Predicting Outcome after Severe Traumatic Brain Injury

MISSING IMAGE

Material Information

Title:
Predicting Outcome after Severe Traumatic Brain Injury
Physical Description:
1 online resource (84 p.)
Language:
english
Creator:
Watson,William D
Publisher:
University of Florida
Place of Publication:
Gainesville, Fla.
Publication Date:

Thesis/Dissertation Information

Degree:
Doctorate ( Ph.D.)
Degree Grantor:
University of Florida
Degree Disciplines:
Psychology, Clinical and Health Psychology
Committee Chair:
Heaton, Shelley C
Committee Members:
Price, Catherine
Kerkhoff, Thomas R
Leonard, Christiana M
Hannay, H.

Subjects

Subjects / Keywords:
brain -- injury -- model -- prognostic -- traumatic
Clinical and Health Psychology -- Dissertations, Academic -- UF
Genre:
Psychology thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract:
Outcomes for patients who survive a severe TBI vary considerably. Researchers strive to predict not only mortality but also long term outcome. A recent model has proven to predict crude outcome after severe TBI. The present study sought to replicate and extend these findings to provide more detailed characterization of outcome at 6-months post-injury. Participants consisted of 546 adults ages 18-89 with severe TBI. Acute physiological data was gathered for use in the extended IMPACT model and included: patient?s age, pupil response, GCS motor score, CT characteristics, as well as events of hypoxia and hypotension. Outcome measures obtained at 6-month post injury included mortality, favorable/unfavorable outcome on the Glasgow Outcome Scale (GOS), Disability Rating Scale (DRS), and Digit Span. This study replicated the IMPACT study findings and showed that the model is sensitive and specific for predicting mortality and favorable outcome in this sample. The model shows promise for using acute physiological data to predict detailed long-term outcome. The model showed limited utility when predicting cognitive working memory ability 6 months after injury. The current model shows promise for creating more precise classifications of injury severity to help reduce the heterogeneity among severe TBI patients for research purposes and clinical trials. From a clinical standpoint, more precision in severe TBI classification could provide a more accurate clinical picture than is currently available which would enable quicker responses and better exchange of information among physicians. The IMPACT model could provide a powerful tool to communicate information among clinicians and to patient?s families.
General Note:
In the series University of Florida Digital Collections.
General Note:
Includes vita.
Bibliography:
Includes bibliographical references.
Source of Description:
Description based on online resource; title from PDF title page.
Source of Description:
This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility:
by William D Watson.
Thesis:
Thesis (Ph.D.)--University of Florida, 2011.
Local:
Adviser: Heaton, Shelley C.

Record Information

Source Institution:
UFRGP
Rights Management:
Applicable rights reserved.
Classification:
lcc - LD1780 2011
System ID:
UFE0043085:00001


This item is only available as the following downloads:


Full Text

PAGE 1

1 PREDICTING OUTCOME AFTER SEVERE TRAUMATIC BRAIN INJURY By WILLIAM D. WATSON A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2011

PAGE 2

2 2011 William D. Watson

PAGE 3

3 To my mom

PAGE 4

4 ACKNOWLEDGMENTS I thank the chair and members of my committee for their mentorship. I thank my family and friends for their encouragement throughout this proc ess.

PAGE 5

5 TABLE OF CONTENTS page TABLE OF CO NTENTS .................................................................................................. 5 LIST OF TABLES ............................................................................................................ 7 LIST OF FIGURES .......................................................................................................... 8 ABSTRACT ..................................................................................................................... 9 CHAPTER 1 BACKGROUND .......................................................................................................... 11 Introduction ............................................................................................................. 11 Existing Prognostic Models ..................................................................................... 12 Predictors Used in Prognost ic Models .................................................................... 14 Glasgow Coma Scale (GCS) ............................................................................ 14 Age ................................................................................................................... 16 Pupil Response ................................................................................................ 17 Structural Neuroimaging ................................................................................... 17 Secondary injury ............................................................................................... 19 Measurement of Outcome after Traumatic Brain Injury .......................................... 19 Global Functional Outcome .............................................................................. 20 Cognitive Outcome ........................................................................................... 23 2 OVERVIEW AND STATEMENT OF THE PROBL EM ................................................ 27 Introduction ............................................................................................................. 27 Using the International Mission for Prognosis and Analysis of Clinical Trials (IMPACT) Study Model to Predict Outcome ........................................................ 27 Aim 1 ................................................................................................................ 30 Aim 2 ................................................................................................................ 31 Aim 3 ................................................................................................................ 31 3 METHODS ................................................................................................................. 33 General Methods .................................................................................................... 33 Participants ............................................................................................................. 33 Procedure ............................................................................................................... 34 Measurement of Physiological Predictor Variables ................................................. 34 Admission Glasgow Coma Scale (GCS) Motor Scale ...................................... 34 Pupil Response ................................................................................................ 35 Head CT Classification ..................................................................................... 35 Hypoxia ............................................................................................................ 36 Hypot ension ..................................................................................................... 36

PAGE 6

6 IMPACT Model Sum Scores ............................................................................. 36 Measurement of Global and Cog nitive Outcome .................................................... 39 Global Functional Outcome Measures ............................................................. 40 Glasgow Outcome Scale (GOS) ................................................................ 40 Disability Rating Scale (DRS) .................................................................... 40 Cognitive Outcome Measure: Digit Span .......................................................... 40 Analyses Utilized .................................................................................................... 41 Aim 1 Replication of IMPACT Core and Extended Models Predicting Mortali ty and Unfavorable Outcome on GOS ................................................ 41 Aim 2 Extension of IMPACT Model Predicting Full GOS and DRS at Six Months .......................................................................................................... 42 Aim 3 Extension of IMPACT Predicting Working Memory Outcome at Six Months .......................................................................................................... 43 4 RESULTS ................................................................................................................... 44 Sample Characteristics ........................................................................................... 44 Aim 1a Replication of IMPACT Core and Extended models Predicting Mortality .... 47 Aim 1b. Replication of IMPACT Core and Extended Models Predicting Unfavorable Outcome .......................................................................................... 51 Aim 2a Extension of IMPACT Model Predicting Full Range of Functional 6month outcomes (GOS). ...................................................................................... 55 A im 2b Extension of IMPACT model predicting DRS at six months ........................ 60 Aim 2c Comparison of IMPACT prediction of GOS and DRS ................................. 61 Aim 3 Extension of IMPACT Predicting Working Memory Outcome at 6months ... 61 5 DISCUSSION ............................................................................................................. 63 Study Summary ...................................................................................................... 63 Hypothesis 1 ........................................................................................................... 64 Hypothesis 2 ........................................................................................................... 67 Hypothesis 3 ........................................................................................................... 69 Implications ............................................................................................................. 70 Study Strengths ...................................................................................................... 72 Study Limitations .................................................................................................... 73 Study Summary and Future Directions ................................................................... 73 LIST OF REFERENCES ............................................................................................... 76 BIOGRAPHICAL SKETCH ............................................................................................ 84

PAGE 7

7 LIST OF TABLES Table page 1 1 Impact models .................................................................................................... 14 4 1 Demographic characteristics of the sample (N = 546) ........................................ 44 4 2 Patient characteristics of present sample and IMPACT study ............................ 46 4 3 Patient outcomes of present sample and IMPACT study ................................... 47 4 4 Logistic regression analysis for mortality (n = 546) ............................................. 48 4 5 Sensitivity and specificity of the IMPACT model for mortality prediction ............. 48 4 6 Logistic regression analysis for Unfavorable outcome (n = 546) ........................ 51 4 7 Sensitivity and specificity of the IMPACT model for Unfavorable outcome prediction ............................................................................................................ 52 4 8 Sensitivity and specificity at the classification values of the IMPACT Core and Extended models for mortality and Unfavorable outcome prediction .................. 55 4 9 Comparison of observed versus predicted categories in ordinal regression model using IMPACT total Score predicting GOS category ............................... 57 4 10 Summary of ordinal regression of IMPACT total score predicting 6month GOS ................................................................................................................... 58 4 11 Sensitivity and specificity of the IMPACT model for less than Good Recovery outcome prediction ............................................................................................. 58 4 12 Sensitivity and specificity at the classification values of the IMPACT Core and Extended models for Good Recovery prediction ................................................ 59 4 13 S ummary of proportional odds model for Core and Extended IMPACT model predicting GOS ................................................................................................... 60 4 14 Summary of hierarchical regression analysis with IMPACT scores predicting DRS scores ........................................................................................................ 61 4 15 Digit Span scores of the current study sample (N=76) ....................................... 62 4 16 Summary of hierarchical regression analysis with education, IMPACT scores predicting Digit Span total scores (N=76) ........................................................... 62

PAGE 8

8 LIST OF FIGUR ES Figure page 2 1 IMPACT model score chart for 6month outcome after TBI ................................ 28 3 1 IMPACT model score chart for 6month outcome after TBI ................................ 37 4 1 ROC curve for Core model predicting mortality .................................................. 50 4 2 ROC curve for Core + Extended model predicting mortality ............................... 50 4 3 ROC curve for the Core IMPACT model predi cting Unfavorable outcome ......... 53 4 4 ROC curve for Core + Extended IMPACT model predicting Unfavorable outcome .............................................................................................................. 53 4 5 Plot of observed cumulative percentages IMPACT Core + Extended scores ..... 56

PAGE 9

9 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy PREDICTING OUTCOME AFTER SEVERE TRAUMATIC BRAIN INJURY By William D. Watson August 2011 Chair: Shelley C. Heaton Major: Psychology Outcomes for patients who survive a severe TBI vary considerably. Researchers strive to predict not only mortality but also long term outcome. A recent model has proven to predict crude outcome after severe T BI. The present study sought to replicate and extend these findings to provide more detailed characterization of outcome at 6months post injury. Participants consisted of 546 adults ages 1889 with severe TBI. Acute physiological data was gathered for use in the extended IMPACT model and included: patients age, pupil response, GCS motor score, CT characteristics, as well as events of hypoxia and hypotension. Outcome measures obtained at 6month post injury included mortality, favorable/unfavorable outcome on the Glasgow Outcome Scale (GOS), Disability Rating Scale (DRS), and Digit Span. This study replicated the IMPACT study findings and showed that the model is sensitive and specific for predicting mortality and favorable outcome in this sample. The model shows promise for using acute physiological data to predict detailed longterm outcome. The model showed limited utility when predicting cognitive working memory ability 6 months after injury.

PAGE 10

10 The current model shows promise for creating more precise classifications of injury severity to help reduce the heterogeneity among severe TBI patients for research purposes and clinical trials. From a clinical standpoint, more precision in severe TBI classification could provide a more accurate clinical picture than is currently available which would enable quicker responses and better exchange of information among physicians. The IMPACT model could provide a powerful tool to communicate information among clinicians and to patients families.

PAGE 11

11 CHAPTER 1 BACKGROUN D Introduction Patients who suffer severe traumatic brain injuries (TBI) are frequently admitted to the hospital in very critical condition. These early stages are pivotal for patient survival, and physicians must often make quick decisions regarding the use of certain therapeutic interventions. For this reason, prognostic models have been developed to use data that can be obtained promptly upon admission to detect a patients declining medical status and determine whether more aggressive interventions should be initiated. Researchers have recently used data a from large multisite sample from the International Mission for Prognosis and Analysis of Clinical Trials project to create a prognostic model (IMPACT model) that can effectively predict mortality at s ix months after a TBI (Steyerberg et al., 2008). However, once patients have been stabilized, the concerns of the practitioners and families change from keeping a patient alive to maximizing their recovery. As patients continue to recover, plans are made f or their discharge from the hospital and decisions are made regarding rehabilitation services and supervision needs (De Guise et al., 200 6 ). This task is becoming increasingly more difficult as the length of stay in the hospital has been continually decreasing over time for TBI patients (Ashley & Ashley, 2004). Discharge planners now have less time and less information to make suitable plans for rehabilitation or home environment. More sensitive prognostic information is important in the counseling of patients and loved ones at this important time regarding outcome beyond general Favorable or Unfavorable including issues of independent living, employment, and leisure activities (Perel et al., 2006).

PAGE 12

12 A prognostic model that could predict more detailed outcome would be a valuable asset to assist in the process of discharge planning. This could give families a better idea of what to expect long term in reference to a wider range of global functioning and a more detailed picture of cognitive functioning. The current study proposal builds upon the previously validated prognostic model of TBI outcome based on the extensive IMPACT study. The IMPACT model as it has been produced only predicts to gross outcome. The current proposal seeks to determine whether more spe cific levels of long term functioning including cognitive abilities can be predicted. The following sections will briefly review previous efforts to identify acute clinical variables that contribute to the predictive ability of prognostic models in TBI. Research on each of the individual predictors used in the IMPACT model will be presented. A great deal of research has been conducted regarding the predictor variables; however, methods of measuring outcome have been less well developed. Therefore, research regarding the issues of feasibility, ecological validity, and psychometric properties of different methods of outcome measurement will be discussed. Existing Prognostic Models Existing prognostic models have revealed various combinations of acute clinical variables associated with general outcome after TBI. Several early outcome studies found that the admission Glasgow Coma Scale (GCS) score, when combined with age and pupil reactions were accurate in predicting mortality and morbidity (Jennett et al., 1976; Choi et al., 1983; Narayan et al., 1981). Studies then began to build on these predictors and examined the contributions of other variables such as CT data (Choi et al., 1983), period of post traumatic amnesia (Bishara et al., 1992; Ellenberg et al., 1996), intracranial pressure (Narayan et al., 1981) and cerebral perfusion pressure

PAGE 13

13 (Changaris et al., 1987). These outcome studies, like others carried out in acute care settings have focused on mortality and morbidity (Elf et al., 2002, Jeremitsky et al., 2003). Over the years, many prognostic models have been developed, but few have been widely used due to small sample sizes and poor methodology (Young and Andrews, 2008). The creators of the IMPACT model (Steyerberg et al., 2008) argued that many attempts at developing prognostic models for TBI have fallen short due to small sample sizes, lack of external validation, and impractical presentation of the model for clinicians. To remedy these problems, they developed a prognostic model for TBI using a metaanalysis of 11 studies in the very large multi site IMPACT database using a sample of 8,509 patients (Maas et al. 2007). These studies included three observational studies and eight randomized control trials conducted between 1984 and 1997. They included pat ients aged 14 years and older with moderate to severe TBI as defined by a GCS less than or equal to 12. In order to externally validate their model, they used another large database of 6,681 TBI patients from the Medical Research Council Corticosteroid Ran domization after Significant Head Injury trial (MRC CRASH trial collaborators, 2008) who participated between 1999 and 2004. This was a large international trial of the administration of methylprednisolone after head injury. To address the potential confound of the administration of the drug, secondary analyses were performed with only placebo patients. Twenty six potential predictors including demographic, injury severity, CT, biochemical, and secondary insult variables were tested. They found that age, GC S motor score, and pupil response were the most predictive of general outcome at the six month time point (Steyerberg et al., 2008). These three predictors constituted

PAGE 14

14 their Core model which is consistent with many previous models for TBI (Jennett et al 1976; Perel et al 2006). An Extended model was also developed that included the three predictors from the core model, but added information on secondary insults (e.g., hypoxia, hypotension) as well as CT characteristics. This Extended model provided additional prognostic information to the Core model (see Table 11 ). With this extensive database, the IMPACT model for TBI addressed many of the limitations of previous models to predict mortality at six months post injury. Table 1 1 Impact m odels Core Model E xtended model Age Core Model GCS motor score Pupil response CT characteristics Marshall Classification Traumatic Subarachnoid hemorrhage (tSAH) Epidural hematoma (EH) Secondary Injury Hypotension Hypoxia Predictors Used in Prognos tic Models Each of the individual predictors used in the IMPACT prediction formula has been extensively examined for its utility in the prediction of outcome after severe TBI. Glasgow Coma Scale (GCS) The GCS is comprised of three scales including eye op ening, verbal response and motor response. These three values can be considered separately, but most frequently they are summed to create a single score. The lowest possible GCS is 3 (deep coma or death), and the highest is 15 (fully conscious; Teasdale and Jennett, 1974). Since its introduction in 1974, the GCS has become one of the most widely used

PAGE 15

15 tools for assessing brain injury severity because it can be quickly performed, it is repeatable, and it serves as an effective means to communicate brain injur y information with other members of the treatment team in a standardized way (Jennett 2002; McNett 2007). Emergency personnel frequently use the GCS score with head injured patients to make clinical decisions in the very acute phases (i.e. appropriate faci lities to take a patient, appropriate treatment protocols; Jennett, 2002). The developers of the scale argue that the use of the GCS in frequent monitoring of brain injured patients has led to better outcomes in many cases because it enables clinicians to detect deterioration earlier so they can intervene more quickly (Jennett 2002). These features of the GCS have lead researchers to study the GCS extensively for its validity in predicting functional outcome and mortality (McNett, 2007). Studies have shown that the GCS in combination with other variables can be useful in predicting outcome after TBI when broad outcome categories are used. The GCS has served its intended purpose of providing a quick and easy method for assessing TBI severity. Nevertheless, th ere are some challenges in using GCS as a predictor. Researchers have often used different GCS assessment time points for their prognostic models. GCS scores have been used from admission to the first hospital, some admission to the study hospital, or a co mbination of these time points (Marmarou et al., 1999; Murray et al., 1999). In fact, patients can vary significantly across early time points due to a variety of confounding variables (McNett, 2007). When obtained within the first few hours after injury, GCS scores can be influenced by the presence of drugs or alcohol. Later scores can be complicated by sedation, intubation or tracheostomy. The developers of this scale assert that the most valid time to measure GCS is after

PAGE 16

16 resuscitation (Jennett, 2002), and of course prognostic variables are most valuable when they can be accurately obtained early in the intervention process. One drawback the IMPACT is that the studies drawn from to create the database ranged from within 4 hours of injury to 3 days post injury in the time point the GCS was administered (Marmarou et al., 2007). They attempted to obtain time points as close as possible to post resuscitation GCS, but at the same time efforts were made to maximize the number of patients in the database (Marmarou et al., 2007). Prognostic models are most effective when the predictors are gathered in consistent ways at consistent time points. Age A patients age is another variable commonly used in the prediction of outcome after severe TBI. Increasing age has been associated with both a higher likelihood of mortality and a poorer general outcome from severe TBI (Howard et al,. 1989, Mosenthal et al., 2002). Howard and colleagues (1989) noted that despite comparable GCS scores, outcome was worse in the elderly, suggesting that GCS scores can underestimate the severity of brain injury in the elderly. The reasons for these worse outcomes remain unclear; however, several studies have suggested differences in the mechanism of injury, therapeutic interventions, complication rate, and premorbid disease (Hukkelhoven et al., 2003). Hukkelhoven and colleagues (2003) hypothesized that the brain has a lower capacity to heal as it ages resulting in a decreasing number of functioning neurons. Additionally, older individuals may experience a cumulative effect of minor insults to the brain as age increases (Hukkelhoven, 2003). These mechanisms suggest that the elderly may experience more significant loss of cognitive functioning after TBI. Age is therefore an important variable to include in a prognostic model as it is

PAGE 17

17 likely to contribute significantly to cognitive outcome and it has demonstrated utility in addition to the GCS score. Pupil Response Another acute clinical variable that has been commonly used in prognostic models is pupil response upon admission. Pupil response is fundamental to neurological assessment because poor reactivity or dilation of the pupils may be related to compression of the oculomotor nerve and indicate brainstem dysfunction (Marmarou et al., 2007). Vari ation in pupil response may be indicative of different underlying causes such as uncal herniation or reduced blood flow to the brain stem (Ropper et al., 1991; Ritter et al., 1999). Damage to the brain stem can quickly become lethal as it can result in pro blems with autoregulation or changes in blood pressure (Ritter et al., 1999). Several studies have shown a strong relationship between absence of pupil response and poor outcome after TBI (Andrews et al., 2002; Marmarou et al., 2007). Additionally, the com bination of fixed dilated pupils with low GCS scores has been shown to be a powerful predictor of mortality (Jain et al., 2008). Thus, pupil response has been shown to be a robust prognostic indicator and is widely accepted as an important clinical variabl e to include in outcome prediction models for severe TBI. Structural Neuroimaging As highlighted in the IMPACT model there are additional prognostic variables that can strengthen the Core model including information regarding structural neuroimaging. Computed tomography (CT) of the brain has commonly been used as an additional acute indicator of brain injury characteristics. CT imaging is used in the acute stage of TBI to detect gross structural brain abnormalities and identify the need for medical or surg ical intervention, because it can be done quickly and inexpensively

PAGE 18

18 (Haydel et al., 2000). Despite its lower resolution, CT is preferable to magnetic resonance (MR) imaging for acute TBI patients, because unlike MR, it can even be done with common metal based life support equipment in place (Bigler 2005). Head CT has proved to be an excellent clinical tool in detecting some lesions, hemorrhages, skull fractures, and swelling (Bigler 2005). Marshall et al. (1992) created a CT rating scale that provides a bas is for evaluating the severity of injury during the acute stage. This classification system identifies 6 categories describing the extent and nature of the brain injury based on CT scan abnormalities. The Marshall CT classification system has become widely accepted for descriptive purposes and has been shown to be a strong predictor of general outcome in TBI (Hukkelhoven et al., 2005; Lobato et al., 1983; Servadei et al., 2000). O ne drawback to this system is that although it distinguishes between the pres ence and absence of brain hemorrhage requiring surgical intervention from those that do not, it does not describe critical information about the location of bleeding. More specifically it is understood that bleeding occurring above the dura has different p rognostic implications compared to bleeding below the dura (Maas et al., 2005). Maas and colleagues (2005) found that the inclusion of additional CT characteristics in a model can add prognostic value to the Marshall classification alone. For example, the presence of traumatic subarachnoid hemorrhages (tSAH) and epidural hematomas (EH) has been shown to be a strong predictor of outcome and mortality in TBI and have been recommended for inclusion in prognostic models of severe TBI (Maas et al., 2005; Servade i et al., 2002). Despite the utility of CT in identifying gross structural pathology in the acute stage, these findings have not been shown to relate well to cognitive

PAGE 19

19 outcome at the time of discharge from rehabilitation, which has made the accurate predic tion of outcome from acute CT findings alone difficult (Dikmen et al. 2001; Temkin et al. 2003). While CT findings cannot alone predict cognitive outcome after severe TBI, they do make a significant contribution to a prognostic model. Secondary injury Dama ge to the brain sustained from a severe TBI is classified as primary if it is related to the initial trauma and secondary if it is related to any events that occur afterward. Following the primary injuries, the brain becomes more vulnerable to secondary damage (McHugh et al., 2007). Of the types of secondary injury that have been studies thus far, hypoxia and hypotension have proven most valuable predictors of outcome in prognostic models of severe TBI (Andrews et al., 2002). Associations have been found fo r both of these two markers of secondary insults and adverse outcome after TBI (Manley et al., 2001; Walia and Sutcliffe, 2002; Chesnut, 1993). Both of these injury markers are relatively common on admission and their combined effects are substantially greater than the sum of the individual effects (McHugh et al., 2007). Therefore, these indicators of secondary injury will likely contribute to the prognostic value to a model of cognitive outcome after severe TBI. Measurement of Outcome after Traumatic Brain Injury Extensive research has been conducted on ways to measure and combine various acute physiological variables in severe TBI to create prognostic models with great statistical power. Research on outcome measurement, on the other hand, has lagged behind. The ideal measure of TBI outcome would be easily administered, ecologically valid, reliable, and sensitive to a wide range of outcomes. Unfortunately,

PAGE 20

20 such a measure does not yet exist. Each of the existing measures that are commonly used in TBI outcome research comes with its own pros and cons. Many studies simply ask whether patients returned to work after their injury (Wehman et al., 2005). Although this outcome is certainly ecologically valid and can be easily assessed, it is wrought with confounds t hat may call its validity into question such as a patients premorbid employment or the role of other physical injuries unrelated to the TBI. Scales have been created that classify TBI outcome into global categories of functioning such as the Glasgow Outcome Scale (Jennett and Teasdale, 1975). These measures tend to be easy to administer but often suffer from poor sensitivity and inter rater reliability (McCauley et al., 2001). Efforts have been made to create outcome measures that remove these flaws; however, there is often a tradeoff such that when a test becomes more sensitive and reliable, it becomes more difficult to administer. Neuropsychological tests provide a means of measuring cognitive functioning after TBI with excellent psychometric properties and high sensitivity to subtle differences. However, these tests often require a lot of time to administer and some must be administered and interpreted by trained professionals. The following sections outline some specific examples of measures used in TB I outcome research and their strengths and weaknesses. Global Functional Outcome As described earlier, most of the prognostic models developed thus far have predicted recovery in a general way using crude variables such as mortality or global outcome as measured by the Glasgow Outcome Scale (GOS, Jennett and Bond 1975). The GOS has long been a widely used measure of both short and long term recovery from TBI that provides a quick assessment of functional outcomes with clinical

PAGE 21

21 relevance (Satz 1998). This s cale categorizes outcome into 5 general categories: (1) D eath; (2) P ersistent Vegetative State; (3) Severe Disability ; (4) Moderate Disability ; and (5) Good Recovery (Jennett and Bond 1975). The scale is not intended to provide detailed information about t he specific difficulties faced by individual patients, but to give a general index of overall outcome (Marshall, 1992). It has been recommended for use in clinical trials as it was intended to compare patients in a simple easily interpreted way (Clifton e t al., 1993). GOS has traditionally been obtained through a short unstructured interview which has led to poor inter rater reliability (Anderson et al., 1993). Different versions of this scale have been developed to address some of its shortcomings by incl uding a structured interview to increase reliability and expanding the number of categories to increase sensitivity (Wilson et al., 1998). For the purpose of prognostic model development, the GOS is often collapsed further into fewer categories. In TBI outcome prediction studies the scale has been traditionally dichotomized into Unfavorable outcome ( Death, Vegetative State or Severe Disability ) and Favorable outcome ( Moderate Disability or Good Recovery ; McHugh et al., 2007). These levels of outcome are generally adequate for studies in acute care settings where doctors early focus is primarily on mortality and morbidity (Ashley & Ashley, 2004). However, once patients have stabilized, more sensitive measures of outcome become necessary to assess the full r ange of outcome possible after TBI (Clifton, 1993). Although the GOS is widely accepted as a standard way of classifying recovery after TBI some have argued that this scale may sacrifice detail for simplicity in its broad classifications of outcome (Clifton, 1993). This scale has been shown to be insensitive

PAGE 22

22 to clinically significant changes in patient states (Hall, 1985; Struchen et al., 2001). Because of this insensitivity it has proven difficult to document patient recovery over time using this scale (Cl ifton, 1993). Moreover, many patients with good GOS scores have been shown to have significant functional deficits as measured by more specific neurobehavioral tests (Satz, 1998). In fact, the GOS has been criticized for emphasizing the physical symptoms a nd while disregarding mental changes that are often more important in determining disability after TBI (Anderson et al., 1993). The developers of this scale contend that it serves its intended purpose of facilitating inter center communication and comparis on regarding TBI outcome (Jennett and Bond 1975). However, Wilson and colleagues (1998) noted that other more detailed measures of impairment and disability should supplement the GOS when used as a primary outcome measure. The Disability Rating Scale (DR S) is another accepted measure of global outcome after TBI that may be more sensitive to changes in functioning (Rappaport et al., 1982). The DRS assesses more detailed levels of feeding, grooming, and toileting abilities and includes related issues such as dependence on others and psychosocial adaptability (Rappaport et al., 1982). The DRS can thereby characterize patients on a full range of outcomes from coma to reentry into the community. The DRS has demonstrated excellent inter rater reliability, and t est retest reliability (Novack, 1991; Gouvier, 1987). Furthermore, the predictive validity of the DRS has been demonstrated through correlations with length of hospital stay, disposition at discharge, and DRS one year later (Eliason and Topp, 1984; Rappapo rt et al., 1982). Additionally, the DRS has been shown to have ecological validity as it was able to predict return to competitive

PAGE 23

23 employment (Cope et al., 1991). Overall, the DRS has proven to be a sensitive, functional, reliable, and ecologically valid m eans of monitoring patients with traumatic head injury during the course of their recovery. Cognitive Outcome The GOS and DRS provide adequate estimates of general outcome after TBI that have demonstrated ecological validity. However, other types of outcom e measurement would be helpful in discharge planning and determining on whom to target limited rehabilitation resources (Clifton, 1993). Some researchers have argued that due to the complexity of recovery from TBI, multiple dimensions of outcome should be explored. Novack et al. (2001) proposed a theoretical framework for the nature of outcome after TBI that includes cognitive, emotional, and functional status. In this study the relationships among the multiple components of outcome were evaluated, and it w as found that injury characteristics could not only be used to predict functional skills but also cognitive status. Additionally, cognitive status was found to have a particularly strong relationship with measures designed to capture overall functional outcome, suggesting that cognitive abilities play a significant role in recovery (Novack, 2001). Further, Rassovsky and colleagues (2006) found that TBI severity exerted an indirect rather than direct impact on functional outcome. This study explored potenti al mediators of the relationship between TBI severity and overall outcome, and they determined that cognitive functioning was the most significant mediator (Rassovsky 2006). This research supports the role of neuropsychological testing in TBI outcome research. Such testing can contribute a reliable forecast of functional ability to both the family and rehabilitation team.

PAGE 24

24 Specifically, neuropsychological measures have provided a valuable characterization of the nature of common cognitive impairments after TBI. Most commonly these patients have demonstrated deficits in attention, reduced capacity for new learning, and slowed information processing (Dikmen, 1993 ; Millis, 2001). Additionally, deficits in motor skills, executive functions, and overall intellect ual skills have been identified (Dikmen, 1993 ). The extent of these impairments has been found to be related to the severity of the TBI (Millis, 2001). However, these studies have commonly relied on crude measures of injury severity. Additionally, these st udies revealed great variability of performance within severity levels. Dikmen et al. (1993) argued that this heterogeneity may be taken as a sign of failure to appropriately classify severity. Studies that have specifically examined cognitive functioning have not utilized the full range of injury characteristics. More detailed assessment of injury characteristics used in prediction may help to understand the variability seen in neuropsychological outcome among patients with severe TBI. Neuropsychological batteries can provide an in depth look at the nature of cognitive deficits, but they require a significant investment of time and energy from both the patients and examiners. Because of these heavy demands on patients time and effort, following them long t erm becomes more difficult even when compensation is provided and extensive efforts are made to make participation more convenient for patients. Longitudinal studies that use longterm cognitive outcome measures have notoriously high attrition rates. Studi es have reported having almost 50% attrition rate for six month cognitive outcome, which is considerably higher than is most often achieved with noncognitive outcome measures (Struchen et al., 2001 ). Additionally, a

PAGE 25

25 significant portion of survivors of TBI remain to disabled to complete a full traditional battery of neuropsychological tests (Clifton 1993). Thus, there are considerable obstacles that of prevent the use neuropsychological tests as primary outcome measures in TBI recovery research. Researchers have sought to remedy this problem by reducing the size of the neuropsychological battery and selecting tests that give the most bang for your buck (Clifton et al., 1993; Satz et al., 1998). The optimal test of cognitive functioning would be brief and eas ily administered so that a nonneuropsychologist could give it at bedside or even over the phone. Ideally, this test would also minimize physical confounds related to motor or visual impairments. Additionally, the test would be sensitive to the full range of outcomes that can occur after TBI and not be subject to floor effects. With hundreds of cognitive assessment tools to choose from there are some tests that come closer to reaching these criteria than others. As described above the areas of common defici ts after TBI include information processing speed, attention, memory, and executive functioning. Tests of processing speed typically have a heavy motor component that may introduce a confounding factor for TBI patients with impaired motor functioning in li mbs. Furthermore, many tests of executive functioning are difficult to perform and are subject to floor effects in a severe TBI population. A simple test of attention and working memory such as Digit Span may be the best candidate to fit the above criteri a as it is easy to administer, sufficiently sensitive, and has no motor or visual component. Digit Span tests consist of both forward and backward conditions, each of which involves slightly different mental activities. Forward span tasks are typically con sidered measures of attention and immediate recall (Lezak

PAGE 26

26 et al., 2004). Forward digit repetition is often impaired in the first months following head trauma, but over time performance returns to normal (Lezak, 1979). Backward span tasks, with the added requirement of mentally reversing the numbers, are regarded as measures of working memory (Lezak et al., 2004) The backward condition has been found to be very vulnerable to diffuse damage and has been shown to be impaired following TBI (Fork et al., 2005). Both forward and backward digit spans measure the adverse effects of brain injury on cognitive abilities of attention and working memory. Given its ability to measure these deficits, previous research has classified the Digit Span task among psychological assessment tools with demonstrated robustness in the prediction of everyday functioning (Tupper and Cicerone, 1990). However, it has been argued that level of education should be taken into account when using this outcome, as it has been demonstrated that education is a strong predictor of both forward and backward digit span (Ostrosky Nevertheless, t he addition of a cognitive measure of outcome such as Digit Span could provide a powerful supplement to the previously described global measures that are ecologically valid.

PAGE 27

27 CHAPTER 2 OVERVIEW AND STATEMENT OF THE PROBLEM Introduction TBI research has long sought to create a means to accurately predict long term outcome. The review of the literature presented above suggests that the I MPACT model has proven to be useful in predicting mortality or crude general outcome after severe TBI (Steyerberg et al., 2008). However, this model has not demonstrated the ability to predict more detailed outcomes. Part of the challenge of this process i s the choice of outcome measures. The state of measurement development leaves us with a tradeoff between ecological validity and sound psychometric properties. The overall goal of this study is to replicate the main findings of the IMPACT model and to extend these findings by using measures that will provide more detailed characterization of outcomes. Using the International Mission for Prognosis and Analysis of Clinical Trials (IMPACT) Study Model to Predict Outcome The creators of the IMPACT model have developed a scoring system that allows clinicians to calculate the probability of prognosis for a particular patient. To create this scoring system, the predictor variable from the Core and Extended IMPACT models were used in proportional odds logistic regr ession analysis with 6 month GOS as the outcome measure. The regression coefficients from these analyses were rounded and scaled to create a score chart for each of the models (see Figure 2 1)

PAGE 28

28 Figure 2 1. IMPACT model score chart for 6 month outcome afte r TBI Characteristics Value Score (b ) Age (years) 0 30 39 1 40 49 2 50 59 3 60 69 4 70 + 5 GCS Motor Score None/extension 6 Abnormal flexion 4 Normal flexion 2 Localizes/obeys 0 Pupil Response Both pupils reacted 0 One pup il reacted 2 No pupil reacted 4 Total Score Core Model _________ Hypoxia Yes or suspected 1 No 0 Hypotension Yes or suspected 2 No 0 CT classification I 2 II 0 III/IV 2 V/VI 2 tSAH Yes 2 No 0 EH Yes 2 No 0 Extended Mod el Sub total _________ Total Score Extended Model _________ The IMPACT model score chart is used to calculate the probability of an Unfavorable outcome at six months post injury and is derived from a logistic regression equation in the form of: P( Y ) = (1 /(1+e-LP). In this case P( Y ) is the probability of an Unfavorable outcome at six months post injury and e is the base of natural logarithms. LP refers to the linear predictor in a logistic regression model in the form of : LP = b0 + b1X1+ b2 X2 + ... + bn The constant ( b0) or intercept of the line is based on previously published calculation of proportion of Unfavorable outcomes in severe TBI, and ( Xn) represent the series of predictor variables with coefficients (or weights) attached to each predictor ( bn) This formula shows that the predictors (X1Xn) are on a scale related to

PAGE 29

29 the probability of Unfavorable outcome with relative weights ( b1b2) derived from the regression coefficients rather than odds ratios (Moon et al., 2002). This means that for a pati ent with a particular profile, the regression coefficients bn are multiplied by each of the patients corresponding Xn values which can then simply be added together creating the sum scores. The resulting equations become: LPcore = 1.62 + 0.299 x Sum score of Core model LPextended = 2.10 + 0.276 x (Sum score of Core model + subscore CT) The resulting value can then be logistically transformed according to equation one to calculate that patients probability of Unfavorable outcome P( Y ). The regression co efficients have been scaled and rounded to the nearest integer so that scores for each predictor can be easily obtained and used for both the Core and Extended models. In this way relative risks of the individual predictor variables can be combined into a single prognostic variable that is indicative of the risk of an Unfavorable outcome at six months post injury. By entering a patients raw prediction data into this IMPACT model score chart and then using the resulting Total scores in these equations, one can establish the probability of Unfavorable outcome at six months. The discriminatory ability of the core model had an area under the receiver operating characteristic curve (AUC) between .66 and .84 at cross validation with the CRASH database. The extended model improved the AUC by .05 suggesting these models were both highly effective in predicting general outcome categories. This finding could suggest that there is little added value to including the additional information

PAGE 30

30 contained in the Extended model. However, this has not been established for other measures of outcome. Unfortunately, the studies that were merged to form the IMPACT database lacked uniformity in data collection for some of the measures. The GCS scores used in the studies were administ ered at time points that ranged from within 4 to 72 hours post injury, but when possible the post stabilization GCS was used (Marmarou et al., 2007). Pupil response was likewise measured at various times which coincided with administration of the GCS. The same time points were used for both. Information regarding preenrollment hypotension and hypoxia was gathered for 9 of the studies. Hypotension was defined as systolic blood pressure < 90 mmHg and hypoxia was defined as partial pressure of oxygen (pO2) < 60 mmHg. Head CT scans obtained on admission that provided Marshall classifications as well as information regarding tSAH and EDH were obtained in seven of the studies. Although some discrepancies exist from merging data from multiple studies, the data in the IMPACT database underwent rigorous quality control measures and is reported to be at a very high standard (Marmarou et al., 2007 ). The IMPACT model from Steyerberg et al., (2008) was selected for this study because it was based on a large sample, exter nally validated, and had sound methodology. Aim 1 The first aim of the current study was to replicate the Core and Extended IMPACT models using the same global outcome classifications more specifically mortality and dichotomous GOS scores classified as Fav orable ( Moderate Disability or Good Recovery ) or Unfavorable ( Death Vegetative State or Severe Disability ). It was

PAGE 31

31 hypothesized that the results would be similar those reported in Steyerberg et al., 2008. Specifically, it was predict ed that: Hypothesis 1a: The Core model would have statistically significant associations with the 6 month mortality and dichotomous GOS outcome, similar to those observed in Steyerberg et al., 2008. Hypothesis 1b: The addition of the Extended model sum score would contribute significantly to the core model and predict more of the variance in the 6 month mortality as well as in the dichotomous GOS. Aim 2 Aim 2 sought to extend the findings of the IMPACT model and test its ability to measure a wider range of outcomes that have demonstrated ecological validity. To achieve this goal, the full range of the GOS (as opposed to the simple Favorable/ Unfavorable dichotomy) and the DRS were used as outcome measures. It was hypothesized that: Hypothesis 2a: The IMPACT model would have a statistically significant association with the 6 month GOS such that poorer GOS scores would be related to a lower model Total score. Hypothesis 2b: The IMPACT model would have a statistically significant association with the 6 month DRS such that poorer DRS scores would be related to a lower model Total score. Hypothesis 2c: The IMPACT model would predict outcome on the DRS with more accuracy than it would on the full range of the GOS because of the level of detail and continuous range of the DRS. Aim 3 Aim 3 sought to extend the IMPACT model by testing its ability to measure cognitive outcome in the domain of working memory at six months post injury. To achieve this goal, Digit Span would be used as the outcome measure and analyses would be controlled for level of education. It was hypothesized that:

PAGE 32

32 Hypothesis 3: The IMPACT model would have a statistically significant association with the 6 month Digit Span score after controlling for level of education such that poorer Digit Span total scores would be related to a lower model Total score.

PAGE 33

33 CHAPTER 3 METHODS General Methods The current study utilized data from several studies, both ongoing and archival, across two research sites (University of Florida and University of Houston). The ongoing multi si te study (PI: Robiscek, University of Florida) funded by the National Institute of Health (NIH) seeks to systematically identify and validate a panel of biochemical markers for TBI (N= 79) It has been reviewed and approved by the Institutional Review Boar ds at both institutions. Additionally, archival data was used from several NIH funded studies conducted by Dr. Hannay, PhD at the University of Houston (N= 467) Participants A total of 546 participants were recruited from two Level 1 trauma centers: Ben Taub Ge neral Hospital in Houston Texas (N=503) and Shands at the University of Florida in Gainesville, Florida (N=43) Participants were patients who had sustained severe closed head injuries identified in the emergency departments or intensive care units of these hospitals. Patients who were or may have been pregnant were excluded from the studies Patients were also excluded if they had an autoimmune disease, life threatening injuries in any organ system other than the head or spine, or a severe preexist ing chronic disease. Additionally, the studies excluded patients with a history of severe psychiatric conditions (e.g. psychotic disorders or bipolar disorder); however, patients with less severe conditions (e.g. depression or anxiety disorders) were not e xcluded. For the archival studies conducted at the University of Houston, patients 1 4 years old and older while the ongoing multisite biomarker study excluded patients under 18 years old. Patients were included if they had a severe TBI as defined by an

PAGE 34

34 adm ission GCS of 8 or less. Additionally, in the ongoing multisite study of biomarkers, participants were included if they a ventriculostomy placed as part of their routine care, and had neuromonitoring probes for intracranial pressure. Procedure Consent for patient participation in the study was initially obtained from a legally authorized representative (LAR; typically a significant other or relative) as soon as possible upon the patients arrival at the hospital. Consent was also obtained from someone close to the patient to participate in the study as a collateral source of information regarding the patients background and pre injury functioning as well as their functioning later in the recovery process. In this longitudinal study patients abilities to co nsent for themselves were continually assessed based on the Galveston Orientation and Amnesia Test ( GOAT; Levin et al., 1979). If patients were deemed capable of providing consent (based upon GOAT scores in the normal range 76100) they were asked to do so. Measurement of Physiological Predictor Variables Methods for the collection of raw physiological data are presented below. This raw data was processed according to methods described following the explanations of how each variable was obtained. Admission Glasgow Coma Scale (GCS) M otor S cale For patient data collected at the University of Houston (N=503) the GCS was obtained post resuscitation of any hypotension or hypoxia but before surgery or other interventions. For patient data collected at the Unive rsity of Florida (N=43) the post resuscitation GCS was ascertained through retrospective review of charts approximating as closely as possible the timing of the Houston data. For the purposes

PAGE 35

35 of the IMPACT model used in this study only the motor scale was used consisting of 6 levels: (1) No response; (2) Abnormal ExtensionDecerebrate Posture;(3) Abnormal FlexionDecorticate Posture; (4) Withdraws from pain; (5) Localizes pain; and (6) Follows commands. The GCS exam was administered by a research nurse and observed by a research assistant. The exams were administered in a uniform way according to national guidelines for management of severe TBI (Badjatia et al., 2007). Regardless of the apparent neurological status of the patient, verbal directions were give n first. If a patient was nonresponsive and unable to follow simple commands, a painful stimulus (pinch to inner arm or thigh, or nail bed pressure) was applied to induce a motor response in each limb not known to have paralysis Pupil Response Pupil exam was performed post resuscitation. This was typically done by a physician or nurse as part of routine clinical exam by momentarily holding a flashlight over each eye individually. For the purposes of this study, constriction of pupil was recorded as present for both eyes, one eye, or neither eye. Head CT Classification Head CT scans taken on admission to the hospital as part of the patients regular care were used in this study. All CT scans were read by a clinicians trained in neuroradiology. The Marshall C T classification system was used to classify the findings of the CT (see Table 3 1 ). Additionally, CT scans were examined for the presence of tSAH and EH. The characteristics of EH and tSAH were scored as present or absent, without further differentiation.

PAGE 36

36 Table 3 1 Marshall computed tomographic c lassification Category Definition I. Diffuse Injury I (no visible pathology) No visible intracranial pathology seen on CT scan II. Diffuse Injury II Cisterns are present with midline shift of 0 5 mm and/or l esion densities present; no high or mixed density lesion III. Diffuse Injury III (swelling) Cisterns compressed or absent (swelling) with midline shift 0 5 mm, no high or mixed lesion > 25 cc IV. Diffuse Injury IV (Shift) Midline shift > 5 mm, no high ( shift) or mixed density lesion > 25 cc V. Evacuated mass lesion Any lesion surgically evacuated VI. Non evacuated mass lesion High or mixed density lesion > 25 cc, not surgically evacuated Hypoxia Hypoxia was defined as any documented occurrence of a partial pressure of oxygen (pO2) < 60 mmHg from time of injury until successful resuscitation. This was recorded as present or absent for the purposes of this study Hypotension The admission blood pressures were taken up on admission to the hospital and assessed for the presence of hypotension. Hypotension was defined as systolic blood pressure below 90 mm Hg (Butcher et al 2007). This was recorded as present or absent for the purposes of this study IMPACT M odel S um S cores As indicated above, the variables c ontained in the IMPACT Core and Extended models were used in proportional odds logistic regression analyses with the full range of GOS at 6 month post injury as the outcome measure. The regression coefficients from these analyses were rounded and scaled to create a weight for each of the variables in the model. These weighted scores (see Figure 3 1 ) were then added to yield a sum

PAGE 37

37 score for the Core model and Extended model which were then added to yield an overall sum score for the model. Figure 3 1 IMPAC T model score chart for 6 month outcome after TBI Characteristics Value Score (b ) Age (years) 0 30 39 1 40 49 2 50 59 3 60 69 4 70 + 5 GCS Motor Score None/extension 6 Abnormal flexion 4 Normal flexion 2 Localizes/obeys 0 P upil Response Both pupils reacted 0 One pupil reacted 2 No pupil reacted 4 Total Score Core Model _________ Hypoxia Yes or suspected 1 No 0 Hypotension Yes or suspected 2 No 0 CT classification I 2 II 0 III/IV 2 V/VI 2 tSAH Yes 2 No 0 EH Yes 2 No 0 Extended Model Subtotal _________ Total Score Extended Model _________ For the current study, sum scores for the model were calculated from each participants acute data. These weighted and summed scores were used as the independent variables in the analyses to follow rather than the raw values of the individual predictors. The regression coefficients used to develop the weights in the score chart were calculated based on the outcome measure of the GOS. These same weights will be appl ied to the raw predictors for the analyses of the other outcome measures of DRS and Digit Span. A new regression equation was developed for each

PAGE 38

38 of the outcome measures tested. These analyses yield ed R2 values that told how much of the variance in the outc ome measure was explained by the Sum scores from the indicate how much an outcome score would change for each point increase in the IMPACT model Sum score. Ultimately, if these analyses were significant a clinician would be able to plug a patients raw predictor values into the IMPACT score chart to yield a Total score that could then be put into a simple regression formula to calculate an estimated level of outcome as measur ed by the full GOS, DRS, or Digit Span. The weights used in these analyses were applied to outcomes other than the one they were based on. Although several weaknesses to the this method have been cited (Rubin, et al., 1992), positive results obtained from hundreds of studies using this method suggest that it is robust to these limitations, much in the same way statistical methods can be robust to violations of some assumptions. This methodology was justified because this study represents an attempt to take a powerful model based on a large sample and examine whether the same well validated model can be extended to predict different types of outcome with more specificity including measures with good ecological validity such as the DRS and measures with sound psychometric properties such as the Digit Span test. These methods have been used in similar studies (Charlson et al., 1994 ). Additionally, it should be noted that the sample used in the analysis of Digit Span contained only those p atients who survived si x months, were functionally able to complete the Digit Span subtest and had available data on their level of education (N=76) This analysis again use d the weights calculated from the IMPACT model

PAGE 39

39 regression coefficients despite the fact that the samples differ ed in terms of the range of outcome included. This is justified because the sample used in the analysis for the current study represents a subset of the overall sample used in the IMPACT study. Therefore the prediction model should still apply to this truncated portion of the sample. These methods of applying a model to a subset of the original sample have also been used in other similar studies (Hemmelgarn et al., 2003; Sarbjit et al., 2005). Measurement of Global and Cognitive Outcome Outcome was ass essed at 6 months post injury using the GOS, DRS and the Digit Span test. These different outcome measures were used to address different aims and hypotheses. Information to complete these measures was obtained from the patients and their families. Researc h assistants who obtained this information were extensively trained. The quality of the data collected was monitored in weekly sessions to determine the adequacy and reliability of the responses on these measures. If the research group did not agree that enough information had been obtained to accurately determine the DRS or GOS scores, additional data were obtained until a consensus was reached. The research assistants were in contact with the patients and their families when they were admitted to the study and maintained regular contact with the patients through the first six months post injury. Additionally, a battery of neuropsychological tests including the Digit Span test (WAIS III; Wechsler., 1997) was administered at six months post injury to those patients who were functioning at a level that permitted testing. Some patients completed the battery in an outpatient visit to the study hospital. If patients were unable to travel to the research hospital, efforts were made to travel to the patient. Other s completed the tests as inpatients in a hospital or rehabilitation facility.

PAGE 40

40 Global Functional Outcome Measures Glasgow Outcome Scale (GOS) The GOS is a five point scale assesses outcome us ing 5 broad categories: (1) Death; (2) Vegetative State; (3) Seve re Disability ; (4) Moderate Disability ; and (5) Good Recovery (Jennett and Bond 1975). The scale was developed as a means of standardizing outcome classifications across centers. This instrument is one of the most commonly used outcome measures utilized in neurosurgical studies of head injury outcome. For Aim 1 these scores will be grouped as Favorable (Moderate Disability and Good Recovery ) and Unfavorable outcome ( Death, Vegetative State, and Severe Disability ) according to the methods of the IMPACT study For Aim 2, all five categories were used for the analyses. Disability Rating Scale (DRS) The DRS is a 30 point scale (0 = no problems; 30 = dead) that is based on the sum of ratings on eight items that assess four areas: awareness/arousal, cognitive abi lity for self care, level of physical dependence on others, and estimated ability for work, school, or homemaking activities (Rappaport et al., 1982). This instrument was developed to assess disability of severe head trauma patients from coma through diffe rent levels of awareness and functioning to community reentry. The raw score on this measure were used for the analyses in Aim 2. Cognitive Outcome Measure: Digit Span The Digit Span test from the Wechsler Adult Intelligence Scale Third Edition (WAIS III; Wechsler et al., 1997) is made up of two different tasks: Digits Forward and Digits Backward. For these tests the examiner reads random number sequences of increasing length aloud at the rate of one number per second. For the Digits Forward

PAGE 41

41 test, the pat ient was asked to repeat each sequence exactly as it was given. These sequences ranged in length from 2 to 9 digits. For the Digits Backward task, patients were asked to repeat the digits exactly in reverse order. These sequences ranged in length from 2 to 8 digits. For both tasks, two trials are given of each span length and the subject received a raw score point for each correct trial. T esting is discontinued when the patient makes errors on both trials of a particular span length. Although these tasks m easure slightly different mental activities, they both correspond to the cognitive domain of working memory. This study is less concerned with the specificity of the underlying cognitive construct and more concerned with the sensitivity of the measure to brain injury. Therefore, the raw scores from each condition were combined to from one total score for this studys analysis to maximize the range of outcomes that will be achieved and minimize the floor effects that might have resulted from using just the backward span. Analyses Utilized Aim 1 Replication of IMPACT C ore and E xtended M odels P redicting M ortality and Unfavorable O utcome on GOS The first set of analyses will use mean, standard deviation, median and frequencies to describe data distribution. Univ ariate and multivariable forced entry logistic regression analysis will be used to examine the relative contributions of the Core and Extended IMPACT models to outcome prediction. The Total score of the Core model were used in the first block and the Total score of the Extended model were added in the second block with the scores as continuous variables and mortality at 6 months as the dichotomous outcome. Additionally, similar analyses were run with a dichotomized version of the GOS Unfavorable (Death, Veg etative State or Severe

PAGE 42

42 Disability ) versus Favorable ( Moderate Disability to Good Recovery ) outcome at six months post injury as the dichotomous variable and the core and extended model scores again serving as continuous variables. The odds ratio (OR) and 95% confidence intervals ( CI) were also calculated. One tailed p values less than 0.05 were considered to be statistically significant. Aim 2 Extension of IMPACT M odel P redicting F ull GOS and DRS at S ix M onths This set of analyses examined the extended I MPACT model using the full range of the GOS at six months post injury. The GOS is an ordered categorical variable and it can not be justified to treat it as continuous. Therefore, these analyses used proportional odds ordinal logistic regression with the si x month GOS as an ordinal outcome. This again involved logistic regression with the Total score of the Core model in the first block and the Total score of the Extended model added in the second block as continuous variables. The odds ratios (OR) and 95% c onfidence intervals ( CI) were again calculated. Statistical analyses were performed running the SPSS 17 statistical package on a personal computer. Onetailed p values less than 0.05 were considered to be statistically significant. Nagelkerkes R2 was calculated as an approximation to the percentage of the variability in the GOS that can be explained by the different levels of the model. Nagelkerkes R2 was be used to measure the added predictive value of each level of the model. For the second part of this aim, a multiple regression analysis was used to determine if the IMPACT model could predict outcome on the continuous DRS measure at six months post injury. For this analysis, the Total score of the Core model was entered in the first block and the subsco re of the Extended model was added in the second block as continuous variables. The R2 was used to measure the amount of the

PAGE 43

43 variability in the DRS that can be explained by the model. One tailed p values less than 0.05 were considered to be statistically s ignificant. Finally, the R2 from this analysis were compared to the Nagelkerkes R2 to determine whether the DRS or GOS is better predicted by the models. Aim 3 Extension of IMPACT P redicting W orking M emory O utcome at S ix M onths Finally, this set of analys es use d the total Digit Span score as a continuous outcome measure to examine the proportion of the variance explained by the IMPACT model. This analysis used a multiple regression with the Core model score in the first block and the Extended model score a dded in the second block as continuous variables. Resulting R2 statistics were reported to measure the amount of the variability in the Digit Span total score that could be explained by the model. One tailed p values less than 0.05 were considered to be st atistically significant. All statistical analyses were performed running the SPSS 17 statistical package on a personal computer.

PAGE 44

44 CHAPTER 4 RESULTS Sample Characteristics A total of 546 patients were included in the present study. Demographic characteristi cs of the sample are presented in Table 4 1. Patients ages ranged from 1491 years, 83% were male, and the average education level was approximately 12 years. A higher proportion were Hispanic (42%) than White (36%), which is representative of the population demographics of the Houston area where much of the data was collected. Table 4 1 Demographic characteristics of the s ample (N = 546) Characteristic N (%) M SD Range Age 35.4 15.2 14 91 Education 11.8 2.6 3 20 Sex Male 453 (83.0) F emale 93 (17.0) Race Hispanic 228 (41.8) White 196 (35.9) Black 111 (20.3) Asian/PI 11 (2.0) M = Mean, SD = standard Deviation Table 4 2 presents the clinical characteristics of the current study sample as well as those i ncluded in the IMPACT study. Significant group differences were observed for several patient characteristics such that the current study sample generally showed more severe injury characteristics and had worse outcomes. The current study sample had more patients with no motor response (GCS=1) and fewer who demonstrated abnormal flexion (GCS=3 ) or withdrawal from pain (GCS=4 ) Additionally, the current study sample had more patients who exhibited higher levels of neurocognitive status (i.e., GCS=5 localizati on or GCS=6 obeying commands) than were present in the

PAGE 45

45 IMPACT study sample The present study s ample also exhibited more severe injuries as characterized by pupil reactivity, as there were more patients with neither pupil responding and fewer with both pup ils responding as compared to the IMPACT sample. Furthermore, the present study sample had more instances of hypoxia than the IMPACT sample. Regarding head CT characteristics, the present studys patients again showed more severe injuries as there were mor e instances of evacuated mass lesions (Marshall 5) and traumatic subarachnoid hemorrhages as compared to the IMPACT group with fewer patients who had no visible pathology (Marshall 1) or minor pathology (Marshall 2). In terms of outcome, the present study had a higher mortality rate (GOS=1), more patients rated as Severe Disability (GOS=3) and fewer patients rated as Good Recovery (GOS=5). In summary, the current study sample exhibited more severe injury characteristics and poorer resulting outcomes as compared to the IMPACT study.

PAGE 46

46 Table 4 2 Pati ent characteristics of present s ample and IMPACT study Characteristic Present Study IMPACT Study P (n = 546) (n = 8509) IMPACT Core Model predictors Age, years: Median (25 75 %ile) 35 (25 46) 30 (21 45) GCS Motor Score None (1) 130 (24%) 1395 (16%) <.001 Extension (2) 70 (13%) 1042 (12%) .692 Abnormal flexion (3) 37 (7%) 1085 (13%) <.001 Normal flexion (4) 72 (13%) 1940 (23%) <.001 Localizes/obeys (5/6) 227 (42%) 2591 (30%) <.001 Pu pillary reactivity Both pupils reacted 323 (59%) 4486 (63%) .003 One pupil reacted 54 (10%) 886 (12%) .698 No pupil reacted 169 (31%) 1754 (25%) <.001 IMPACT Extended Model predictors Hypoxia Yes or suspected 187 (34%) 1116 (20%) <.001 Hypotension Yes or suspected 86 (16%) 1171 (18%) .193 CT classification I 3 (1%) 360 (7%) <.001 II 154 (28%) 1838 (35%) <.001 III 120 (22%) 863 (17%) <.001 IV 9 (2%) 187 (4%) .393 V 231 (42%) 1435 (28%) <.001 VI 29 (5%) 509 ( 10%) .521 Subarachnoid hemorrhage (tSAH) Yes 386 (71%) 3313 (45%) <.001 Epidural hematoma Yes 73 (13%) 999 (13%) .253 The 6 month post injury outcomes for both samples are also presented in Table 4 3. In the current study sample, roughly two thirds survived six months post injury, however only 36% of those were classified as having a Favorable outcome. The average DRS score was 14.2, which would be classified as Severe Disability (Rappaport, 1982). The standard deviation for this scale was 12.5 indicating a high level of variability in outcome in the current study sample.

PAGE 47

47 Table 4 3 Patient outcomes of present s ample and IMPACT study Characteristic Present Study IMPACT Study P (n = 546) (n = 8509) Outcome Mortality Survived 35 7 (65%) 6113 (72%) .001 Death 189 (35%) 2396 (28%) .001 6 month GOS Favorable 196 (36%) 4427 (52%) <.001 Good Recovery 98 (18%) 2761 (32%) <.001 Moderate Disability 98 (18%) 1666 (20%) .351 Unfavorable 350 (64%) 4082 (48%) <.001 Sev ere Disability 159 (29%) 1335 (16%) <.001 Vegetative State 2 (0%) 351 (4%) <.001 Death 189 (35%) 2396 (28%) <.001 Aim 1a Replication of IMPACT Core and Extended models Predicting Mortality Logistic regression was used to model classification of m ortality status based on IMPACT scores. The Core IMPACT model score was entered in the first block to determine its contribution to the prediction of mortality. The Extended IMPACT model score was then added to the second block to assess its added contribution to this model. The Core IMPACT model was significantly associated with mortality, with an increase of one point resulting in a 30% increase in the odds of mortality (OR = 1.30, 95% CI = 1.261.41; p < .001). Similarly, when the Extended IMPACT model s core was added an increase of one point for these additional acute physiological predictors also results in a 30% increase in the odds of mortality when the Core IMPACT model score remains constant (OR = 1.30, 95% CI = 1.151.47; p < .001) (Table 44). Nag elkerke R statistics were calculated for the logistic regression models to estimate the amount of variance predicted by the models as a measure of goodness of fit (Nagelkerke, 1991). Nagelkerkes R for the Core IMPACT model was 0.274 indicating that this score explained 27.4% of the variance in mortality after severe TBI in

PAGE 48

48 this sample. The addition of the Extended IMPACT model score improved the predictive ability by 3.7 thus explaining 31.1% of the variance in patient mortality outcome. Table 4 4 Lo gis tic regression analysis for m ortality (n = 546) Note : Step 1 RN = .274, RCS = .199; Step 2 RN = .311, RCS = .225 The Core IMPACT model was able to differentiate survivors from nonsurvivors. Details o f the correct classifications produced by the regression models are shown in Table 4 5. Sensitivity, defined as the proportion of deaths classified correctly by the regression model was 55% for the Core model while specificity, defined as the proportion of survivors classified correctly by the Core IMPACT model was 85%. The addition of the Extended IMPACT model reduced the sensitivity to 53% while specificity improved to 86%. Table 4 5 Sensitivity and specificity of the IMPACT model for mortality predictio n Outcome Prediction Model Sensitivity Specificity Mortality Core 55% 85% + Extended 53% 86% The discriminative ability of the IMPACT model was examined by a receiver operating characteristic (ROC) analysis. Sensitivity and specificity values (express ed [1 Specificity]) were plotted graphically to obtain ROC curves for each prediction model. Patients scores were compared along their full range of cut off points, and the measure B SE B 2 d f p Odds Ratio 95% CI Step 1 Constant 2.213 .201 121.801 1 .000 .109 Core .259 .026 96.997 1 .000 1.30 1.26 1.41 Step 2 Constant 2.974 .287 107.372 1 .000 .051 C ore .236 .027 77.641 1 .000 1.27 1.20 1.33 Extended .264 .063 17.373 1 .000 1.30 1.15 1.47

PAGE 49

49 of the area under the ROC curve (AUC) was used as a quantitative indicat or of the utility of the model to discriminate between survivors and non survivors. As a point of interpretation reference, an AUC probability of 0.50 would indicate that the model fails to discriminate survivors from nonsurvivors, and a perfect test woul d have an AUC probability of 1.00 (Ritchie & Fuhrer, 1992). ROC curve analysis confirmed the discriminatory ability of the Core model for mortality (AUC 0.77). When the Extended model was added, the performance increased to 0.79 (Figure 1). ROC curves for both the Core model alone as well as with the addition of the extended score are presented in Figures 41 and 4 2. Adding the Extended model did not result in a significant improvement to the discriminative capabilities (AUCs) suggesting that CT and secondary injury variables contributed little to mortality prediction.

PAGE 50

50 Figure 4 1. ROC curve for Core model predicting m ortality AUC = .773 Figure 42. ROC c urve for Core + Extended model predictin g m ortality AUC = .792

PAGE 51

51 Aim 1b. Replication of IMPACT Core and Extended Models Predicting Unfavorable Outcome Logistic regression was again used to model classification of Unfavorable outcome (based on GOS) using the Core and Extended IMPACT models as predictors. The Core IMPACT model score was entered in the fi rst block with the Extended IMPACT model score added in the second block. The Core IMPACT model score was significantly associated with Unfavorable outcome, with an increase of one point resulting in a 33% increase in the odds of an Unfavorable outcome (OR = 1.33, 95% CI = 1.261.41; p < .001). An increase of one point to the Extended IMPACT model score resulted in a 30% increase in the odds of Unfavorable outcome (OR = 1.30, 95% CI = 1.171.49; p < .001) when the core model remained constant (Table 46). T hus, the Core and Extended IMPACT models showed the ability to predict Unfavorable outcome. Nagelkerkes R for the Core IMPACT model was .291 indicating that this score explained 29.1% of the variance in Unfavorable outcome after severe TBI in this sample. The addition of the Extended IMPACT model score improved the predictive ability of this model by 4%, explaining a total of 33.1% of the variance in patient outcome at this level of classification. Table 4 6 Logistic r egre ssion analysis for Unfavorable outcome (n = 546) B SE B Wald df P Odds Ratio Confidence Interval Step 1 Constant 0.78 0.16 25.25 1 .000 .46 Core 0.29 0.03 96.45 1 .000 1.33 1.26 1.41 Step 2 Constant 1.51 0.23 41.45 1 .000 .22 Core .26 0.03 77.52 1 .000 1.30 1.23 1.38 Extended .28 0.06 19.22 1 .000 1.32 1.17 1.49 Note : Step 1 RN = .291, RCS = .212; Step 2 RN = .331, RCS = .241

PAGE 52

52 The Core IMPACT model was able to differentiate Favorable from Unfavorable outcomes. Details of the correct classification s produced by the regression models are shown in Table 47. The Core IMPACT model score correctly classified 82% of the patients with Unfavorable outcomes and 62% of those with Favorable outcomes. The addition of the Extended IMPACT model reduced the sensi tivity to 81% while specificity remained at 62%. Table 4 7 Sensitivity and specificity of the IMPACT model for Unfavorable outcome p rediction Outcome Prediction Model Sensitivity Specificity Unfavorable Outcome Core 82% 62% + Extended 81% 62% The dis criminative ability of the IMPACT model for Favorable vs. Unfavorable outcome was again examined by ROC analysis. ROC curve analysis demonstrated that the Core and Extended IMPACT models showed discriminatory ability for Unfavorable vs. Favorable outcome ( Figures 4 3 and 44).

PAGE 53

53 Figure 4 3 ROC curve for the Core IMPACT model predicting Unfavorable outcome AUC = .783 Figure 4 4. ROC c u rve for Core + Extended IMPACT model predicting Unfavorable o utcome AUC = .800

PAGE 54

54 The IMPACT model score with the greatest classification value can be defined as the one with the greatest summed sensitivity and specificity values allowing for both predictive values to be equally emphasized. The sensitivity and specificity for various classification values of the Core and Extended IMPACT model scores are presented in Table 4 8 In this sample, there was no clear cut off score that maximized both sensitivity and specificity when predicting mortality. For the core model, it can be said that patients with scores lower than 6 are l ikely to survive while those with scores greater than 9 have a much poorer prognosis. The Extended IMPACT model showed that patients with IMPACT scores of less than 8 are likely to survive and those with scores lower than 6 will likely have a Favorable out come.

PAGE 55

55 Table 4 8. Sensitivity and specificity at the classification v alues o f the IMPACT Core and Extended models for mortality and Unfavorable outcome prediction Classification Values Sensitivity (%) Specificity (%) Mortality Core 4.5 78 60 5.5 71 64 6.5* 66 74 7.5* 62 78 8.5* 9.5 55 48 85 88 Extended 6.5 7.5* 86 81 55 61 8.5 73 65 9.5 71 69 10.5 65 74 11.5 60 81 Unfavorable Outcome Core 1.5 2.5* 90 82 43 62 3.5 74 66 4.5 68 75 5.5 62 79 6.5 55 87 Extended 4.5 5.5 88 82 47 62 6.5 75 68 7.5 70 75 8.5 61 82 9.5 60 83 Aim 2a Extension of IMPACT Model Predicting Full Range of Functional 6month outcomes (GOS). Before building the model, data was examined and plotted to visualize the ordinal regr ession model. In these analyses, the categories GOS=1 ( Death ) and GOS=2 ( Vegetative State ) were combined, because the percentage of patients in a vegetative state at 6 month post injury was low ( 0. 4%). Figure 45 is a cumulative percentage plot of the IMPA CT total scores with separate curves for each of the GOS levels. Figure 45 indicates that a larger percentage of patients with lower GOS scores had higher

PAGE 56

56 IMPACT scores while a larger percentage of patients with higher GOS scores had low IMPACT scores. Th is shows preliminary evidence of a clear relationship between higher IMPACT model scores and more adverse GOS ratings. Moreover, the figure shows graphically that GOS scores of 4 and 5 are very close suggesting they will be difficult to distinguish from one another using the IMPACT model. Figure 4 5. Plot of observed c umulative percentages IMPACT Core + Extended scores Ordinal logistic regression analyses evaluated the IMPACT model total scores in their prediction of the full range of the GOS. The defaul t Logit link function was used for these analyses. The IMPACT model significantly predicted GOS classification at six p < .001. Nagelkerkes R = .34, and the Test of Parallel lines was not statistically significant, suggesting that the parallel slopes assumption of ord inal regression was met. Table 49 presents the classification table resulting from this analysis, a matrix of the observed GOS classifications versus the values predicted by the statistical model. The model showed sign ificant prediction of conscious survival (GOS 1 2 vs. 3 5) and Favorable outcome (GOS 13 vs. 4 5) at the threshold of Good

PAGE 57

57 Recovery vs. less than Good Recovery the model was not significant indicating poor performance at this upper range of the GOS. Table 12 shows that 246 patients (45%) were classified into the correct GOS rat ings based on model predictions; however, the model did not predict Moderate Disability (GOS= 4 ) This again suggests that the IMPACT model performed poorly when predicting the upper range of GOS scores. Nevertheless, IMPACT scores were significantly and negatively associated with observed GOS classific ations (see Table 4 10 ). Table 4 9. Comparison of observed versus predicted categories in ordinal regression model using IMPACT total S core predicting GOS c ategory Predicted GOS Category Observed Classificatio n 1 or 2 3 5 GOS 1+2 191 125 (66%) 56 (29%) 10 (5%) 3 159 63 (40%) 64 (40%) 32 (20%) 4 98 18 (18%) 34 (35%) 46 (47%) 5 98 10 (10%) 31 (32%) 57 (58%) N=546. The perce ntages listed in the table represent the predicted response/observed classification. Values in bold font highlight predictions that match the observed classification.

PAGE 58

58 Table 4 10. Summary of ord inal regression of IMPACT total score predicting 6month GOS B SE B Wald 2 Df p GOS Conscious Survival vs. Death/VS 3.04 .21 200.4 5 1 .000 Favorable vs. Unfavorable Outcome 1.41 .18 65.00 1 .000 Good Recovery vs. Less than Good .23 .18 1.80 1 .180 IMPACT Total score .25 .02 168.4 0 1 .000 Nagelk erke R2 = .34; p < .001 At the level of Good Recovery (GOS=5) versus less than Good Recovery (GOS=1 4), the ordinal regression of the IMPACT model was not significant. Sensitivity and specificity of the classifications into these GOS categories produced by the Core and Extended IMPACT models are shown in Table 4 11. Table 4 11. Sensitivity and specificity of the IMPACT model for less than Good Recovery outcome prediction Outcome Prediction Model Sensitivity Specificity Unfavorable Outcome Core 0% 100% + Extended 10% 98% Sensitivity (defined as the proportion of patients with less than Good Recovery who were classified correctly by the regression model) was 0% for the core model while specificity (defined as the proportion of patients with Good Recovery classified correctly by the core model) was 100%. This indicates that the IMPACT model classifies all patients as having less than Good Recovery when analyzed at this threshold. Additionally, ROC curve analysis indicated that the classification scores that maximize the sensitivity and specificity of the model at this threshold were the same scores as those determined for the threshold of Favorable vs. Unfavorable outcome. This

PAGE 59

59 indicates that the model of the model would classify the same patients at the tw o thresholds and suggests that this level of dichotomizing the patients gives no additional information. Taken together these results show that the IMPACT model cannot correctly distinguish between the higher GOS categories of Moderate Disability (GOS=4) a nd Good Recovery (GOS=5). Table 4 12. Sensitivity and specificity at the classification v alues of the IMPACT Core and Extended models for Good Recovery p rediction Classificatio n Values Sensitivit y (%) Specificit y (%) Good Recovery Core 1.5 2.5* 79 7 4 59 66 3.5 67 72 4.5 61 82 5.5 54 82 6.5 46 90 Extended 4.5 5.5 79 7 4 61 70 6.5 67 73 7.5 61 79 8.5 5 5 8 7 9.5 52 88 Table 4 13 summarizes the performance of both the Core and Extended IMPACT models at each dichotomy of the G OS: Good Recovery ( GOS 1 4 vs. 5 ), F avorable outcome ( GOS 1 3 vs. 4 5 ), and Conscious survival (GOS 1 2 vs. 3 5) as well as with a proportional odds model, with their 95% confidence intervals. These results showed that the dichotomous odds ratios were rem arkably consistent regardless of how the GOS was grouped. The slopes of all the regression lines were similar so the assumed common slope was used to obtain the proportional odds model. This implies that the proportional odds model gives an excellent summary of the association between the models and the GOS (McHugh et al., 2007). Thus the IMPACT model demonstrated a

PAGE 60

60 significant relationship with the full range of GOS with stronger predictive ability at the lower range of ratings. As demonstrated previously, the findings regarding mor tality and Unfavorable outcome are particularly robust. Thus it is likely that these results compensate for weaker performance at the threshold of Good Recovery when assessing the model for the entire range of GOS. Table 4 13. Summary of proportional odds m odel for Core and Extended IMPACT model predicting GOS C ore Extended OR (95% CI) OR (95% CI) Dichotomous OR Less than good versus Good Recovery 1. 30 1.21 1. 40 1.3 5 1.16 1.56 Unfavorable versus favorable outcome 1 33 1.26 1. 41 1.3 2 1.17 1.49 Death/VS versus conscious survival 1. 30 1.26 1. 41 1. 3 0 1.15 1.4 7 Proportional Odds 1. 30 1.23 1. 40 1.3 3 1.16 1.50 Aim 2b Extension of IMPACT model predicting DRS at six months A linear regression was performed using DRS scores at 6 months post injury as the dependent variable and the IMPACT model total scores as independent predictor. Natural log transformations were performed on DRS scores to produce model residuals that were normally distributed. Table 16 pr esents the details of this regression analysis. As shown in the table, the IMPACT model was a significant predictor of DRS scores accounting for 32.0% of the variance [ F (1,544) = 256.02, p < .001].

PAGE 61

61 Table 4 14. Summary of hierarchical regression analysis with IMPACT scores predicting DRS s cores Note: R = .320 ( p < .001). p < .001 After converting these figures back from the natural log transformation, the equation obtained was: Disability Rating Scale score = 2.83 + (1.14)IMPACT Score This suggests that each point increase in the IMPACT score results in a little over a point increase in DRS score. Aim 2c Comparison of IMPACT prediction of GOS and DRS The linear regression analysis of the IMPACT m odel and the DRS produced and R2 of .320. The logistical regression for the model the models prediction of the full range GOS scores produced a Nagelkerkes R2 of .340 which is slightly higher but generally comparable. These results suggest that the IMPAC T explains a similar portion of the variance with either of these functional outcome measures. Aim 3 Extension of IMPACT Predicting Working Memory Outcome at 6months Table 4 15 presents the Digit Span characteristics of the present study sample. On aver age, patients who were able and available to complete this test performed within normal limits on both longest forward and backward spans (Lezak et al., 2004). These scores demonstrate that the majority of patients who were able to perform this task were able to do so within the range of normal performance. B SE B Constant 1.04 0.08 IMPACT 0.13 0.01 .57*

PAGE 62

62 Table 4 15. Digit Span scores of the current study s ample (N=76) Characteristic M SD Range Longest Forward Span 5.8 1.3 3 9 Longest Backward Span 4.1 1.4 0 7 Digit Span Total Score 12.5 4.6 3 25 For patients capable of performing some neuropsychological testing (n=76), a hierarchical regression analysis was performed using Digit Span total scores at 6 months post injury as the dependent variable and using the IMPACT model total scores as an independent predictor. The continuous variable of tota l Digit Span score was normally distributed. Education was entered as an independent predictor in the first block to control for its contribution to performance on the Digit Span total score (Table 4 16). The percentage of variance explained by education was 10.6%. The IMPACT model scores did not contribute significantly to the model explaining only 2.7% of the variance in Digit Span performance. Table 4 16. Summary of hierarchical regression analysis with education, IMPACT scores predicting Digit Span total s cores (N=76) B SE B Step 1 Constant 5.41 2.65 Education .64 .23 .33* Step 2 Constant 6.86 2.83 Education .62 .23 .32* IMPACT .21 .15 .16 Note : Step 1 R = .106 ( p = .007); Ste 027 ( p = .1 63 ). p < .05

PAGE 63

63 CHAPTER 5 DISCUSSION Study Summary The first aim of the current study was to replicate Steyerberg et als IMPACT prognostic models of severe TBI predicting global outcome classifications of mortality and a dichotomous GOS classification of Favorable vs. Unfavorable outcome. The current study then sought to extend the findings of the IMPACT model by testing its ability to predict a wider range of more ecologically valid measures of outcomes using the full range of the GOS and the DRS rating scales. Finally, the current study sought to explore whether the IMPACT model could additionally predict cognitive outcome among survivors. More specifically, performance on Digit Span at 6 months post injury was used as a measure of cognitive outcome in the domain of working memory. Several significant group dif ferences in patient predictor characteristics were detected between the current study sample and the original IMPACT study (Steyerberg 2008). Overall, these differences showed the current study sample had sustained more severe injuries. This finding was expected given the restriction of the current study sample to patients with severe TBI while the IMPACT study included patients with moderate TBI as well. Regardless, for the di fferences that were detected, the percentages of patients with each predictor found in the present study sample fell within the range of percentages reported for the individual studies that were compiled to create the IMPACT database. Regarding outcomes fo r participants in the current study, the mortality rate was again within the range reported for the studies included in the IMPACT model and was consistent with previous research for patients with severe TBI (Hukkelhoven et al., 2003; Jennett et al., 1976; Murray et al., 1999; Nakamura et al.,

PAGE 64

64 2006). In terms of the distribution GOS outcome found in the current study sample, a greater proportion of the patients died or had a Severe Disability at six months post injury with fewer having a Good Recovery for patients in the IMPACT study. Although these GOS outcome distributions were again statistically different from the IMPACT sample, the pattern and frequency were similar to those found in the research literature on specifically severe TBI (Thornhill 2000, Mo ppett 2000). The current sample DRS distribution showed that the mean outcome of the current study sample would be classified as a Severe Disability (Rappaport et al., 1998); however, there was a high level of variability as has been noted in the outcome l iterature for patients with severe TBI (Struchen et al., 2001; Pastorek et al., 2004). In terms of cognitive functioning the mean Digit Span scores in the current study sample fell below average but still within the normal range according to Lezak et al. ( 2004). The scores were commensurate with previous studies (Fork et al., 2005), and as Lezak et al. report it is not surprising to find a decrease in Digit Span scores in the first months following head trauma (2004). Hypothesis 1 This study first sought t o replicate the IMPACT prediction model findings of Steyerberg et al. (2008) and hypothesized that the IMPACT model would significantly predict the risk of mortality patients six months after severe TBI. The study findings supported this hypothesis and dem onstrated that the combination of acute physiological variables in the IMPACT core model (including age, motor GCS, and pupil reactivity) can effectively discriminate between survivors and non survivors at 6 months post injury. This finding is consistent w ith the some of the earliest prognostic models ( Teasdale and Jennett, 1976; Jennet et al., 1976) as well as many years of subsequent research that show that these three common predictors create a robust prognostic

PAGE 65

65 model for mortality prediction (Jain et al ., 2008; Choi et al., 1983; Clifton et al., 1993; Hukkelhoven et al., 2005). This finding is further significant because it not only reiterates the predictive ability of these three commonly used predictors, but it also demonstrates that they retain the pr edictive ability when weighted and combined into a single composite score based on the IMPACT model. The results further showed that the addition of information regarding secondary injuries of hypoxia and hypotension as well as head CT characteristics in t he IMPACT Extended Model further strengthened its predictive ability for mortality. The findings of the IMPACT models predictive ability in the current study provide further external validation for the model. To further replicate the IMPACT study, the cur rent study hypothesized that the model could discriminate between patients who had Favorable outcomes (GOS ratings = 4 5) from those who had Unfavorable outcomes (GOS ratings = 1 3). The study findings confirmed this hypothesis for both the Core and Extended IMPACT models, with both demonstrating adequate discriminative ability between patients with Favorable and Unfavorable outcomes. The largest amount of prognostic information was contained in the Core model, but the Extended model did provide a small deg ree of additional prognostic value. The discriminative ability of the IMPACT model demonstrated in the current study was adequate and comparable to the findings of the original IMPACT study as a similar proportion of the variance in Unfavorable outcome was explained (Steyerberg et al., 2008) Although several prognostic models have been created for TBI over the years, there has been little research regarding the generalizability of these models. The current study findings demonstrate generalizability of the IMPACT model across different

PAGE 66

66 populations. Although the mean patient characteristics of the current study sample were still within the range reported for the studies the IMPACT patients were drawn from, they were in some cases significantly different from the means of the IMPACT study. As mentioned above, patients in the current study sample had more severe TBIs on average and were a more contemporary sample than the IMPACT studies (19952010 vs. 19841997). The fact that the IMPACT model demonstrated ade quate discriminative ability despite these differences in sample characteristics speaks to the robustness of the prognostic model and its applicability to contemporary outcome prediction. However, current study suggests that the risk of adverse outcome aft er severe TBI has not improved significantly in the past decade. While this is a discouraging finding, it is consistent with other research (Stein et al., 2010; Moppet et al., 2000; Nakamura et al., 2006) and highlights the continued need for medical advances for this patient population. Further analyses showed additional ways that the IMPACT models ability to predict mortality and Unfavorable outcome could be useful for clinical and research purposes. The results showed that to maximize specificity and s ensitivity an IMPACT classification value of 8 or greater can be used to identify patients at a high risk of Death and an IMPACT value of 6 or higher can be used to identify patients at a high risk of an Unfavorable outcome. These classification values cou ld serve as ways to improve the classification of TBI severity or serve as cut off points for the purposes of inclusion in clinical trials or research studies. Additionally, if applied appropriately, these scores could be useful in aiding with clinical dec ision making. The IMPACT model with its webbased interface was intended to be an easy tool for clinicians to use predicting outcome

PAGE 67

67 for their patients (Maas et al., 2008). However, studies have shown that this type of prognostic information has often been misused and implemented to make treatment limiting decisions (Murray et al., 1993). While the sensitivity at these classification values is fairly high (81% and 82% respectively), this classification still commits type II error, which falsely classifies about 20% of patients as dying or having an Unfavorable outcome. Therefore, it should be emphasized as Steyerberg et al. argues, that these statistical models should only be used to augment and not replace clinical judgment (2008). Hypothesis 2 The current study also sought to extend the findings of the IMPACT model and hypothesized that the IMPACT model could predict a broader range of outcomes as measured by the full range of the GOS. The results of the proportional odds model seem to confirm this hypothesis, as it was demonstrated that higher IMPACT model scores shift GOS classifications in an adverse direction for the entire distribution of GOS. An increase of one point in the total IMPACT score was associated with increasing the odds of adverse outcome by approximately 33% at each possible dichotomy of the GOS. However, further analyses showed that while the IMPACT model performs well when predicting lower ratings of GOS such as mortality and Unfavorable outcome, it demonstrated poorer predictive abilit y for the highest GOS rating of Good Recovery. As discussed earlier the outcome dichotomy most commonly used in the severe TBI literature is Unfavorable (GOS = 1 3) and Favorable (GOS = 4 5). The IMPACT model performed best when this threshold was used to dichotomize the results. This is consistent with previous research that has shown acute prognostic models are

PAGE 68

68 successful at discriminating at this threshold (Choi et al., 1983; Elf et al., 2002; Hukkelhoven et al., 2005). The model demonstrated poorer sens itivity when predicting Good Recovery vs. less than Good Recovery suggesting that the model is not as good at classifying patients with better outcomes. The model predicted almost all patients would fall below the threshold for Good Recovery. For this threshold, the sensitivity and specificity were optimal when using the same cut off scores as indicated for the Favorable/Unfavorable dichotomy. This suggests that the model would make the same classifications of patients at either threshold and that the model could not adequately predict which patients would have a Good Recovery (GOS = 5). Therefore, the hypothesis that the IMPACT model would predict the full range of GOS was not confirmed. Patients in a Vegetative State (GOS =2) were grouped with non survivor s in the analyses, and the model did not show significant predictive ability for patients with Good Recovery Thus, in terms of GOS ratings, the IMPACT model could only significantly predict mortality and Unfavorable outcome dichotomies as already demonstr ated in previous analyses. A further study hypothesis was that the IMPACT model would be able to predict to the full range of the DRS. This hypothesis was confirmed as the model accounted for 31.1% of the variance in DRS at 6 months post injury. This is a sizable portion of the variance in DRS scores and is significantly higher than has been found in previous studies using just individual acute variables to predict DRS scores at 6 months post injury (Struchen et al., 2001). Additionally, the findings showed that IMPACT models ability to predict the full range of more detailed outcome as measured by the DRS was comparable to its ability to predict mortality and Unfavorable outcome. The Nagelkerke

PAGE 69

69 pseudo R2 estimates for the logistic regression analyses of m ortality and Unfavorable outcome were .311 and .331 respectively, whereas the R2 for the linear regression on DRS was .320. This suggests that when paired with an effective outcome measure that is collected appropriately and measures a range of outcomes, t he IMPACT model can provide more information than was demonstrated in the Steyerberg et al study (2008). This provides further support that it is reasonable to use the combination and weighting of acute predictors in the IMPACT model as a single composite score when predicting functional outcome after TBI. Hypothesis 3 Finally, this study sought to extend the Steyerberg findings even further and hypothesized that the IMPACT model would predict cognitive outcome (specifically working memory skills) after TB I. This hypothesis, however, was not supported, as the IMPACT model demonstrated little utility when applied to the basic cognitive test of Digit Span. The extension of the model to Digit Span did not show to be effective beyond what was predicted for thes e patients given their levels of education. This analysis may have suffered from decreased power as only 98 (18%) of the patients in the current study were available and able to complete the digit span test, and only 76 (14% of the total) of those patients provided information regarding their education level. Thus the majority of patients in the current study were unable to complete the Digit Span test due to a variety of reasons such as medical complications, being in an acute confusional state, being othe rwise unresponsive or because they were unavailable. This caveat aside, the current finding of poor prediction of Digit Span performance is consistent with the finding that the IMPACT model demonstrated worse ability to predict Good Recovery on the GOS as compared to the lower ratings. With this weakness of the

PAGE 70

70 model, it is unsurprising that it also struggles to differentiate among those with best outcomes in terms of their cognitive ability. This suggests that for more precise prediction of cognitive funct ioning after severe TBI, it is likely that an effective prognostic would have to include information beyond the acute physiological data. There are likely to be many factors that contribute to some patients differentially retaining or regaining the cogniti ve abilities assessed by the Digit Span test at 6 months post injury. It seems likely that the variability observed among patients classified as having a Favorable outcomes are less a result of the acute injury severity and instead have more to do with pos t acute events and treatments, as well as interventions such as rehabilitation that take place after discharge from the hospital. This is a significant weakness of this study as none of this information could be taken into account. Thus the results suggest that the IMPACT model can predict whether a patient will have a Favorable outcome from the earliest time points. However, it appears that whether the patient will end up in the higher or lower ends of that range depends more on what happens during the six months that follows the injury rather than merely the status of the patient on the day of admission. Implications One application of the IMPACT model could be to offer a more precise classification of TBI than is traditionally provided from GCS scores alone. As evidenced by the current study sample and demonstrated in previous research, considerable heterogeneity can still exist in terms of admission characteristics and outcome among patients classified as severe TBI using primarily the GCS (i.e., GCS 3 8) (Hukkelhoven et al., 2003; Jennett et al., 1977; Murray et al., 1999; Nakamura et al., 2006). This heterogeneity muddles the picture for clinical and research purposes. Clinically, more

PAGE 71

71 accurate classification could be used to better link a specific sever ity of injury with an appropriate therapeutic intervention. Additionally, a more accurate classification of severity may be used to better evaluate the efficacy of health care (both acute medical treatment as well as rehabilitation services) in TBI as outc ome assessment could use targets that were more relevant to the initial injury severity. Clinicians could also use these predicted risk estimates as a method of measuring their efforts to continually improve their services and challenge themselves to obtai n better results than predicted. Similarly, more accurate classification of severe TBI patients according to their predicted risk could play an important role in designing research studies and clinical trials. Without a more accurate way to classify sever e TBI, clinical trials must include patients with such poor prognosis that they will do poorly no matter how good the treatment is. Thus, inclusion of these worst of the patients with severe TBI simply decreases the statistical power of these studies (Cali f f et al., 1997). One noteworthy caveat to the use of this model to exclude patients from research studies and clinical trials is that the cut off scores for the IMPACT model found in the current study (i.e., 6 for Unfavorable outcome and 8 for mortality) would likely exclude most older adults because ages over 60 already add 5 points to the overall score. Thus, the generalizability of these research studies and clinical trials to populations of older adults would be limited. Overall, the current study hi ghlights that use of the IMPACT model in early patient assessment could provide physicians with powerful prognostic information beyond mere predictions of mortality or Unfavorable outcome. This could be important for providing families and caregivers with more information about what to expect in

PAGE 72

72 terms of recovery. Previous studies have reported that t he majority of caregivers and family members want to be told as much information as possible regarding the patients prognosis within the first 24 to 48 hours after injury (Shields, 1998). However, clinicians need to be aware of the limits and statistical uncertainty of this prognostic information and exercise care not to overemphasize the predictions when speaking with families, because this type of information is likely to be remembered (Carey et al., 2001). Armed with more information families can begin to prepare for the road ahead and learn what role they can play to best optimize the recovery of their loved one. Study Strengths A dominant strength of the current study is in the selection of outcome measures. The DRS provided a measure with a broad range of information that is still relatively easy to collect. Relatedly, this study benefited from strong methods of outcome measure collection, as both GOS and D RS ratings were rigorously quality controlled. While the current study was limited to use of the original GOS, future studies could validate and extend the findings using the structured interview version of the GOS or the Extended Glasgow Outcome Scale (GO S E) to provide more breadth of potential outcomes and more structured means of measurement. Another important strength inherent in the current study was the variability of the sample including a distribution of races that included a significant proportion of Hispanic patients, as well as data collection that occurred across several years and across sites. This variability helped to demonstrate that the IMPACT is quite generalizable and still produces accurate predictions in spite of differences in settings time periods, and ethnic groups.

PAGE 73

73 Study Limitations A limitation of the current study was the inconsistency in the collection of GCS information. The majority of the GCS ratings for the current study sample were obtained post resuscitation. In other words GCS prediction values were within one hour of admission after doctors have indicated that hypoxia and hypotension are stabilized, but before any surgery or other interventions. However, a small number of the data points were gathered via chart review and were an approximation of this time point. While this is a weakness in the data, it still fits within the parameters of the IMPACT study and clinical applicability as this is a limitation inherent in this clinical setting. In order to improve the applicati on of prognostic models that contain GCS information, these processes should be better standardized to gather the best possible acute data. Finally, there was a high level of attrition for the collection of the cognitive outcome variable examined in this study. Therefore it is possible that a number of patients were capable of performing the Digit Span test, but were unavailable to complete it. Additionally, the data was incomplete in reporting whether any kind of impairments prevented the patients from co mpleting this test. This information could be helpful to better characterize the sample and further explain these findings. Unfortunately, o btaining neuropsychological test data can be much more difficult than gathering functional outcome information ( Clif ton 1993 ). Study Summary and Future Directions The current study sought to replicate and extend a well validated prognostic model of outcome after severe TBI (Steyerberg et al., 2008) and t o examine its utility in predicting functional outcome on measures that capture a broader range of ability. The study showed that the IMPACT model has utility beyond predicting mortality and

PAGE 74

74 Unfavorable outcome, especially the predictors are when paired with an outcome measure like the DRS that is scrupulously collected and controlled for data quality. The model demonstrated weakness at the higher end of the functioning spectrum and failed to consistently and appropriately classify patients who had more favorable outcomes. Moreover the model showed limited utility when predicting cognitive working memory ability 6 months after injury. These findings indicate the model remains useful as an acute predictor composite, but they suggest that other variables would likely contribute more to models of prediction among patients w ith better outcomes. Future studies should examine the models utility in combination with post acute measures such as protein biomarkers, MRI, and the same physiological variables it measures acutely, through the post acute phase. As the higher functioni ng patients progress in their recovery, predictions of finder detail outcome (e.g., cognitive) become more urgent for discharge placement and family planning. Therefore studies that tested the IMPACT model along with additional data on post acute physiological events and treatment as well as discharge placement variables regarding rehabilitation services would likely provide more information regarding which patients have the best chance of a full recovery. Nevertheless, the current model shows promise for creating more precise classifications of injury severity to help reduce the heterogeneity among severe TBI patients for research purposes and clinical trials. Future studies should examine the properties of this model for this purpose of proposing sub cla ssifications. From a clinical standpoint, more precision in severe TBI classification could provide a much more accurate snapshot than is currently available which would enable quicker responses

PAGE 75

75 and better exchange of information among physicians. The IMPA CT model could provide a powerful tool to communicate information among clinicians and to patients families.

PAGE 76

76 LIST OF REFERENCES Anderson, S. I., Housley, A. M., Jones, P. A., Slattery, J., & Miller, J. D. (1993). Glasgow Outcome Scale: an inter rater rel iability study. Brain Inj, 7(4), 309 317. Andrews, P. J., Sleeman, D. H., Statham, P. F., McQuatt, A., Corruble, V., Jones, P. A., et al. (2002). Predicting recovery in patients suffering from traumatic brain injury by using admission variables and physio logical data: a comparison between decision tree analysis and logistic regression. J Neurosurg, 97 (2), 326336. Ashley, M. J., & Ashley, S. M. (2004). Discharge Planning in Traumatic Brain Injury Rehabilitation. In M. J. Ashley (Ed.), Traumatic brain injury: Rehabilitative treatment and case management (2nd ed., pp. 727755). Boca Raton, FL: CRC Press. Badjatia, N Carney N Crocco T J. Fallat M E Hennes H M Jagoda, A S Jernigan, S Lerner E B Letarte, P B Moriarty T Pons P T Sass er S Scalea, T M Schleien, C & Wright D W. ( 2007 ) Guidelines for prehospital management of traumatic brain injury. 2nd ed. New York, NY: Brain Trauma Foundation. Bigler, E. D. (2005). Structural imaging. In J. M. Silver, T. W. McAllister & S. C. Y udofsky (Eds.), Textbook of traumatic brain injury. (pp. 79 105). Arlington, VA: American Psychiatric Publishing Inc. Bishara, S. N., Partridge, F. M., Godfrey, H. P., & Knight, R. G. (1992). Posttraumatic amnesia and Glasgow Coma Scale related to outcom e in survivors in a consecutive series of patients with severe closedhead injury. Brain Inj, 6(4), 373380. Butcher I Maas A I Lu J. Marmarou, A M urray, G.D., Mushkudiani, N.A., ( 2007) Prognostic value of admission blood pressure in traumatic brain injury: results from the IMPACT study Journal of Neurotrauma 24 294 302 Changaris, D. G., McGraw, C. P., Richardson, J. D., Garretson, H. D., Arpin, E. J., & Shields, C. B. (1987). Correlation of cerebral perfusion pressure and Glasgow Coma Scale to outcome. J Trauma, 27 (9), 10071013. Charlson, M., Szatrowski, T.P., Peterson, J., & Gold, J. (1994). Validation of a combined comorbidity index. Journal of Clinical Epidemiology 47(11) 12471251. Chesnut, R. M., Marshall, L. F., Klauber, M. R., Blun t, B. A., Baldwin, N., Eisenberg, H. M., et al. (1993). The role of secondary brain injury in determining outcome from severe head injury. J Trauma, 34 (2), 216222.

PAGE 77

77 Choi, S. C., Ward, J. D., & Becker, D. P. (1983). Chart for outcome prediction in severe h ead injury. J Neurosurg, 59(2), 294297. Clifton, G. L., Kreutzer, J. S., Choi, S. C., Devany, C. W., Eisenberg, H. M., Foulkes, M. A., et al. (1993). Relationship between Glasgow Outcome Scale and neuropsychological measures after brain injury. Neurosurgery, 33(1), 34 38; discussion 3839. Cope, D. N., Cole, J. R., Hall, K. M., & Barkan, H. (1991). Brain injury: analysis of outcome in a post acute rehabilitation system. Part 1: General analysis. Brain Inj, 5 (2), 111125. de Guise, E., LeBlanc, J., Feyz, M., & Lamoureux, J. (2006). Prediction of outcome at discharge from acute care following traumatic brain injury. J Head Trauma Rehabil, 21(6), 527536. Dikmen, S., Machamer, J., Miller, B., Doctor, J., & Temkin, N. (2001). Functional status examination: a new instrument for assessing outcome in traumatic brain injury. J Neurotrauma, 18(2), 127 140. Dikmen, S., Machamer, J., & Temkin, N. (1993). Psychosocial outcome in patients with moderate to severe head injury: 2 year follow up. Brain Inj, 7(2), 113 124. Elf, K., Nilsson, P., & Enblad, P. (2002). Outcome after traumatic brain injury improved by an organized secondary insult program and standardized neurointensive care. Crit Care Med, 30 (9), 2129 2134. Eliason, M. R., & Topp, B. W. (1984). Predictive v alidity of Rappaport's Disability Rating Scale in subjects with acute brain dysfunction. Phys Ther, 64(9), 1357 1360. Ellenberg, J. H., Levin, H. S., & Saydjari, C. (1996). Posttraumatic Amnesia as a predictor of outcome after severe closed head injury. P rospective assessment. Arch Neurol, 53(8), 782 791. Fork, M., Bartels, C., Ebert, A. D., Grubich, C., Synowitz, H., & Wallesch, C. W. (2005). Neuropsychological sequelae of diffuse traumatic brain injury. Brain Inj, 19 (2), 101108. Gouvier, W. D., Blanto n, P. D., LaPorte, K. K., & Nepomuceno, C. (1987). Reliability and validity of the Disability Rating Scale and the Levels of Cognitive Functioning Scale in monitoring recovery from severe head injury. Arch Phys Med Rehabil, 68(2), 9497. Hall, K., Cope, D N., & Rappaport, M. (1985). Glasgow Outcome Scale and Disability Rating Scale: comparative usefulness in following recovery in traumatic head injury. Arch Phys Med Rehabil, 66(1), 35 37.

PAGE 78

78 Haydel, M. J., Preston, C. A., Mills, T. J., Luber, S., Blaudeau, E ., & DeBlieux, P. M. (2000). Indications for computed tomography in patients with minor head injury. N Engl J Med, 343(2), 100 105. Hemmelgarn, B. R., Manns, B. J., Quan, H., & Ghali, W. A. (2003). Adapting the Charlson Comorbidity Index for use in patients with ESRD. Am J Kidney Dis, 42(1), 125132. Howard, M. A., 3rd, Gross, A. S., Dacey, R. G., Jr., & Winn, H. R. (1989). Acute subdural hematomas: an agedependent clinical entity. J Neurosurg, 71 (6), 858863. Hukkelhoven, C. W., Steyerberg, E. W., Habb ema, J. D., Farace, E., Marmarou, A., Murray, G. D., et al. (2005). Predicting outcome after traumatic brain injury: development and validation of a prognostic score based on admission characteristics. J Neurotrauma, 22(10), 10251039. Hukkelhoven, C. W., Steyerberg, E. W., Rampen, A. J., Farace, E., Habbema, J. D., Marshall, L. F., et al. (2003). Patient age and outcome following severe traumatic brain injury: an analysis of 5600 patients. J Neurosurg, 99(4), 666673. Jain, S., Dharap, S. B., & Gore, M. A. (2008). Early prediction of outcome in very severe closed head injury. Injury, 39 (5), 598 603. Jennett, B., & Bond, M. (1975). Assessment of outcome after severe brain damage. Lancet, 1 (7905), 480484. Jennett, B., Teasdale, G., Braakman, R., Minderhoud, J., & Knill Jones, R. (1976). Predicting outcome in individual patients after severe head injury. Lancet, 1 (7968), 10311034. Jennett, B., Teasdale, G., & Knill Jones, R. (1975). Prognosis after severe head injury. Ciba Found Symp(34), 309 324. Jerem itsky, E., Omert, L., Dunham, C. M., Protetch, J., & Rodriguez, A. (2003). Harbingers of poor outcome the day after severe brain injury: hypothermia, hypoxia, and hypoperfusion. J Trauma, 54 (2), 312319. Levin, H. S., O'Donnell, V. M., & Grossman, R. (197 9). The Galveston orientation and amnesia test. The Journal of Nervous and Mental Disease, 367(11), 675684. Lezak, M. D. (1979). Recovery of memory and learning functions following traumatic brain injury. Cortex, 15(1), 63 72.

PAGE 79

79 Lezak, M. D., Howieson, D B., Loring, D. W., Hannay, H. J., & Fischer, J. S. (2004). Neuropsychological assessment (4th ed.). New York, NY: Oxford University Press. Lobato, R. D., Cordobes, F., Rivas, J. J., de la Fuente, M., Montero, A., Barcena, A., et al. (1983). Outcome from severe head injury related to the type of intracranial lesion. A computerized tomography study. J Neurosurg, 59 (5), 762774. Maas, A. I., Hukkelhoven, C. W., Marshall, L. F., & Steyerberg, E. W. (2005). Prediction of outcome in traumatic brain injury wit h computed tomographic characteristics: a comparison between the computed tomographic classification and combinations of computed tomographic predictors. Neurosurgery, 57(6), 1173 1182; discussion 1173 1182. Maas, A. I., Marmarou, A., Murray, G. D., Teasd ale, S. G., & Steyerberg, E. W. (2007). Prognosis and clinical trial design in traumatic brain injury: the IMPACT study. J Neurotrauma, 24(2), 232 238. Maas, A. I., Steyerberg, E. W., Butcher, I., Dammers, R., Lu, J., Marmarou, A., et al. (2007). Prognost ic value of computerized tomography scan characteristics in traumatic brain injury: results from the IMPACT study. J Neurotrauma, 24 (2), 303314. Manley, G., Knudson, M. M., Morabito, D., Damron, S., Erickson, V., & Pitts, L. (2001). Hypotension, hypoxia, and head injury: frequency, duration, and consequences. Arch Surg, 136(10), 11181123. Marmarou, A., Lu, J., Butcher, I., McHugh, G. S., Murray, G. D., Steyerberg, E. W., et al. (2007). Prognostic value of the Glasgow Coma Scale and pupil reactivity in t raumatic brain injury assessed prehospital and on enrollment: an IMPACT analysis. J Neurotrauma, 24(2), 270 280. Marmarou, A., Lu, J., Butcher, I., McHugh, G. S., Mushkudiani, N. A., Murray, G. D., et al. (2007). IMPACT database of traumatic brain injury : design and description. J Neurotrauma, 24(2), 239 250. Marmarou, A., Nichols, J., Burgess, J., Newell, D., Troha, J., Burnham, D., et al. (1999). Effects of the bradykinin antagonist Bradycor (deltibant, CP 1027) in severe traumatic brain injury: result s of a multi center, randomized, placebocontrolled trial. American Brain Injury Consortium Study Group. J Neurotrauma, 16 (6), 431444. Marshall, L. F., Marshall, S. B., Klauber, M. R., Van Berkum Clark, M., Eisenberg, H., Jane, J. A., et al. (1992). The diagnosis of head injury requires a classification based on computed axial tomography. J Neurotrauma, 9 Suppl 1, S287292.

PAGE 80

80 McCauley S R ., Levin H. S., Vanier M., Mazaux JM., Boake C., Goldfader P.R., Rockers, D., Butters M., Kareken D.A., Lambert J., & Clifton G. L., (2001). The neurobehavioural rating scalerevised: sensitivity and validity in closed head injury assessment Journal of Neurology Neurosurgery and Psychiatry, 71(5), 643651. McHugh, G. S., Butcher, I., Steyerberg, E. W., Lu, J., Mushkudiani, N., Marmarou, A., et al. (2007). Statistical approaches to the univariate prognostic analysis of the IMPACT database on traumatic brain injury. J Neurotrauma, 24(2), 251 258. McHugh, G. S., Engel, D. C., Butcher, I., Steyerberg, E. W., Lu, J. Mushkudiani, N., et al. (2007). Prognostic value of secondary insults in traumatic brain injury: results from the IMPACT study. J Neurotrauma, 24(2), 287293. McNett, M. (2007). A review of the predictive ability of Glasgow Coma Scale scores in headinj ured patients. J Neurosci Nurs, 39 (2), 6875. Millis, S. R., Rosenthal, M., Novack, T. A., Sherer, M., Nick, T. G., Kreutzer, J. S., et al. (2001). Long term neuropsychological outcome after traumatic brain injury. J Head Trauma Rehabil, 16(4), 343 355. Moons, K. G., Harrell, F. E., & Steyerberg, E. W. (2002). Should scoring rules be based on odds ratios or regression coefficients? J Clin Epidemiol, 55(10), 10541055. Moppet, K. (2007). Traumatic brain injury: assessment, resuscitation and early management British Journal of Anaesthesia 99(1):1831. Mosenthal, A. C., Lavery, R. F., Addis, M., Kaul, S., Ross, S., Marburger, R., et al. (2002). Isolated traumatic brain injury: age is an independent predictor of mortality and early outcome. J Trauma, 52 (5), 907911. Murray, G. D., Butcher, I., McHugh, G. S., Lu, J., Mushkudiani, N. A., Maas, A. I., et al. (2007). Multivariable prognostic analysis in traumatic brain injury: results from the IMPACT study. J Neurotrauma, 24(2), 329337. Murray, L. S., Teasdal e, G. M., Murray, G. D., Miller, D. J., Pickard, J. D., & Shaw, M. D. (1999). Head injuries in four British neurosurgical centres. Br J Neurosurg, 13(6), 564569. Narayan, R. K., Greenberg, R. P., Miller, J. D., Enas, G. G., Choi, S. C., Kishore, P. R., e t al. (1981). Improved confidence of outcome prediction in severe head injury. A comparative analysis of the clinical examination, multimodality evoked potentials, CT scanning, and intracranial pressure. J Neurosurg, 54(6), 751 762.

PAGE 81

81 Novack, T. A., Bergqui st, T. F., Bennett, G., & Gouvier, W. D. (1991). Primary caregiver distress following severe head injury. Journal of Head Trauma Rehabilitation, 6 (4), 6976. Novack, T. A., Bush, B. A., Meythaler, J. M., & Canupp, K. (2001). Outcome after traumatic brain injury: pathway analysis of contributions from premorbid, injury severity, and recovery variables. Arch Phys Med Rehabil, 82(3), 300 305. Ostrosky, F. & Lozano, A. (2006). Digit Span: Effect of education and culture. International Journal of Psychology, 41(5), 333 341 Perel, P., Arango, M., Clayton, T., Edwards, P., Komolafe, E., Poccock, S., et al. (2008). Predicting outcome after traumatic brain injury: practical prognostic models based on large cohort of international patients. BMJ, 336(7641), 425 429. Perel, P., Edwards, P., Wentz, R., & Roberts, I. (2006). Systematic review of prognostic models in traumatic brain injury. BMC Med Inform Decis Mak, 6, 38. Rappaport, M., Hall, K. M., Hopkins, K., Belleza, T., & Cope, D. N. (1982). Disability rat ing scale for severe head trauma: coma to community. Arch Phys Med Rehabil, 63(3), 118123. Rassovsky, Y., Satz, P., Alfano, M. S., Light, R. K., Zaucha, K., McArthur, D. L., et al. (2006). Functional outcome in TBI I: neuropsychological, emotional, and b ehavioral mediators. J Clin Exp Neuropsychol, 28(4), 567 580. Ritchie, K. and Fuhrer, R. (1992) A comparative study of the performance of screening tests for senile dementia using receiver operating characteristics analysis. Journal of Clinical Epidemioogy 45, 627637. Ritter, A. M., Muizelaar, J. P., Barnes, T., Choi, S., Fatouros, P., Ward, J., et al. (1999). Brain stem blood flow, pupillary response, and outcome in patients with severe head injuries. Neurosurgery, 44(5), 941 948. Ropper, A. H., Cole, D., & Louis, D. N. (1991). Clinicopathologic correlation in a case of pupillary dilation from cerebral hemorrhage. Arch Neurol, 48(11), 11661169. Rubin, H R. & Wu A W. (1992). The Risk of Adjustment Medical Care, 30( 11) 973975. Satz, P., Zaucha, K., Forney, D. L., McCleary, C., Asarnow, R. F., Light, R., et al. (1998). Neuropsychological, psychosocial and vocational correlates of the Glasgow Outcome Scale at 6 months post injury: a study of moderate to severe traumatic brain injury patients. Brain Inj, 12 (7), 555567.

PAGE 82

82 Servadei, F., Murray, G. D., Teasdale, G. M., Dearden, M., Iannotti, F., Lapierre, F., et al. (2002). Traumatic subarachnoid hemorrhage: demographic and clinical study of 750 patients from the European brain injury consortium survey of head injuries. Neurosurgery, 50(2), 261 267; discussion 267269. Servadei, F., Nasi, M. T., Giuliani, G., Cremonini, A. M., Cenni, P., Zappi, D., et al. (2000). CT prognostic factors in acute subdural haematomas: the value of the 'worst' CT scan. Br J Ne urosurg, 14(2), 110116. Shields, C.E. (1998). Giving Patients Bad News. Oncology 25(2) 381390. Stein, S.C., Georgoff, P., Meghan, S., Mizra, K., Sonnad, S.S. (2010). 150 Years of Treating Severe Traumatic Brain Injury: A Systematic Review of Progress i n Mortality Journal of Neurotrauma 27(1) 111. Steyerberg, E. W., Mushkudiani, N., Perel, P., Butcher, I., Lu, J., McHugh, G. S., et al. (2008). Predicting outcome after traumatic brain injury: development and international validation of prognostic scores based on admission characteristics. PLoS Med, 5(8), e165; discussion e165. Struchen, M. A., Hannay, H. J., Contant, C. F., & Robertson, C. S. (2001). The relation between acute physiological variables and outcome on the Glasgow Outcome Scale and Disabil ity Rating Scale following severe traumatic brain injury. J Neurotrauma, 18(2), 115 125. Teasdale, G., & Jennett, B. (1974). Assessment of coma and impaired consciousness. A practical scale. Lancet, 2(7872), 8184. Temkin, N. R., Machamer, J. E., & Dikme n, S. S. (2003). Correlates of functional status 3 5 years after traumatic brain injury with CT abnormalities. J Neurotrauma, 20(3), 229241. Thornhill S Teasdale, G M Murray G D McEwen J. Roy C W & Penny K I. ( 2000). Disability in young peo ple and adults one year after head injury: prospective cohort study. British Medical Journal 320: 1631 5 Tupper, D.E., & Cicerone, K.D. (1990). The Neuropsychology of Everyday Life: Assessment and Basic Competencies New York, NY: Springer. Walia, S., & Sutcliffe, A. J. (2002). The relationship between blood glucose, mean arterial pressure and outcome after severe head injury: an observational study. Injury, 33 (4), 339344. Wechsler D., (1997). WAISIII administration and scoring manual San Antonio, T X: The Psychological Corporation.

PAGE 83

83 Wehman, P., Targett, P., West, M. & Kregel, J. (2005). Productive work and employment for persons with TBI: What have we learned after 20 years? Journal of Head Trauma Rehabilitation 20(2), 115 127. Wilson, J. T., Pettig rew, L. E., & Teasdale, G. M. (1998). Structured interviews for the Glasgow Outcome Scale and the extended Glasgow Outcome Scale: guidelines for their use. J Neurotrauma, 15 (8), 573585. Young, B., Runge, J. W., Waxman, K. S., Harrington, T., Wilberger, J ., Muizelaar, J. P., et al. (1996). Effects of pegorgotein on neurologic outcome of patients with severe head injury. A multicenter, randomized controlled trial. JAMA, 276 (7), 538543. Young, N. H., & Andrews, P. J. (2008). Developing a prognostic model f or traumatic brain injury --a missed opportunity? PLoS Med, 5(8), e168.

PAGE 84

84 BIOGRAPHICAL SKETCH William Daniel Watson received a Bachelor of Science in p sychology and a Bachelor of Arts in German at the University of Utah. While at the University of Utah he also studied chemistry and completed the honors program. He then began his graduate studies at University of Florida where he earned a Master of Science in clinical psychology. There, he continued his studies toward a d octorate degree in clinical p sycho logy with a focus on neuropsychology.