optimism.⁶ This uncertainty extends to mild oaTBI cases (GCS 13–15) who have more heterogeneous functional recoveries⁷ and higher mortality than younger adults with mild TBI.⁸ Pre-injury health status may provide useful information that can refine prediction of recovery following oaTBI.⁹

A comprehensive geriatric assessment is a multidimensional diagnostic process guiding personalized care for older people, ideally including assessment of function and independent capabilities, mobility, cognition, and medical comorbidities.¹⁰ Hospital-based older-adult initiatives and emerging guidelines¹¹ have incorporated brief targeted assessments to guide care. Borrowing from the fields of geriatric medicine and dementia, many existing brief measures could prove useful in the context of oaTBI: questionnaires assessing independent mobility and activities of daily living;¹² burden of comorbidities,¹³ cognitive status,¹⁴ and perhaps the most holistic measure, frailty status.¹⁵ Frailty describes available reserve to withstand acute stressors¹⁶ and may be of particular value as an oaTBI functional recovery predictor.¹⁷

Two extensively validated TBI outcome prediction models, CRASH¹⁸ (Corticosteroid Randomization After Significant Head Injury Cohort) and IMPACT⁴ (International Mission for Prognosis and Analysis of Clinical Trials), use acute injury features and routine emergency department (ED) blood tests to predict the odds of achieving unfavorable 6-month outcome, defined as recovery to severe disability or worse on the Glasgow Outcome Scale (GOS<4) or GOS-Extended (GOSE<5), where the lowest score means death, and the highest corresponds with complete return to pre-injury functional status.¹⁹ CRASH and IMPACT outcome prediction models, developed on mostly younger patient cohorts, may not optimally capture injury severity in oaTBI,²⁰ and do not capture heterogeneity in pre-injury health. Few prior studies have reported on pre-injury health measures as predictors of functional outcomes in oaTBI and all have excluded patients with pre-existing cognitive impairment or a high burden of comorbidities, thus masking the real-world heterogeneity in pre-injury health and limiting generalizability.¹⁷ This study focused primarily on capturing a specific dimension of previously unmeasured pre-injury heterogeneity in oaTBI - namely overall function, cognitive status, and frailty. We hypothesized that adding pre-injury characteristics to existing TBI functional outcome prediction models would improve performance for predicting severe disability or death (GOSE<5), moderate disability or worse (GOSE<7), or less than complete recovery (GOSE<8) in oaTBI.

To test this hypothesis, we leveraged data from the TRACK-GERI prospective cohort study of older adults presenting to two Level 1 Trauma Centers with acute TBI. Notably, rather than excluding participants with pre-injury cognitive impairment/multi-morbidity, TRACK-GERI used measures of pre-injury health to quantify this heterogeneity. The TRACK-GERI cohort thus serves as an ideal cohort in which to test head-to-head the added prognostic value of several different pre-injury health measures. The aims of this study were: (1) Determine the performance of CRASH and IMPACT outcome prediction models in prediction of six-month functional outcome in TRACK-GERI, a real-world, representative cohort of oaTBI; (2) Select the best-performing existing model; (3) Determine the independent association of six different pre-injury health measures with six-month functional outcome and estimate the added prognostic value of each measure to the best-performing existing model in TRACK-GERI.

Methods

Study design and participants

TRACK-GERI was a prospective observational cohort study conducted at two Level 1 trauma centers with a planned enrollment of N=270 older adults with acute TBI. Ethical approval for both sites was obtained from UCSF Institutional Review Board (#19–28925). Each participant with TBI was co-enrolled with a study partner informant. The first N=30 participant-informant dyads were enrolled into a single center pilot study (02/2018–04/2019) to establish feasibility of the protocol and anticipated rate of enrollment; the remaining N=240 dyads were subsequently enrolled into a two-center study (10/2020–09/2024). Inclusion criteria were: (1) age≥65 years; (2) English-speaking (or Chinese-speaking during the final 6-months of recruitment at the UCSF site); (3) arrived in the ED within 72 hours of TBI; (4) had an acute head CT scan performed as part of clinical care. Exclusion criteria: participants in custody, under a medically evaluated psychiatric hold, an acute major event concurrently with the TBI (e.g., acute stroke, seizure or cardiac arrest). All participants and study partners provided an informed consent. For individuals unable to consent on their own, the process was pursued through a surrogate.

Measurements

Baseline demographics (e.g., age, marital status) were collected. Injury mechanism and severity metrics were recorded: ED arrival Glasgow Coma Scale (GCS), pupil reactivity, hypoxia or hypotension, blood glucose and hemoglobin, mechanism of injury and ED disposition. The first clinical head CT was evaluated by a single neuroradiologist (co-author E.L.Y) for TBI neuro-imaging, coding elements needed for the calculating of the Marshall CT score classification: presence and size of mass or mixed density lesions, degree of compression of basal cisterns, midline shift, subarachnoid hemorrhage, and epidural hematoma.⁴ IMPACT and CRASH models’ scores were calculated. In TRACK-GERI, Pre-injury health measures were collected extensively: Functional Activities Questionnaire (FAQ; assessment of cognitively-mediated instrumental activities of daily living),¹² Charlson Comorbidity index (CCI; summary score of specific comorbidities),¹³ Clinical Dementia Rating (CDR) global and sum of boxes scores,²¹ and Groningen Frailty Indicator (GFI; multi-dimensional frailty scale).²² Independent performance of activities of daily living (ADL) was also recorded, see Supplementary eTable 1 for more details.

Six-month outcome was assessed using the GOSE reported by the participant or study partner informant, by in-person or phone interviews. If both completed the GOSE, then the lowest score was used. The GOSE assessment measured the return to pre-injury functional status. For example, a participant who required 16 hours of in-home help pre-injury and then fully recovered post-injury to the point of requiring precisely that same amount of in-home help as pre-injury, would be coded as fully independent inside the home environment. While TRACK-GERI collected both GOSE-TBI and GOSE-All, the present analysis used only GOSE-All to align with CENTER-TBI and the majority of published TRACK-TBI studies and because isolation of TBI-specific functional deficits from functional deficits resulting from concurrently sustained non-TBI trauma in this population was difficult.¹⁹ Poor outcome was defined as in the original CRASH and IMPACT models: GOSE<5 (severe disability or death).^4,18 Sensitivity analyses were conducted, additionally defining poor outcome as GOSE<7 (moderate disability or worse) and GOSE<8 (less than complete recovery).

We calculated predicted outcome using four existing TBI outcome prediction models: (1) CRASH-basic for high-income countries (predictors: age, ED arrival GCS, pupil reactivity and major extra cranial injury)¹⁸, (2) IMPACT-basic (predictors: age, ED arrival GCS motor score, pupil reactivity), (3) IMPACT-extended (predictors: IMPACT-basic plus hypoxia, hypotension, CT results), (4) IMPACT-lab (predictors: IMPACT-extended plus blood glucose and hemoglobin levels).⁴

Statistical analysis

The analytic cohort included participants with 6-month GOSE score plus all predictor variables necessary to estimate the four CRASH and IMPACT models. Included versus excluded participants did not significantly differ in age, CCI, CDR, FAQ, injury mechanism, ED GCS, or ED disposition (all p>0.05; see additional variables in Supplementary eTable 2)

Baseline demographic, injury, and pre-injury health measures were compared between GOSE>5 versus GOSE≥5 subgroups, using Mann-Whitney test, t-test and Chi-Square test. Due to small sample size, we were unable to estimate IMPACT and CRASH model coefficients for our cohort. Therefore, to avoid model overfitting and to avoid exceeding the recommended number of predictors per outcome event according to Harrell’s rule of 10,²³ we derived IMPACT and CRASH total scores based on published models’ coefficients, and used them as a single predictor of 6-month functional outcome. Model selection analysis, examining each score’s performance in discriminating GOSE<5, was estimated using univariable logistic regression models, and receiver operating characteristic (ROC) curves were used to derive the area under the ROC curve (AUC) with 95% confidence intervals (CI). AUCs <0.7 were considered poor, 0.7–0.8 fair, 0.8–0.9 good, and 0.9–1.0 excellent.²⁴ Calibration intercept and slope were calculated to quantify systematic under- or over-estimation. To contextualize our results within the existing published literature, we compared model performance in TRACK-GERI to model performance reported in a prior meta-analysis of N=12 studies of TBI in all-ages²⁵ using MedCalc, an online calculator for comparing independent AUC’s, based on the Hanley & McNeil method.²⁶

The best-performing model (i.e., highest AUC) was then used in model optimization analysis; multivariable logistic regression models that included the best-performing model’s coefficient score and each pre-injury measure to determine its contribution to overall outcome prediction (primary: GOSE<5; sensitivity: GOSE<7 and GOSE<8). Each model’s performance was reported using AUC and Nagelkerke pseudo R². The models AUC's before and after adding pre-injury health measures were compared using DeLong test.²⁷ Independent association of each pre-injury measure with poor outcome was reported using odds ratios (OR) with 95% CI. Inclusion criteria for these subsequent analyses were having all variables needed to estimate the selected model coefficient score plus all pre-injury measures (i.e., FAQ, ADL, CCI, CDR, and GFI; see Figure 1). Thus, with a final sample size of N=103 participants included in the model optimization analysis, our analysis had 80% power (α=0.05) to detect the following minimum odds ratios for the association of each pre-injury health measure with 6-month GOSE<5: FAQ 1.09, ADL 2.51, CCI 1.43, CDR Global 3.09, CDR sum of boxes 1.23, GFI 1.25 (calculated using G*Power).²⁸

Figure 1: Flow diagram showing cohort included in each analysis.

TRACK-GERI cohort recruited a total of 270 patients following oaTBI, of which 160 had completed the 6-month GOSE interviews. N=112 cases had sufficient data available to be included in the main analysis comparing performance of existing outcome prediction models. N=103 cases had sufficient data available to be included in the subsequent analysis of the independent association of pre-injury health variables with outcome and their added value on the best-performing existing outcome prediction model (CRASH-basic).

Abbreviations: Corticosteroid Randomization After Significant Head Injury Cohort (CRASH); International Mission for Prognosis and Analysis of Clinical Trials (IMPACT); Functional Activities Questionnaire (FAQ); Activities of daily living (ADL); Charlson Comorbidity Index (CCI); Clinical Dementia Rating (CDR); Groningen Frailty Index (GFI).

Results

A total of N=112 participants were included in the model selection analysis (Figure 1), following exclusion of cases missing 6-month GOSE or any of the variables required for employing CRASH-basic or one of the three IMPACT models. Demographic, pre-injury measures, and injury variables are shown in Table 1. By 6 months post-injury, N=40 (36%) had GOSE<5; N=59 (53%) had GOSE<7; N=92 (82%) had GOSE<8 (Supplementary eFigure 1). This completion rate of the main outcome (N=160/270, 59%) was comparable to that of cohorts with similar data collection burden, such as TRACK-TBI²⁹ and CENTER-TBI,³⁰ that excluded participants with pre-existing cognitive impairment or excessive disability.

Table 1.

Demographic, pre-injury, and injury characteristics stratified by GOSE<5

Characteristics are shown as N (%) or Mean±SD (range)	All (N=112)	GOSE≥5 (N=72)	GOSE<5 (N=40)	P-value1
Age	77.8±7.9 (65–99)	76.4±7.8 (65–99)	80.5±7.6 (65–97)	0.007
Women	47 (42%)	34 (47%)	13 (33%)	0.130
Race (N=1 missing)
White	92 (83%)	61 (86%)	31 (78%)	0.266
Black	4 (4%)	2 (3%)	2 (5%)
Other/Unknown	15 (13%)	8 (11%)	7 (17%)
Marital Status (N=2 missing)
Married or in partnership	57 (52%)	39 (55%)	18 (46%)	0.761
Widower	28 (25%)	14 (20%)	14 (36%)
Other2	25 (23%)	18 (25%)	7 (18%)
CCI (N=1 missing)	4.8± 1.8 (2–11)	4.6±1.9 (2–11)	5.2±1.4 (3–8)	0.055
GFI score (range 0–12), (N=2 missing), %Frail (GFI≥4)	4.4±2.8 (0–12), 56%	3.7±2.3 (0–11), 50%	5.7±3.3 (1–12), 68%	0.003
FAQ score (N=3 missing)	3.6±6.6 (0–30)	1.6±3.5 (0–18)	7.4±9.1 (0–30)	<0.001
ADL score (N=3 missing)	0.5±0.7 (0–2.7)	0.3±0.5 (0–2.1)	0.7±0.8 (0–2.7)	0.012
CDR Global (N=15 missing)
Normal (0)	45 (46%)	34 (56%)	11 (31%)	0.002
MCI (0.5)	39 (40%)	24 (39%)	15 (42%)
Dementia (≥1)	13 (14%)	3 (5%)	10 (27%)
CDR Sum of Boxes (N=20 missing)	1.7±3.2 (0–18)	0.9±1.6 (0–8)	3.0±4.5 (0–18)	0.004
Injury Mechanism
Fall, Ground level	60 (54%)	38 (53%)	22 (55%)	0.665
Fall, Height	32 (29%)	19 (26%)	13 (33%)
MVA	15 (13%)	13 (18%)	2 (5%)
Other	5 (4%)	2 (3%)	3 (7%)
Major extra cranial injury (% positive) (N=84)3	51 (61%)	29 (57%)	22 (67%)	0.270
Emergency Department Disposition
Discharged from ED	21 (19%)	16 (22%)	5 (13%)	<0.001
Admitted to Ward	54 (48%)	43 (60%)	11 (27%)
Admitted to ICU	37 (33%)	13 (18%)	24 (60%)
GCS score
15	72 (64%)	54 (75%)	18 (45%)	<0.001
14	29 (26%)	14 (19%)	15 (38%)
13	6 (5%)	4 (6%)	2 (5%)
≤12	5 (5%)	0 (0%)	5 (12%)
GCS motor score
1	1 (1%)	0 (0%)	1 (3%)	0.001
2	0 (0%)	0 (0%)	0 (0%)
3	0 (0%)	0 (0%)	0 (0%)
4	2 (2%)	0 (0%)	2 (5%)
5	5 (5%)	1 (1%)	4 (10%)
6	104 (92%)	71 (99%)	33 (82%)
Emergency department pupillary reaction
Equal and reactive	110 (98%)	72 (100%)	38 (95%)	0.057
Unequal/Unreactive	2 (2%)	0 (0%)	2 (5%)
Hypoxia (% positive)	1 (1%)	0 (0%)	1 (3%)	0.178
Hypotension (% positive)	5 (5%)	4 (6%)	1 (3%)	0.453
CT findings (% positive)	74 (66%)	44 (61%)	30 (75%)	0.137
Marshall CT classification
I	37 (33%)	27 (38%)	10 (25%)	0.007
II	64 (57%)	44 (61%)	20 (50%)
III	1 (1%)	0 (0%)	1 (3%)
IV	0 (0%)	0 (0%)	0 (0%)
V	2 (2%)	0 (0%)	2 (5%)
VI	8 (7%)	1 (1%)	7 (17%)
Epidural hematoma (% positive)	3 (3%)	2 (3%)	1 (3%)	0.930
Subarachnoid hemorrhage (% positive)	46 (41%)	25 (35%)	21 (53%)	0.067
Subdural hemorrhage (% positive)	45 (40%)	21 (29%)	24 (60%)	0.001
Glucose (mg/dl)	71.9±23.6 (29.4–219.8)	71.4±25.1 (29.4–219.8)	72.7±20.9 (45.5–134.3)	0.691
Hemoglobin (g/dl)	8.2±1.1 (5.3–11.1)	8.1±1.2 (5.3–11.1)	8.3±0.9 (6.0–9.9)	0.439

Notes:

¹Mann-Whitney test (age, race, marital status, CCI, GFI, FAQ, ADL, CDR total and CDR sum of boxes, injury mechanism, abbreviated injury scale, emergency department disposition, GCS and Marshall CT score), T-test (hemoglobin measure), and Chi-Square test (for gender, hypoxia, hypotension, pupillary reaction, positive CT, sub-arachnoid hemorrhage, epidural hemorrhage, sub-dural hemorrhage).

²Single/Divorced

³Abbreviated Injury Scale≥2 for extra cranial injury, s/p extra cranial surgery or fractures.

Abbreviations: Charlson Comorbidity Index (CCI); Groningen Frailty Index (GFI); Functional Activities Questionnaire (FAQ); Activities of daily living (ADL); Clinical Dementia Rating (CDR); Motor Vehicle Accident (MVA); Emergency Department (ED); Intensive Cara Unit (ICU); Glasgow Coma Scale (GCS); Computerized Tomography (CT).

The performance of each of the four existing models for predicting 6-month GOSE<5 in oaTBI ranged from poor to fair (AUCs 0.68–0.74) and is shown in Table 2 and Figure 2a. Overall, CRASH-basic had the highest AUC (95% CI) at 0.74 (0.64–0.83), considered “fair”. Compared to performance reported in a meta-analysis of N=12 prior studies of the same four outcome prediction models in all-ages TBI (reported AUCs 0.80–0.82),²⁵ both CRASH-basic and IMPACT-extended had comparable performance in our cohort as reported in previous studies, while IMPACT-core and IMPACT-lab had significantly worse performance (p=0.035–0.039, see Table 2).

Table 2.

Performance of four existing models for predicting 6-month poor outcome (GOSE<5) in older adults with TBI versus all-ages TBI

	oaTBI (TRACK-GERI)					All-Ages TBI (meta-analysis)
N=112	Sensitivity (N correctly predicted poor outcome/N observed)	Specificity (N correctly predicted good outcome/N observed)	Calibration Intercept (95% CI)	Calibration Slope (95% CI)	AUC (95% CI)	AUC (95% CI)1	Comparison2
CRASH-basic	40.0% (16/40)	87.5% (63/72)	0.022 (−0.26, 0.30)	0.97 (0.55,1.38)	0.74 (0.64,0.83)	0.82 (0.78,0.86)	p=0.137
IMPACT-core	32.5% (13/40)	87.5% (63/72)	0.004 (−0.35,0.35)	0.99 (0.47,1.52)	0.69 (0.59,0.79)	0.80 (0.77,0.83)	p=0.041
IMPACT-extended	32.5% (13/40)	84.7% (61/72)	0.022 (−0.28,0.32)	0.97 (0.52,1.42)	0.72 (0.63,0.82)	0.82 (0.78,0.85)	p=0.086
IMPACT-lab	22.5% (9/40)	90.3% (65/72)	0.014 (−0.36,0.39)	0.98 (0.41,1.55)	0.68 (0.58,0.78)	0.80 (0.75,0.85)	p=0.039

Notes.

¹Zarei et al; Arch Acad Emerg Med. (2023)

²MedCalc, using Hanley & McNeil method

Abbreviations: Older Adults Traumatic Brain Injury (oaTBI); Area Under Curve (AUC); Corticosteroid Randomization After Significant Head Injury Cohort (CRASH); International Mission for Prognosis and Analysis of Clinical Trials (IMPACT).

Figure 2. Receiver operating characteristic (ROC) curves for each model predicting GOSE<5 in older adults with traumatic brain injury.

3a – ROC curves of the four models, predicting GOSE<5; 3b – ROC curves CRASH basic/CRASH basic with added FAQ; 3c – ROC curves CRASH basic/CRASH basic with added ADL; 3d – ROC curves CRASH basic/CRASH basic with added CCI; 3e – ROC curves CRASH basic/CRASH basic with added CDR global score; 3f – ROC curves CRASH basic/CRASH basic with added CDR sum of boxes; 3g – ROC curves CRASH basic/CRASH basic with added GFI.

Abbreviations: Receiver Operating Characteristic (ROC); Corticosteroid Randomization After Significant Head Injury Cohort (CRASH); International Mission for Prognosis and Analysis of Clinical Trials (IMPACT); Functional Activities Questionnaire (FAQ); Activities of daily living (ADL); Charlson Comorbidity Index (CCI); Clinical Dementia Rating (CDR); Groningen Frailty Index (GFI).

Model optimization analysis of the added value of pre-injury measures to the best-performing model (CRASH-basic) for predicting 6-month poor outcome (primary: GOSE<5; sensitivity: GOSE<7 and GOSE<8) is shown in Table 3 and Figures 2b–g. Several AUC point-estimates improved slightly, particularly with the addition of FAQ for prediction of GOSE<5 (improved from AUC=0.76 to 0.82, fair to good, DeLong p=0.07) and addition of GFI for prediction of GOSE<7 or GOSE<8 (improved from AUC=0.62 to 0.71 and AUC=0.70 to 0.79, DeLong p=0.86 and p=0.60, respectively), however no improvement was statistically significant (all DeLong p>0.05 for difference in AUC). When added to CRASH-basic, measures FAQ, ADL, CDR (global and sum of boxes) and GFI were significantly and independently associated with greater odds of GOSE<5; FAQ and GFI were significantly associated with greater odds of GOSE<7, and only GFI was found to be significantly associated with greater odds of GOSE<8 (see Table 3).

Table 3.

Added value of each pre-injury health measure for predicting poor outcome (GOSE<5, GOSE<7, and GOSE<8) in older adults with TBI

N=103	Sensitivity (N correctly predicted poor outcome/N observed poor outcome)	Specificity (N correctly predicted good outcome/N observed good outcome)	Calib. Intercept (95% CI)	Calib. Slope (95% CI)	Model’s AUC (95% CI)	Model’s Nagelkerke R squared	Odds Ratio (95% CI)2	P-value2
Predicting GOSE<5; outcome events N=33 (32%)
CRASH basic	42.4% (14/33)	90.0% (63/70)	0.015 (−0.26,0.29)	0.978 (0.59,1.36)	0.76 (0.66,0.85)	0.27	-	-
CRASH basic + FAQ1	51.5% (17/33)	90.0% (63/70)	0.011 (−0.20,0.22)	0.984 (0.69,1.28)	0.82 (0.73,0.91)	0.40	1.17 (1.06,1.29)	0.002
CRASH basic + ADL1	48.5% (16/33)	91.4% (64/70)	0.011 (−0.23,0.25)	0.983 (0.65,1.32)	0.76 (0.69,0.88)	0.34	2.45 (1.16,5.15)	0.018
CRASH basic + CCI1	42.4% (14/33)	90.0% (63/70)	0.015 (−0.26,0.29)	0.978 (0.59,1.36)	0.76 (0.66,0.85)	0.27	1.00 (0.76,1.32)	0.990
CRASH basic + CDR global1	48.5% (16/33)	91.4% (64/70)	0.019 (−0.21,0.25)	0.972 (0.65,1.29)	0.80 (0.71,0.89)	0.36	5.29 (1.38,20.29)	0.015
CRASH basic + CDR sum of boxes1	45.5% (15/33)	92.9% (65/70)	0.016 (−0.21,0.24)	0.977 (0.66,1.29)	0.80 (0.71,0.89)	0.37	1.38 (1.07,1.76)	0.012
CRASH basic + GFI1	48.5% (16/33)	94.3% (66/70)	−0.011 (−0.24,0.22)	1.016 (0.69,1.34)	0.79 (0.69,0.89)	0.35	1.30 (1.06,1.59)	0.012
Predicting GOSE<7; outcome events N=52 (50%)
CRASH basic	53.8% (28/52)	58.8% (30/51)	0.012 (−0.41,0.43)	0.977 (0.17,1.79)	0.62 (0.51,0.72)	0.08	-	-
CRASH basic + FAQ1,3	51.9% (27/52)	72.5% (37/51)	0.015 (−0.30,0.33)	0.971 (0.37,1.57)	0.65 (0.55,0.76)	0.14	1.09 (1.0,1.19)	0.045
CRASH basic + ADL1,3	48.1% (25/52)	64.7% (33/51)	0.017 (−0.32,0.36)	0.966 (0.32,1.61)	0.65 (0.54,0.75)	0.12	1.81 (0.92,3.55)	0.085
CRASH basic + CCI1	53.8% (28/52)	58.8% (30/51)	0.012 (−0.41,0.43)	0.976 (0.17,1.78)	0.61 (0.51,0.72)	0.08	0.99 (0.79,1.26)	0.962
CRASH basic + CDR global1,3	57.7% (30/52)	62.7% (32/51)	0.017 (−0.32,0.36)	0.967 (0.32,1.61)	0.64 (0.53,0.74)	0.12	2.41 (0.86,6.72)	0.094
CRASH basic + CDR sum of boxes1,3	53.8% (28/52)	64.7% (33/51)	0.014 (−0.31,0.34)	0.972 (0.37,1.58)	0.65 (0.55,0.76)	0.13	1.23 (0.99,1.54)	0.067
CRASH basic + GFI1	57.7% (30/52)	68.6% (35/51)	0.005 (−0.25,0.26)	0.990 (0.51,1.47)	0.71 (0.61,0.81)	0.18	1.33 (1.10,1.61)	0.003
Predicting GOSE<8; outcome events N=84 (82%)
CRASH basic4	100% (84/84)	0% (0/19)	0.044 (−0.52,0.60)	0.946 (0.27,1.63)	0.70 (0.58,0.82)	0.13	-	-
CRASH basic + FAQ1,4	100% (84/84)	0% (0/19)	0.015 (−0.51,0.54)	0.982 (0.35,1.61)	0.71 (0.59,0.83)	0.15	1.08 (0.94,1.24)	0.295
CRASH basic + ADL1,4	100% (84/84)	0% (0/19)	0.035 (−0.49,0.56)	0.957 (0.32,1.59)	0.72 (0.60,0.83)	0.15	1.67 (0.62,4.50)	0.307
CRASH basic + CCI1,4	100% (84/84)	0% (0/19)	0.035 (−0.52,0.59)	0.957 (0.28,1.63)	0.70 (0.58,0.82)	0.13	1.03 (0.74,1.42)	0.873
CRASH basic + CDR global1,4	100% (84/84)	0% (0/19)	0.047 (−0.51,0.60)	0.942 (0.27,1.62)	0.70 (0.59,0.82)	0.13	1.41 (0.34,5.81)	0.639
CRASH basic + CDR sum of boxes1,4	100% (84/84)	0% (0/19)	0.042 (−0.51,0.60)	0.949 (0.27,1.62)	0.70 (0.58,0.82)	0.13	1.06 (0.81,1.39)	0.678
CRASH basic + GFI1	95.2% (80/84)	15.8% (3/19)	0.008 (−0.63,0.38)	0.991 (0.54,1.44)	0.79 (0.68,0.90)	0.25	1.47 (1.10,1.97)	0.010

Notes:

¹Range of preinjury variables: FAQ 0–30; ADL 0–2.7; CCI 2–11; CDR global 0–3; CDR sum of boxes 0–18; GFI 0–12

²Significance level and Odds Ratio of pre-injury variable in the model

³CRASH basic was statistically insignificant in the regression model

⁴Although CRASH basic was a statistically significant predictor, the model has a specificity of 0% (predicted all cases as GOSE<8)

Abbreviations: Area Under Curve (AUC); Functional Activities Questionnaire (FAQ); Activities of daily living (ADL); Charlson Comorbidity Index (CCI); Clinical Dementia Rating (CDR); Groningen Frailty Index (GFI).

Discussion

In this study, examined outcome prediction models demonstrate poor to fair accuracy in a real-world cohort of older adults with acute TBI, with CRASH-basic model, comprised of age, GCS, pupil reactivity, major extracranial injury, and country income, performing best with AUC of 0.74. We quantified pre-injury health and functional heterogeneity and examined their independent associations with 6-month outcomes, assessing their added value to existing oaTBI prediction models: cognitively mediated function (FAQ) and frailty (GFI) were independently associated with 6-month GOSE after adjustment for CRASH-basic, demonstrating the feasibility of capturing clinically meaningful pre-injury health in oaTBI. However, adding these measures did not significantly improve the CRASH-basic model. We hypothesize that this finding may be the result of limited power due to small sample size and suboptimal characterization of injury severity in predominantly GCS 13–15 cases. Future larger studies should evaluate oaTBI-specific outcome prediction models that incorporate early post-TBI sequelae.

Four studies have reported validation of CRASH or IMPACT models for prediction of 6-month GOSE<5 in oaTBI.^31–34 Of these, Staples et al^31,32 focused on mild oaTBI (GCS 13–15) with no exclusions for pre-injury health: IMPACT lab model calibration was overly optimistic in old vs. young patients, but had good overall performance (AUC=0.84, CI 0.77–0.91),³¹ while the CRASH-CT model had both good calibration and performance (AUC=0.83, CI 0.79–0.88).³² The other two studies excluded individuals with pre-existing cognitive impairment or multi-morbidity: Yue et al³³ reported that in oaTBI with GCS 3–14, CRASH-CT was superior to IMPACT lab (c-statistic 0.88, CI 0.80–0.93, and 0.62 CI 0.43–0.79, respectively). Roe et al³⁴ showed that in severe oaTBI (GCS<9), CRASH models show inappropriately pessimistic predictions (e.g., predicted 85% unfavorable outcome vs. 44% observed, AUC scores not presented). Thus, prior studies on the performance of existing prediction models in real-world all-severity oaTBI are limited.

CRASH and IMPACT were developed primarily in younger patients with more severe TBI (GCS <14), making them a priori suboptimal for cohorts dominated by GCS 13–15. Although newer models for mild TBI are emerging,^17,35 these rely on measures not available in TRACK-GERI and could not be applied here, thus we evaluated and adapted CRASH/IMPACT. Notably, these models do not capture pre-injury heterogeneity beyond age and rely on injury severity markers (e.g., pupillary reactivity) with limited discriminatory value in mild TBI, constraining their ability to distinguish recovery trajectories in this cohort.

In our study, two pre-injury health measures were consistently associated with oaTBI 6-month outcome: (1) GFI, a multidimensional frailty measure;²² and (2) FAQ, a measure of cognitively-mediated instrumental ADLs.¹² Notably, the CCI, an index of medical comorbidities that is a well-established predictor of mortality in older adults,¹³ showed no added value: CCI neither improved the model’s performance nor was associated with 6-month outcome recovery with the three thresholds tested. These contrasting results highlight a critical concept: measures like the GFI or FAQ capture elements of pre-injury health that are distinct from pre-existing diagnosed comorbidities, able to meaningfully quantify aging heterogeneity in the context of oaTBI. Some studies have attempted to capture frailty in the context of oaTBI outcome using a simple “multi-morbidity” approach, using recorded diagnostic codes or self-reported medical conditions.^9,36 This approach will not capture the five core elements of frailty originally characterized by Fried et al:³⁷ unintentional weight loss, exhaustion, weakness, slow walking, and low physical activity. Indeed, simple tabulation of diagnosed diseases will also miss most cases of mild cognitive impairment (MCI), a severely under-diagnosed condition in the general population,³⁸ found to be extremely common in the TRACK-GERI cohort (i.e., 40% of participants).

Measures like the GFI and FAQ capture the end-result of comorbidity in a way that is agnostic to specific diagnosis. Using feasible assessments and surveys like GFI and FAQ offer a standardized method to assess pre-injury status, beyond subjective bed-side observations. Frailty is broadly conceptualized as increased vulnerability or having lower functional reserve needed to endure acute stressors.¹⁶ Substantial evidence demonstrates frailty is an important predictor of outcome following acute events,^39,40 beyond age and injury severity.⁴¹ A meta-analysis of 12 studies examined frailty's value in predicting TBI outcome,⁹ however included studies were based on a wide age range,³⁶ only severe⁴² or only mild TBI cases,¹⁷ narrow inclusion criteria (e.g., cases requiring craniotomy), or report only short-term outcome.^43–45

At least two prior studies have reported on the association of frailty, measured using the GFI, with outcome after mild oaTBI (GCS 13–15): Abdulle et al reported that higher GFI score was significantly associated with increased odds of incomplete recovery (GOSE<8) 1–3 years post-injury (OR=1.87, 95% CI 1.51–2.31).¹⁷ Thuss et al reported GFI was significantly associated with reduced odds of complete recovery (GOSE=8) 3-months post-injury (OR=0.76, 95% CI 0.63–0.92), but not at 6-months post-injury (OR-0.91, 95% CI 0.77–1.08), speculating that this inconsistency could be attributed to the lingering effect of the injury on perceived quality of life or post-traumatic complaints.⁷ Considering that frailty detected as early as mid-life has been associated with important general health outcomes in large-scale population-based studies,⁴⁶ future studies of any-severity TBI in both older and middle-aged adults should consider incorporating a measure of frailty as a means to resolve an aspect of otherwise un-measured pre-injury heterogeneity.

With respect to the FAQ, our findings suggest that cognitively-mediated function assessed using this a brief survey should also be considered in the context of oaTBI recovery outcomes, and it may be preferred over the more extended semi-structured 30-minute CDR interview for most studies of oaTBI that do not require a definitive cognitive diagnosis category (normal, MCI, dementia). The CDR (both CDR global and CDR sum-of-boxes) was only significantly associated with 6-month GOSE<5 in models adjusted for CRASH-basic, while FAQ was significantly associated with both GOSE<5 and GOSE<7, suggesting it adds value particularly for teasing apart higher thresholds of functional recovery. Incorporating FAQ into large-scale TBI cohorts and trials has potential to facilitate the inclusion of older adults with pre-injury cognitive impairment, support pre-planned sub-group analysis by pre-injury cognitive status, and to quantify this otherwise unmeasured source of heterogeneity.

TRACK-GERI is a real-world prospective cohort study of oaTBI that quantified multiple dimensions of pre-injury health, including patients with pre-existing cognitive impairment, multi-morbidity, and/or extreme frailty. This analysis compares associations between individual pre-injury measures and oaTBI outcomes, and may guide future studies to assess, rather than exclude, patients with pre-injury cognitive impairment and frailty. The independent effect of the individual pre-injury measures on the outcome demonstrates their value for future developed oaTBI outcome prediction models.

Limitations include low enrollment of participants with GCS<13, limiting generalizability to moderate-severe oaTBI. Additionally, our primary outcome of GOSE<5 should be interpreted with some caution in older adults with high levels of pre-existing disability, because small absolute changes in functional status may result in large absolute changes in GOSE in this context. This caution, along with major differences between TRACK-GERI and IMPACT/CRASH cohort characteristics, is why we additionally chose to show GOSE thresholds of <7 and <8. As in most prior long-term outcome studies of acute TBI,³³ the study was limited by missing data and high attrition, which was comparable to other studies reporting on oaTBI.⁴⁷ We found no significant differences between included and excluded participants; however, residual selection bias remains possible. Without a non-TBI trauma control group (e.g., orthopedic injury), we cannot isolate decline attributable to intracranial injury from the broader effects of trauma or hospitalization, particularly given the impact of extracranial injuries in mild TBI.⁴⁸ Although CRASH-basic model includes “major extracranial injury,” this binary measure may not adequately capture the type or sequelae of such injuries in oaTBI. Future studies should better characterize extracranial injury patterns and their consequences in oaTBI, which are likely to differ from those in younger adults. Recovery following oaTBI is multifactorial, reflecting interactions among pre-existing, injury-related, and post-injury biological, clinical, and social factors.⁴⁹ This analysis tackled a very small slice of this milieu and sought to efficiently quantify pre-injury health using brief, established geriatric assessments. What we have learned is that measured pre-injury cognitively mediated function and frailty were efficiently measured and independently associated with 6-month outcomes: this finding alone may provide a framework for more inclusive clinical trials, stratifying patients based on these frequently un-quantified sources of heterogeneity. However, further refinement of oaTBI outcome prediction models is needed. Lastly, additional unmeasured factors, such as detailed CT features, extra-cranial injury, blood-based biomarkers, and care-limiting decisions, may further improve prognostic accuracy.⁵⁰

Frailty is underused as a clinical measure in TBI research and care.⁹ The GFI’s brevity and multi-dimensionality²² make it a promising composite measure of pre-injury reserve and resiliency that is not currently captured using standard TBI Common Data Elements.⁵¹ These findings should not yet guide individual patient care but support the inclusion of older adults with TBI and pre-existing cognitively-mediated function impairment or frailty in large-scale studies to build a much-needed evidence base for this population. The independent associations of FAQ and GFI with 6-month GOSE, after adjustment for CRASH-basic, underscore their value in capturing clinically meaningful pre-injury health and provide a foundation for refining prognostic models and developing targeted interventions. Next steps towards resolving the residual unexplained heterogeneity in oaTBI outcome calls for large-scale studies across the spectrum of injury-severity and pre-injury health, refining age-appropriate injury severity measures, including age-appropriate blood biomarkers,⁵² and age-appropriate outcome measures.²⁰ Currently, there is a lack of consensus in the field on the specific means to measure age-related heterogeneity in TBI clinical research and this is limiting progress. Researchers seeking to measure frailty in older adult or neurotrauma or other surgical populations must choose from many different potential tools.⁹ There is a need to establish consensus on pre-injury health, injury severity, frailty, and outcome assessments for oaTBI to accelerate synergistic international research and trials.⁵³

Conclusions

Using a real-world cohort of older adults with heterogenous pre-injury health from the TRACK-GERI study, this study demonstrates that pre-injury health is independently associated with 6-month outcome in older adults with TBI, independent of injury severity (e.g., CRASH-basic model) metrics. Frailty and cognitively-mediated function (measured by GFI and FAQ) were statistically significantly associated with outcome across recovery thresholds: each point raised the odds of unfavorable outcome by 30–47% and 9–17% respectively. Importantly, a simple composite of multi-morbidity (CCI) was not associated with outcome, suggesting that “frailty indices” comprised of a summary of diagnosed medical conditions will do little to resolve outcome heterogeneity in real-world older-adult TBI. Functional reserve, and not multi-morbidity, is key in older adult’s recovery following a TBI. Brief, feasible tools like the GFI and FAQ efficiently capture aging heterogeneity, and may improve outcome prediction in older adults, reinforcing the urgent need for and feasibility of inclusive clinical trials in oaTBI.

Data Sharing Statement

All raw de-identified TRACK-GERI data is deposited at regular intervals to the Federal Interagency Traumatic Brain Injury Research (FITBIR) informatics system where qualified researchers may request access. The curated, de-identified, dataset used in the current analysis is available upon request following review and approval by the TRACK-GERI investigators.

Coinvestigators (Group Information: The TRACK-TBI Investigators/Authors, alphabetical listing):

Andrea L.C. Schneider, MD, PhD	Perelman School of Medicine, University of Pennsylvania, USA)
Claudia Robertson, MD	Department of Neurosurgery, Baylor College of Medicine, Texas, USA
David Schnyer, PhD	Department of Psychology, University of Texas, USA
John K. Yue, MD	University of California, San Francisco, USA
Michael McCrea, PhD, ABPP	Clinical Neuroscience Research Center, Wisconsin Institute of Neuroscience, USA
Pratik Mukherjee, MD PhD	University of California, San Francisco, USA
Ramesh Grandhi, MD, MS	University of Utah, Salt Lake City, USA
Sabrina R. Taylor, PhD, CCRP	University of California, San Francisco, USA
Shawn R. Eagle, PhD, ATC, CSCS	School of Medicine, University of Pittsburgh, USA
Vijay Krishnamoorthy, MD	Department of Anesthesiology, Duke University School of Medicine, USA