Downloaded from http://jnnp.bmj.com/ on November 18, 2015 - Published by group.bmj.com

JNNP Online First, published on February 18, 2015 as 10.1136/jnnp-2014-309479 General neurology

RESEARCH PAPER

Cost prediction following traumatic brain injury: model development and validation Gershon Spitz,1,2 Dean McKenzie,3,4 David Attwood,5 Jennie L Ponsford1,2 1

School of Psychological Sciences, Monash University, Melbourne, Australia 2 Monash-Epworth Rehabilitation Research Centre, Epworth Hospital, Melbourne, Australia 3 Clinical Trials and Research Centre, Epworth Healthcare, Melbourne, Australia 4 School of Public Health and Preventive Medicine, Monash University, Melbourne, Australia 5 Transport Accident Commission, Geelong, Australia Correspondence to Gershon Spitz, School of Psychological Sciences, Monash University, Clayton VIC 3800, Australia; [email protected] Received 16 September 2014 Revised 21 January 2015 Accepted 28 January 2015

ABSTRACT Objective The ability to predict costs following a traumatic brain injury (TBI) would assist in planning treatment and support services by healthcare providers, insurers and other agencies. The objective of the current study was to develop predictive models of hospital, medical, paramedical, and long-term care (LTC) costs for the first 10 years following a TBI. Methods The sample comprised 798 participants with TBI, the majority of whom were male and aged between 15 and 34 at time of injury. Costing information was obtained for hospital, medical, paramedical, and LTC costs up to 10 years postinjury. Demographic and injuryseverity variables were collected at the time of admission to the rehabilitation hospital. Results Duration of PTA was the most important single predictor for each cost type. The final models predicted 44% of hospital costs, 26% of medical costs, 23% of paramedical costs, and 34% of LTC costs. Greater costs were incurred, depending on cost type, for individuals with longer PTA duration, obtaining a limb or chest injury, a lower GCS score, older age at injury, not being married or defacto prior to injury, living in metropolitan areas, and those reporting premorbid excessive or problem alcohol use. Conclusions This study has provided a comprehensive analysis of factors predicting various types of costs following TBI, with the combination of injury-related and demographic variables predicting 23-44% of costs. PTA duration was the strongest predictor across all cost categories. These factors may be used for the planning and case management of individuals following TBI.

INTRODUCTION

To cite: Spitz G, McKenzie D, Attwood D, et al. J Neurol Neurosurg Psychiatry Published Online First: [ please include Day Month Year] doi:10.1136/ jnnp-2014-309479

The lifetime cost of each traumatic brain injury (TBI) in Australia is estimated to be $A2.5 million (US$2.1 million) for moderate and $A4.8 million (US$3.9 million) for severe injuries, with total costs estimated at $A8.6 billion (US$7.0 billion).1 These include costs for hospital and medical care, allied health therapy, equipment and modifications, longterm care (LTC) and loss of productivity.1 In comparison, the annual cost of TBI in Europe is estimated to be €33 billion (US$39.2 billion) and in the USA, this is estimated to be US$76.5 billion which comprises acute medical care, rehabilitation and loss in productivity.2 3 There is currently no clear basis on which to predict such costs for a given individual. The ability to make such a prediction would be of considerable assistance in planning treatment and support services by healthcare providers, insurers and other governmental agencies. No previous studies have developed comprehensive predictive cost models following TBI. Models

based on demographic and injury-related factors would allow for rapid and low cost predictions at the acute stage following injury. Studies to date have largely examined factors individually. Higher severity of injury has been most consistently related to higher costs.4–8 Other studies, however, have unexpectedly found that individuals with mild TBI accumulate greater costs compared to those with injury severity in the moderate range.4 9 10 Sustaining a TBI due to a gunshot, blow to the head or motor vehicle accident is also related to higher costs.6 8 Higher hospital costs are incurred due to comorbid physical injuries to the neck, thorax, abdomen and spine, surgical intervention for intracranial complications, and lower functional independence.6 11 12 The study by Morris et al12 is the only one to have developed a multivariate model, albeit one confined to hospital costs. They found that older age, worse injury severity and presence of other coexisting injuries were significantly related to higher hospital costs. Nevertheless, the model of cost developed in this study was not directly compared to other potentially viable models of cost and did not consider costs other than hospital charges. The objective of the current study was to develop competing predictive multivariate models of costs accrued from the time of the initial accident and over the subsequent 10 years following injury. Costs were predicted for hospital, medical, paramedical and LTC following complicated mild to severe TBI in the state of Victoria, Australia. This study builds on a recent study by our group which conducted some preliminary descriptive and exploratory analyses of cost indices.13

METHOD Study population and design The study was approved by the Human Ethics Committees of Epworth Healthcare and Monash University. Participants were recruited between 1987 and 2003 from consecutive admissions to a TBI rehabilitation centre in the context of a no-fault accident compensation system administered by the Transport Accident Commission (TAC) or Worksafe, at the time of inpatient admission. There were 798 eligible patients who had reached 10 years postinjury, for whom costing data was available from the TAC and for whom complete data were available for chosen demographic and injury-related variables.

Measures Cost variables Costing information was obtained from the TAC for hospital, medical, paramedical and LTC costs

Spitz G, et al. J Neurol Neurosurg Psychiatry 2015;0:1–8. doi:10.1136/jnnp-2014-309479

Copyright Article author (or their employer) 2015. Produced by BMJ Publishing Group Ltd under licence.

1

Downloaded from http://jnnp.bmj.com/ on November 18, 2015 - Published by group.bmj.com

General neurology between 1 and 10 years postinjury. Hospital costs included acute and rehabilitation hospital costs. Medical costs related to specialist medical or surgical consultations, as well as medical visits —including pathology, radiology and psychiatry. Paramedical costs included physiotherapy, speech pathology, occupational therapy, psychology, social work, dental costs, vocational services—such as equipment, training and counselling—and assistance in the home, such as gardening, domestic services and childminding. LTC comprised costs for attendant care, integration aides in schools, special accommodation, special equipment, living expenses and other community support requirements. Costs were inflated to match October 2013 cost values using quarterly inflation factors provided by the TAC. These inflation factors are derived from the quarterly consumer price index (CPI) and average weekly earnings (AWE) published by the Australian Bureau of Statistics (ABS). Owing to variable inflationary pressures, the inflation factors for past payments vary across benefit types. For example, income benefits are linked to wage inflation (AWE), whereas treatment benefits are largely impacted by price movements (CPI). Most of the TAC’s treatment benefits are considered to increase by 1% above price movements (CPI+1%). For attendant care costs, the factors take into account the claim-specific historical mix of payments between TAC contracted and non-contracted providers as well as historical changes in hourly care rates and are calculated from historical payment files. No discounting was applied to costs.

Demographic and injury-related variables Demographic and injury-severity variables were collected at the time of admission to the rehabilitation hospital. Demographic variables included: gender, age at injury, years of education, premorbid marital status (married/de facto vs widowed/divorced/ never married/single/separated), premorbid living location (metropolitan vs country/interstate), premorbid employment (employed/studying vs not in the labour force/unemployed), and premorbid medical history ( premorbid psychological treatment or diagnosis, excessive or problem drinking and other physical limitations). Injury-related variables included: cause of injury (car occupant, pedestrian, motorcycle, bicycle, fall or work related), Glasgow Coma Scale (GCS) score (mild 13–15, moderate 9–12 or severe 3–8), duration of post-traumatic amnesia (PTA), length of acute hospital stay, inpatient CT scan findings (normal vs abnormal), and injury-related physical injuries (back, chest, abdomen, limb or facial). PTA duration was determined from daily administration of the Westmead PTA scale.14 Accident-related physical injuries were categorised as ‘No/minor’ vs ‘Moderate/severe’, whereby the ‘Moderate/severe’ injuries were those resulting in fractures or requiring surgery. As the TAC largely funds TBIs due to motor-vehicle-related accidents, there was a consequent bias towards such injuries in the current sample. Owing to very low frequencies in other aetiologies, cause of injury was not included in modelling analysis.

importance and model selection was conducted on the learning data set and cross-validated on the hold-out data set. Importance of predictors to each cost type was initially assessed using a bootstrapped multiple regression including best subset regression, using the IBM SPSS implementation of bootstrap regression, also known as automatic linear modelling.16 17 Automatic linear modelling comprised 100 iterations of best subset based on Akaike Information Criterion, corrected for smaller sample (AICc).18–20 The AICc was used in the model building process to find a balance between model performance (eg, amount of variation in outcome variable accounted for by the model) and model complexity (eg, number of variables included). Variables that were chosen at least 50% of the time in the best subset regression were selected for subsequent models. Confining the predictors to only those that appear in at least 50% of bootstrap samples has been shown to exhibit good generality to other data sets, based on studies involving real data.21 22 A set of linear regression models was developed based on the results obtained from the initial bootstrapped best subset approach. Predictors were entered sequentially into the model based on the frequency with which they were chosen for each cost type, such that the final model for each cost type comprised all variables selected at least 50% of the time. Models for each cost type were evaluated simultaneously based on their AICc and BIC (Bayesian Information Criterion) to determine the best model.23 Owing to uncertainty in choosing the ‘best’ model, models with a difference of two on either the AICc or BIC were considered equally viable. Competing models within a difference of two were averaged.24 In addition to the AICc-derived and BIC-derived models, three other potentially viable models were considered: a ‘Simultaneous’ model comprising all predictors (reaching at least 50% criterion in the initial bootstrap stage), a model comprising the ‘Top three’ most frequent predictors, and a model comprising the ‘Best (most frequent) predictor’. The performance of the five models developed on the learning sample was assessed on the hold-out sample.25 The multiple correlation for each model in the learning sample was compared with that obtained when the latter model was applied to the hold-out sample using the standard test for correlation coefficient obtained from different samples.26 Cross-validation of the models from the learning to the hold-out sample was confirmed if the difference in model performance was clearly not statistically significant. Final model performance was evaluated by plotting the adjusted R2 of each model as it applied to the hold-out sample by the number of parameters in the model. The adjusted R2 was used as it adjusts for the number of predictors in the model. The adjusted R2 increases only if a new predictor improves the model more than would be expected by chance. The best model derived for each cost type was that which maximised the adjusted R2 while also having the least number of parameters. Adjusted R2 was plotted rather than AICc or BIC, as the former was thought to be a more readily understandable measure.

STATISTICAL ANALYSIS The natural logarithm of all 10-year aggregate costs was calculated due to positive skew and used in all analyses. Statistical models of costs were developed in several stages. Cross-validation of cost models was conducted to test the generality of the resulting models to other data.15 The original data set was divided into a learning sample comprising 75% of observations used to develop the model and a hold-out sample of 25% of observations used to test the model. Predictor 2

RESULTS Patient characteristics The final sample comprised 798 participants with TBI. Almost all individuals accrued some hospital (99%), medical (100%) and paramedical (99%) costs over 10 years following their injury. Less than half of the sample (41%), however, accrued LTC costs. The sample comprised largely younger individuals aged between 15 and 34, males at injury, with less than 12 years Spitz G, et al. J Neurol Neurosurg Psychiatry 2015;0:1–8. doi:10.1136/jnnp-2014-309479

Downloaded from http://jnnp.bmj.com/ on November 18, 2015 - Published by group.bmj.com

General neurology Table 1 Demographic and injury characteristics of participants Variable Age at injury (years) 15–24 25–34 35–44 45–54 55–64 >65 Sex Male Female Education (years) 12 Marital status Married/de facto Widowed/divorced/never married/single/separated Living location Metropolitan Country/interstate Employment Employed/studying Not in the labour force/unemployed Premorbid medical history Psychological disorder/treatment Excessive/problem drinking Other physical limitations Cause of injury Car occupant Pedestrian Motorcycle Bicycle Fall Work related Glasgow Coma Scale score Mild (13–15) Moderate (9–12) Severe (3–8) Length of post-traumatic amnesia (days) Moderate (0–14) Moderate severe (15–28) Severe (29–70) Extremely severe (>70) CT scan Abnormal Normal Injuries Back Chest Abdomen Limb Facial Costs indices Actual costs, $A Mean$ Median$ SD$ Range$ Logarithm Mean Median SD Range

Hospital (n=794), % (n)

Medical (n=798), % (n)

Paramedical (n=791), % (n)

Long-term care (n=324), % (n)

52.8 (419) 23.0 (183) 10.3 (82) 6.4 (51) 5.0 (40) 2.4 (19)

52.8 (421) 22.9 (183) 10.4 (83) 6.5 (52) 5.0 (40) 2.4 (19)

52.7 (417) 22.9 (181) 10.5 (83) 6.4 (51) 5.1 (40) 2.4 (19)

48.8 (158) 21.0 (68) 12.7 (41) 6.8 (22) 7.7 (25) 3.1 (10)

71.7 (569) 28.3 (225)

71.7 (572) 28.3 (226)

71.8 (568) 28.2 (223)

66.4 (215) 33.6 (109)

62.5 (496) 37.5 (298)

62.4 (498) 37.6 (300)

62.3 (493) 37.7 (298)

67.0 (217) 33.0 (107)

25.8 (205) 74.2 (589)

25.8 (206) 74.2 (592)

25.8 (204) 74.2 (587)

26.5 (86) 73.5 (238)

65.4 (529) 34.6 (275)

65.3 (521) 34.7 (277)

65.2 (516) 34.8 (275)

59.6 (196) 40.4 (131)

73.3 (582) 26.7 (212)

73.3 (585) 26.7 (213)

73.5 (581) 26.5 (210)

73.5 (238) 26.5 (86)

9.9 (79) 16.1 (128) 19.9 (158)

10.0 (80) 16.3 (130) 20.2 (161)

9.9 (78) 16.2 (128) 19.7 (156)

11.4 (37) 16.4 (53) 22.8 (74)

60.7 (482) 22.8 (181) 11.3 (90) 3.1 (25) 0.6 (5) 1.4 (11)

60.5 (483) 22.9 (183) 11.4 (91) 3.1 (25) 0.6 (5) 1.4 (11)

60.6 (479) 22.9 (181) 11.4 (90) 3.2 (25) 0.6 (5) 1.4 (11)

59.6 (193) 24.1 (78) 11.7 (38) 3.1 (10) 0.3 (1) 1.2 (4)

21.2 (168) 15.9 (126) 63.0 (500)

21.2 (169) 15.8 (126) 63.0 (503)

21.2 (168) 15.5 (123) 63.2 (500)

13.3 (43) 13.9 (45) 72.8 (236)

44.1 (350) 21.7 (172) 24.2 (192) 10.1 (80)

44.2 21.6 24.2 10.0

43.9 21.7 24.3 10.1

(347) (172) (192) (80)

22.5 (73) 19.1 (62) 37.0 (120) 21.3 (69)

80.2 (637) 19.8 (157)

80.1 (639) 19.9 (159)

80.0 (633) 20.0 (158)

84.6 (274) 15.4 (50)

20.0 (159) 35.1 (279) 21.3 (169) 62.0 (492) 29.6 (235)

19.9 35.0 21.2 61.8 29.6

20.1 35.1 21.1 62.2 29.8

25.6 (83) 43.5 (141) 23.5 (76) 67.3 (218) 29.6 (96)

78 362.0 59 604.2 87 476.9 879.4–1 266 687.2

26 463.3 20 467.3 22 433.3 1.6–160 186.2

37 902.1 21 549.1 47 113.2 53.7–308 826.9

1 101 108.5 6743.5 255 008.4 14.0–1 824 282.5

4.7 4.8 0.4 2.94–6.1

4.3 4.3 0.4 0.21–5.2

4.3 4.3 0.9 1.7–5.5

3.9 3.8 1.2 1.2–6.3

(353) (172) (193) (80)

(159) (279) (169) (493) (236)

(159) (278) (167) (492) (236)

Categorisation of injury severity for duration of post-traumatic amnesia used the cut-offs specified by Nakase-Richardson et al.32

Spitz G, et al. J Neurol Neurosurg Psychiatry 2015;0:1–8. doi:10.1136/jnnp-2014-309479

3

Downloaded from http://jnnp.bmj.com/ on November 18, 2015 - Published by group.bmj.com

General neurology Table 2 Bootstrapped predictor frequency Variable frequency (descending order) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

Hospital

Medical

Paramedical

Long-term care

Duration of post-traumatic amnesia (days) Limb injury

Duration of post-traumatic amnesia (days) Limb injury

Duration of post-traumatic amnesia (days) Living location

Glasgow Coma Scale score Age at injury (years) Physical limitation Psychological problem/ disorder Marital status Problem/excessive drinking Facial injury Living location Back injury Chest injury Gender Abnormal CT Abdomen injury Employment Education (years)

Chest injury Facial injury Employment Back injury

Problem/excessive drinking Marital status Limb injury Employment

Duration of post-traumatic amnesia (days) Psychological problem/ disorder Limb injury Length of hospital stay (days) Abnormal CT Facial injury

Problem/excessive drinking. Length of hospital stay (days) Living location Abdomen injury Psychological problem/disorder Physical limitation Age at injury (years) Marital status Gender Abnormal CT Glasgow Coma Scale score Education (years)

Glasgow Coma Scale score Length of hospital stay (days) Psychological problem/disorder Chest injury Physical limitation Age at injury (years) Facial injury Gender Back injury Education (years) Abnormal CT Abdomen injury

Living location Glasgow Coma Scale score Back injury Employment Age at injury (years) Abdomen injury Marital status Physical limitation Problem/excessive drinking Chest injury Gender Education (years)

Bold variables met the requirement for variables to be chosen at least 50% from 100 iterations. These variables formed the models in the subsequent modelling stage.

Figure 1 Model selection based on Akaike Information Criterion corrected for smaller sample (AICc) and Bayesian Information Criterion (BIC). Models subsumed within the rectangles were averaged due to being within a difference of 2 from either the lowest AICc or BIC model statistic. 4

Spitz G, et al. J Neurol Neurosurg Psychiatry 2015;0:1–8. doi:10.1136/jnnp-2014-309479

Downloaded from http://jnnp.bmj.com/ on November 18, 2015 - Published by group.bmj.com

General neurology Table 3

Model cross-validation using Fisher’s z comparisons of overall model performance (R) between training and testing data sets Models

Hospital Training Testing p Value** Medical Training Testing p Value** Paramedical Training Testing p Value** Long-term care Training Testing p Value**

AICc* R (Adj R2)

BIC† R (Adj R2)

Simultaneous‡ R (Adj R2)

Top three§ R (Adj R2)

Best predictor¶ R (Adj R2)

0.679 (0.45) 0.639 (0.38) 0.39

0.673 (0.45) 0.640 (0.39) 0.48

0.699 (0.48) 0.691 (0.46) 0.85

0.660 (0.44) 0.643 (0.40) 0.72

0.613 (0.38) 0.615 (0.38) 0.97

0.544 (0.29) 0.527 (0.25) 0.77

0.435 (0.18) 0.505 (0.24) 0.63

0.550 (0.29) 0.527 (0.24) 0.69

0.507 (0.26) 0.508 (0.22) 0.99

0.340 (0.12) 0.379 (0.14) 0.59

0.535 (0.28) 0.471 (0.19) 0.30

0.534 (0.28) 0.470 (0.19) 0.30

0.545 (0.20) 0.454 (0.18) 0.14

0.494 (0.21) 0.454 (0.20) 0.53

0.435 (0.15) 0.387 (0.15) 0.48

0.643 (0.39) 0.568 (0.23) 0.36

0.625 (0.38) 0.569 (0.29) 0.50

0.657 (0.41) 0.582 (0.26) 0.35

0.625 (0.38) 0.568 (0.29) 0.50

0.591 (0.35) 0.572 (0.32) 0.83

Bold values were not significant results. Correlations and Adj R2 between predicted and actual costs compared for training and testing data sets. Non-significant differences between correlations indicate cross-validation of models. *Costs predicted using the best (or best averaged) model based on AICc. †Costs predicted using the best (or best averaged) model based on BIC. ‡Costs predicted using the entire set of variables reaching the 50% criterion for each cost type (refer to shaded variables in table 2). §Costs predicted using the top three most frequently occurring predictors in each bootstrapping stage for each cost type. ¶Costs predicted using the single most frequent variable for each cost type. **Correlations for training and testing data sets compared using test based on Fisher’s z transformation. AICc, Akaike Information Criterion corrected for smaller sample; Adj R2, adjusted R2; BIC, Bayesian Information Criterion.

of education. Two-thirds of the sample acquired their TBI as car occupants. Most individuals accrued a severe injury, assessed using both GCS scores and duration of PTA; individuals had an abnormal CT scan, and some form of moderate or major physical injury (table 1).

Predictor importance Duration of PTA was the most important single predictor for each cost type, occurring most frequently using the bootstrapped regression approach (table 2). A combination of premorbid and injury-related variables were found to predict each cost type. Gender and years of education were poor predictors for all cost types. As is shown in table 2, gender and years of education were not chosen more than 50% of the time in any of the cost bootstrapped models. Variables not meeting the 50% selection criterion were excluded from any further analysis.

AICc and BIC model development Multiple competitive AICc-based models were found for each cost type; that is, models with a difference of two from the model with the lowest AICc (figure 1). For this reason, model averaging was conducted for each of the cost types to develop AICc-based models. A preferred BIC-based model was obtained for paramedical and LTC costs. Model averaging for BIC-based models was undertaken for hospital and medical cost types.

Performance between learning and testing data sets: cross-validation of models Cost models were cross-validated by comparing the Pearson r correlations between the model predicted and actual costs (logarithm) for the learning and hold-out data sets (table 3). All models developed on the learning data set (AICc, BIC, Spitz G, et al. J Neurol Neurosurg Psychiatry 2015;0:1–8. doi:10.1136/jnnp-2014-309479

Simultaneous, Top 3 and Best predictor) were found to generalise for the hold-out data set.

Final model selection Final model selection was conducted by examining the relationship between the adjusted R2 and the number of parameters in each model as it applied to the hold-out sample (figure 2). The ‘Simultaneous’ model was the best performing model for hospital cost prediction, explaining 46% of the variability and comprising nine variables. This model comprised duration of PTA, a moderate or major limb injury, GCS score, age at injury, premorbid physical limitation, premorbid treatment for a psychological disorder, marital status, premorbid excessive or problem alcohol use and moderate or major facial injury. The ‘Top three’ model was the best performing model of medical cost prediction, explaining 22% of the variability in costs. This model comprised duration of PTA, moderate or major limb injury or moderate or severe chest injury. The ‘Top three’ model was also the best performing model of paramedical costs, explaining 18% of costs. This model comprised duration of PTA, living location and premorbid problem or excessive alcohol use. The ‘Best predictor’ model was the best model of LTC costs, explaining 32% of costs. This model comprised only the duration of PTA. The final models were applied to the full sample, comprising both the learning and hold-out samples, due to their successful cross-validation performance (table 4). The hospital model explained 44% of the variability in costs using the full sample. Individuals incurred greater hospital costs the longer they were in PTA, if they experienced a moderate or major limb injury, had a lower GCS score, were older at the time of injury, and were not married or de facto prior to injury. Premorbid physical 5

Downloaded from http://jnnp.bmj.com/ on November 18, 2015 - Published by group.bmj.com

General neurology

Figure 2 Final model selection by plotting adjusted R2 by number of parameters in each model. These plots represent the relationship between the adjusted R2 and the number of parameters in each of the five models. The final model chosen was that which maximised the adjusted R2 while having the fewest parameters. Final chosen models for each cost type are demarcated by a rectangle. Best predictor=Best predictor for each cost type (this was post-traumatic amnesia for all models); Top three=top three best predictors for each cost type, Bayesian Information Criterion (BIC) =BIC-based model; AICc= Akaike Information Criterion corrected for smaller sample ; Simulatensous=all variables reaching the 50% criterion in the iadjustnitial boostrapping approach. limitations, psychological treatment, excessive or problem alcohol use, and facial injuries were not significant contributing variables in this final model. The medical model explained 26% of the variability in hospital costs. Greater medical costs were incurred by individuals with a longer PTA duration, and those with moderate/major limb or chest injury. The paramedical model explained 23% of variability in costs. Greater paramedical costs were incurred by those with a longer PTA duration, those living in a metropolitan location prior to injury, and those that reported premorbid excessive or problem alcohol use. The LTC model explained 34% of variability in LTC costs. Individuals with longer PTA duration incurred greater LTC costs.

DISCUSSION This study found that a significant proportion of costs following TBI could be predicted using factors that were readily available at or soon after injury. Nearly half of hospital costs and one-third of LTC costs were predicted, while the models predicting medical and paramedical costs accounted for around a quarter of cost variability. PTA duration was the single best predictor for each cost type. It was the sole predictor of LTC costs, explaining 34% of the variability. This is significant, given that LTC costs represent the 6

highest overall cost category. PTA emerged most frequently in the initial stage of analysis to represent the ‘best subset’ of predictors across all cost subtypes. Although a number of previous studies have shown an association between injury severity and costs, injury severity was generally based on GCS.4–6 In the present study, PTA duration proved to be a stronger predictor of costs than GCS, although GCS was among the initial set of predictors across most cost subtypes. PTA has consistently emerged as a strong predictor of longer term functional outcome in recent studies—more so than GCS.27–31 Although debate continues regarding the cut-offs which best categorise level of injury severity according to PTA, this finding supports the need for greater consistency in measurement of PTA duration as a measure of injury severity in patients surviving past acute injury hospital discharge.32 In addition to PTA duration, presence of other injuries such as moderate or severe limb, facial, chest or back injuries also contributed significantly to cost prediction across all cost subtypes; in particular, limb and chest injuries predicted 26% of medical costs. Previous studies have found comorbid injuries to be associated with higher hospital costs, which in countries with public healthcare systems would also include a high proportion of medical costs.6 11 12 Spitz G, et al. J Neurol Neurosurg Psychiatry 2015;0:1–8. doi:10.1136/jnnp-2014-309479

Downloaded from http://jnnp.bmj.com/ on November 18, 2015 - Published by group.bmj.com

General neurology Table 4

Final model coefficients for each cost type 95% CI

Hospital Intercept Length of PTA (days) Limb injury GCS Age at injury (years) Physical limitations Psychological disorder/treatment Married or de facto Excessive/problem alcohol Facial injury Medical Intercept Length of PTA (days) Limb injury Chest injury Paramedical Intercept Length of PTA (days) Metropolitan living Excessive/problem drinking Long-term care Intercept Length of PTA (days)

Coefficient

SE

Lower

Upper

4.22 0.01 0.16 0.06 0.003 0.05 0.06 −0.05 −0.02 0.02

0.04 0.0003 0.02 0.01 0.001 0.03 0.04 0.03 0.03 0.02

4.129 0.006 0.117 0.032 0.001 −0.0002 −0.007 −0.103 −0.081 −0.027

4.300 0.007 0.201 0.086 0.005 0.107 0.133 −0.001 0.033 0.061

3.95 0.004 0.24 0.17

0.02 0.0004 0.03 0.03

3.908 0.003 0.186 0.115

3.999 0.004 0.289 0.222

3.924 0.008 0.235 −0.221

0.037 0.001 0.039 0.050

3.851 0.007 0.160 −0.319

3.996 0.009 0.311 −0.123

3.147 0.016

0.079 0.001

2.992 0.014

3.303 0.019

This table presents the results of applying the final models to the full sample, comprising the learning as well as hold-out subsamples. Variables were obtained using automated search methods, which CIs do not take into accounts, and so should be employed as a guide only. GCS, Glasgow Coma Scale; PTA, post-traumatic amnesia.

Demographic variables also contributed to cost prediction for certain cost subtypes, namely hospital costs and paramedical costs. However, LTC costs were predicted by PTA duration alone and medical costs were predicted by a combination of PTA and other injuries, older age and not being married or de facto prior to injury contributed significantly to higher hospital costs. Older age brings with it potentially more severe injury effects, comorbidities and other social factors that may prolong hospital stay. Explaining the influence of marital status is more challenging. One could speculate, however, that earlier discharge may be deemed appropriate if patients could be cared for by their partners at home. This predictive model is quite consistent with the multivariate model proposed by Morris et al,12 whereby older age at injury, worse injury severity and other accident-related physical injuries— specifically presence of a moderate or major limb injury—predicted 25% of hospital costs. In addition to being associated with longer PTA duration, greater paramedical costs were associated with living in a metropolitan location and the absence of premorbid excessive alcohol use. The study findings suggest that individuals in metropolitan areas have greater accessibility to paramedical services, compared to those living in regional areas. Post hoc analyses also showed that individuals living in metropolitan areas had fewer days of PTA which may also, in part, explain this finding. This is most likely due to patients with more severe injuries being more likely to be transferred to specialist acute trauma hospitals and the rehabilitation centre located in the metropolitan area. Those with less severe injuries may be managed in local regional areas. Patients with a premorbid history of excessive or problem Spitz G, et al. J Neurol Neurosurg Psychiatry 2015;0:1–8. doi:10.1136/jnnp-2014-309479

alcohol use also incurred fewer paramedical costs. Although we cannot offer a definitive conclusion regarding this finding, it may be speculated that individuals with such a history are less likely to attend and comply with rehabilitation services. Lastly, duration of PTA was the sole predictor of LTC costs, explaining 34% of the variability and further demonstrating the importance of this factor. These findings suggest that the need for long-term services, such as attendant care support, integration aides in schools, special accommodation and/or equipment, is determined largely by injury severity rather than by demographic factors, in the context of a system that provides funds for such services. This study does have some limitations. The study sample was diminished by missing data. The sample with complete data, however, did not differ from that with missing data in terms of injury severity, years of education or gender, although it was significantly younger on average than the group with missing data. In addition, due to patients being recruited from admission to a TBI rehabilitation centre, the study sample largely comprises moderate and severely injured individuals who are likely to require ongoing rehabilitation. Although models predicting LTC costs and hospital costs accounted for the majority of the variance, there was still substantial variance unaccounted for and the amount of variance accounted for by the models predicting medical and paramedical costs was limited to around a quarter of the variance. Other injury-related, demographic and psychosocial factors not captured in the present study may influence these costs. Finally, it must be acknowledged that these costs were generated in the context of a system providing comprehensive funding for necessary hospital, medical, paramedical and LTC services. 7

Downloaded from http://jnnp.bmj.com/ on November 18, 2015 - Published by group.bmj.com

General neurology CONCLUSION This study has provided a comprehensive analysis of factors predicting various types of costs following moderate to severe TBI, with the combination of injury-related and demographic variables predicting 23–44% of costs, depending on cost type. PTA duration was the strongest predictor across all cost categories. Demographic factors including age, marital status, living in city versus regional areas and history of substance use are also important determinants of certain types of costs. These may be used for the planning and case management of individuals following TBI. Further research is required to examine the contribution of other factors to costs of care following moderate to severe TBI. Contributors GS designed the study and finalised objectives and aims, collected data, wrote the statistical analysis plan, cleaned and analysed the data, and drafted and revised the paper. He is the guarantor. DM analysed the data, and drafted and revised the paper. DA monitored data collection, revised study aims and objectives, and revised the draft. JLP designed the study, and finalised objectives and aims, and drafted and revised the paper. She is also the guarantor.

9

10 11 12 13 14

15 16 17 18 19

Funding This study was funded by the Transport Accident Commission (TAC), through the Institute for Safety, Compensation and Recovery Research (ISCRR). The cost data used in the current study was supplied by the TAC.

20

Competing interests DA has reported to be an employee of the Transport Accident Commission.

22

21

Ethics approval Epworth Human Research Ethics Committee. Provenance and peer review Not commissioned; externally peer reviewed.

23 24 25

REFERENCES 1 2 3 4

5

6

7

8

8

Access Economics. The economic cost of spinal cord injury and traumatic brain injury in Australia, report for the Victorian neurotrauma initiative, June. 2009. Finkelstein EA, Corso PS, Miller TR. The incidence and economic burden of injuries in the United States. New York: Oxford University Press, 2006. Olesen J, Gustavsson A, Svensson M, et al. The economic cost of brain disorders in Europe. Eur J Neurol 2012;19:155–62. Bryant ET, Sundance P, Hobbs A, et al. Managing costs and outcome of patients with traumatic brain injury in an HMO setting. J Head Trauma Rehabil 1993;8:15–29. Cowen TD, Meythaler JM, Devivo MJ, et al. Influence of early variables in traumatic brain injury on functional independence measure scores and rehabilitation length of stay and charges. Arch Phys Med Rehabil 1995;76:797–803. Lehmkuhl L, Hall KM, Mann N, et al. Factors that influence costs and length of stay of persons with traumatic brain injury in acute care and inpatient rehabilitation. J Head Trauma Rehabil 1993;8:88–100. Mayer NH, Pelensky J, Whyte J, et al. Characterization and correlates of medical and rehabilitation charges for traumatic brain injury during acute rehabilitation hospitalization. Arch Phys Med Rehabil 2003;84:242–8. McGarry LJ, Thompson D, Millham FH, et al. Outcomes and costs of acute treatment of traumatic brain injury. J Trauma Injury Infect Crit Care 2002;53:1152–9.

26 27

28

29

30

31

32

Homaifar BY, Harwood JE, Wagner TH, et al. Description of outpatient utilization and costs in group of veterans with traumatic brain injury. J Rehabil Res Dev 2009;46:1003–10. Prang KH, Ruseckaite R, Collie A. Healthcare and disability service utilisation following transport-related traumatic brain injury. Brain Inj 2012;26:441–2. High WM, Hall KM, Rosenthal M, et al. Factors affecting hospital length of stay and charges following traumatic brain injury. J Head Trauma Rehabil 1996;11:85–96. Morris S, Ridley S, Lecky FE, et al. Determinants of hospital costs associated with traumatic brain injury in England and Wales. Anaesthesia 2008;63:499–508. Ponsford JL, Spitz G, Cromarty F, et al. Costs of care after traumatic brain injury. J Neurotraum 2013;30:1498–505. Shores E, Marosszeky J, Sandanam J, et al. Preliminary validation of a clinical scale for measuring the duration of post-traumatic amnesia. Med J Australia 1986;144:569–72. Witten IH, Frank E, Hall MA. Data mining: practical machine learning tools and techniques. 3rd edn. San Francisco, CA: Morgan Kaufmann, 2011. Norusis M, Inc. S. IBM SPSS Statistics 19 statistical procedures companion. Boston, MA: Addison-Wesley, 2011. Ryan TP. Modern regression methods. 2nd edn. Hoboken, NJ: Wiley, 2009. Akaike H. Information theory and an extension of the maximum likelihood principle. 2nd International Symposium on Information Theory; 1973; Budapest. Burnham KP, Anderson DR. Model selection and multi-model inference: a practical information theoretic approach. 2nd edn. New York: Springer-Verlag, 2002. Hocking R, Leslie R. Selection of the best subset in regression analysis. Technometrics 1967;9:531–40. Austin PC, Tu JV. Bootstrap methods for developing predictive models. Am Stat 2004;58:131–7. Billah B, Reid CM, Shardey GC, et al. A preoperative risk prediction model for 30-day mortality following cardiac surgery in an Australian cohort. Eur J Cardio-Thorac 2010;37:1086–92. Schwarz G. Estimating the dimension of a model. Ann Stat 1978;6:461–4. Burnham KP, Anderson DR. Multimodel inference—understanding AIC and BIC in model selection. Sociol Method Res 2004;33:261–304. McKenzie DP, McFarlane AC, Creamer M, et al. Hazardous or harmful alcohol use in Royal Australian Navy veterans of the 1991 Gulf War: identification of high risk subgroups. Addict Behav 2006;31:1683–94. Zar JH. Biostatistical analysis. 5th edn. Upper Saddle River, NJ: Prentice Hall, 2006. Brown AW, Malec JF, McClelland RL, et al. Clinical elements that predict outcome after traumatic brain injury: a prospective multicenter recursive partitioning (decision-tree) analysis. J Neurotraum 2005;22:1040–51. Brown AW, Malec JF, Mandrekar J, et al. Predictive utility of weekly post-traumatic amnesia assessments after brain injury: a multicentre analysis. Brain Inj 2010;24:472–8. Walker W, Ketchum J, Marwitz J, et al. A multicentre study on the clinical utility of post-traumatic amnesia duration in predicting global outcome after moderate-severe traumatic brain injury. J Neurol Neurosurg Psychiatry 2010;81:87–9. Sherer M, Struchen M, Yablon S, et al. Comparison of indices of traumatic brain injury severity: Glasgow Coma Scale, length of coma and post-traumatic amnesia. J Neurol Neurosurg Psychiatry 2008;79:678–85. Ponsford JL, Spitz G. Stability of employment over the first 3 years following traumatic brain injury. J Head Trauma Rehabil 2014;doi;10.1097/ HTR.0000000000000033. Nakase-Richardson R, Sherer M, Seel RT, et al. Utility of post-traumatic amnesia in predicting 1-year productivity following traumatic brain injury: comparison of the Russell and Mississippi PTA classification intervals. J Neurol Neurosurg Psychiatry 2011;82:494–9.

Spitz G, et al. J Neurol Neurosurg Psychiatry 2015;0:1–8. doi:10.1136/jnnp-2014-309479

Downloaded from http://jnnp.bmj.com/ on November 18, 2015 - Published by group.bmj.com

Cost prediction following traumatic brain injury: model development and validation Gershon Spitz, Dean McKenzie, David Attwood and Jennie L Ponsford J Neurol Neurosurg Psychiatry published online February 18, 2015

Updated information and services can be found at: http://jnnp.bmj.com/content/early/2015/02/18/jnnp-2014-309479

These include:

References

This article cites 23 articles, 5 of which you can access for free at: http://jnnp.bmj.com/content/early/2015/02/18/jnnp-2014-309479#BIBL

Email alerting service

Receive free email alerts when new articles cite this article. Sign up in the box at the top right corner of the online article.

Topic Collections

Articles on similar topics can be found in the following collections Injury (466) Neurological injury (381) Trauma (467) Trauma CNS / PNS (381)

Notes

To request permissions go to: http://group.bmj.com/group/rights-licensing/permissions To order reprints go to: http://journals.bmj.com/cgi/reprintform To subscribe to BMJ go to: http://group.bmj.com/subscribe/

Cost prediction following traumatic brain injury: model development and validation.

The ability to predict costs following a traumatic brain injury (TBI) would assist in planning treatment and support services by healthcare providers,...
2MB Sizes 0 Downloads 19 Views