Groupings of Persons With Traumatic Brain Injury: A New Approach to Classifying Traumatic Brain Injury in the Post-Acute Period

Sherer, Mark PhD, ABPP, FACRM; Nick, Todd G. PhD; Sander, Angelle M. PhD; Melguizo, Maria MS; Hanks, Robin PhD; Novack, Thomas A. PhD; Tulsky, David PhD; Kisala, Pamela MA; Luo, Chunqiao MS; Tang, Xinyu PhD

Section Editor(s): Caplan, Bruce PhD, ABPP; Bogner, Jennifer PhD, ABPP; Brenner, Lisa PhD, ABPP

Journal of Head Trauma Rehabilitation: March/April 2017 - Volume 32 - Issue 2 - p 125–133
doi: 10.1097/HTR.0000000000000207
Focus on Clinical Research and Practice, Part 1

Objective: To (1) identify groups of persons with traumatic brain injury (TBI) who differ on 12 dimensions of cognitive function: cognitive, emotional, and physical symptoms; personal strengths; physical functioning; environmental supports; and performance validity; and (2) describe patterns of differences among the groups on these dimensions and on participation outcome.

Setting: Three centers for rehabilitation of persons with TBI.

Participants: A total of 504 persons with TBI living in the community who were an average (standard deviation) of 6.3 (6.8) years postinjury and who had capacity to give consent, could be interviewed and tested in English, and were able to participate in an assessment lasting up to 4 hours.

Design: Observational study of a convenience sample of persons with TBI.

Main Measures: Selected scales from the Traumatic Brain Injury Quality of Life measures, Neurobehavioral Symptom Inventory, Economic Quality of Life Scale, Family Assessment Device General Functioning Scale, measures of cognitive function, Word Memory Test, and Participation Assessment with Recombined Tools–Objective (PART-O) scale.

Results: Cluster analysis identified 5 groups of persons with TBI who differed in clinically meaningful ways on the 12 dimension scores and the PART-O scale.

Conclusion: Cluster groupings identified in this study could assist clinicians with case conceptualization and treatment planning.

TIRR Memorial Hermann, Houston, Texas (Drs Sherer and Sander); Department of Physical Medicine and Rehabilitation, Baylor College of Medicine, Houston, Texas (Drs Sherer and Sander); Department of Pediatrics and Biostatistics, University of Arkansas for Medical Sciences, Little Rock (Drs Nick and Tang and Mss Melguizo and Luo); Department of Physical Medicine and Rehabilitation/Center for Trauma Rehabilitation, Harris Health System, Houston, Texas (Dr Sander); Department of Physical Medicine and Rehabilitation, Wayne State University Medical School, Rehabilitation Institute of Michigan, Detroit (Dr Hanks); Department of Physical Medicine and Rehabilitation, University of Alabama at Birmingham (Dr Novack); Department of Physical Therapy, University of Delaware College of Health Sciences, Newark (Dr Tulsky and Ms Kisala); and Kessler Foundation, West Orange, New Jersey (Dr Tulsky).

Deceased.

Corresponding Author: Mark Sherer, PhD, ABPP, FACRM, TIRR Memorial Hermann, 1333 Moursund St, Houston, TX 77030 (Mark.Sherer@memorialhermann.org).

The authors dedicate this article to the memory of Dr Todd G. Nick, a great statistician and a great friend.

This research was supported by US Department of Education National Institute on Disability and Rehabilitation Research grants H133B090023 and H133A120020.

The authors declare no conflicts of interest.

Article Outline

PERSONS WITH TRAUMATIC BRAIN INJURY (TBI) have most commonly been classified on the basis of injury severity1 as determined by ratings on the Glasgow Coma Scale (GCS).2 Ratings on the GCS indicate the depth and duration of impaired consciousness.2 These ratings inform early medical and surgical management of persons with TBI3,4 and assist with establishing prognosis.1 However, persons with similar or identical initial GCS scores can be quite heterogeneous with regard to underlying neuropathology and eventual behavioral outcomes.1 TBI can result in a wide range of cognitive, emotional, physical, and sensory impairments as well as a number of nonspecific neurobehavioral symptoms such as fatigue, irritability, restlessness, and others.5 Incidence and severity of these impairments and symptoms in the post-acute period are not strongly associated with initial injury severity. For these reasons, GCS scores and other indices of TBI severity may have limited utility in informing case conceptualization or guiding treatment in the post-acute period.

Clinicians conduct individualized assessments of persons with TBI to provide guidance for treatment planning and to monitor recovery.6 In addition to specific instruments such as functional rating scales, cognitive tests, and questionnaires, these assessments include review of medical records and interview with the person with injury as well as close others to capture specific concerns.7 Clinicians form case conceptualizations that include such elements as need for supervision, capacity for decision making, ability to return to work, and need for treatment based on knowledge of initial injury severity, medical complications and course of recovery, cognitive impairments, physical functioning, emotional distress, family concerns, and specific symptom complaints of the person with TBI.6 The adequacy of these case conceptualizations and their effectiveness in guiding treatment depends on the training and experience of the clinician. Judgments about how to combine assessment findings to form a case conceptualization and which treatments to provide are likely to be idiosyncratic. While the need for an integrated approach to treatment of cognitive, emotional, and other sequelae of TBI has been recognized for years,8–11 conceptual models that would guide this integrated approach are at an early stage of development.12

This investigation is part of a larger effort to develop an approach to classification of persons with TBI in the post-acute period and to provide guidance to clinicians in forming case conceptualizations based on these assessments while making decisions regarding treatment. Because of the limited usefulness of traditional severity indices for guiding treatment in the post-acute period, we decided to base our classification approach on a broad assessment of patient characteristics. In an initial phase of this effort,13 we administered a wide range of measures of cognitive functioning; performance validity; cognitive, emotional, neurobehavioral, and physical symptoms; personal strengths; physical functioning; and environmental supports to a large, multicenter cohort of persons with complicated mild, moderate, or severe TBI who were 6 or more months postinjury. Analysis of these data resulted in the identification of 12 dimensions that can be used to characterize the experience of persons with TBI in the post-acute period. Furthermore, we developed a methodology for calculating scores on each dimension for a given person with TBI. These dimensions are as follows: (1) Memory, (2) Cognitive Processing Speed, (3) Verbal Fluency, (4) Self-reported Cognitive Symptoms, (5) Independence and Self-esteem, (6) Resilience, (7) Emotional Distress, (8) Postconcussive Symptoms, (9) Physical Symptoms, (10) Physical Functioning, (11) Economic and Family Support, and (12) Performance Validity. Measures needed to calculate scores for each dimension are shown in Table 1.

The current investigation had 2 goals. First, we sought to identify clusters of persons with TBI who differed from each other on their profiles on the 12 dimensions listed earlier and to describe the patterns of differences among these patient clusters on the 12 dimension scores and other variables of interest. Second, we examined differences in participation outcomes among the cluster groupings.

Back to Top | Article Outline

METHODS

Participants for this study are the same as those for the Sherer et al13 investigation. Additional description of study methodology can be found in that article. Study participants were enrolled at 3 sites: (1) TIRR Memorial Hermann, (2) Rehabilitation Institute of Michigan, and (3) Spain Rehabilitation Center. Institutional review board approval was obtained for each site. All participants had definite medical evidence of TBI, capacity to give informed consent, and adequate cognition to complete all study measures in English. Participant age was limited to 18 to 64 years. Potential participants were excluded from study if it was perceived that functional status was attributable to a neurologic or psychiatric condition rather than to the TBI. To increase generalizability, persons with a history of neurologic or psychiatric disorder were not excluded if the symptoms presentation was consistent with TBI. The period of study was from October 2011 to September 2013.

The majority of participants had previously participated on other research projects at the study sites and had indicated interest in participating in additional research. Most participants had previously undergone inpatient TBI rehabilitation at one of the study sites. Potential participants were selected from registries at each site and contacted to determine interest in this project. While participants were included on registries due to systematic, often consecutive, recruitment into TBI-related studies, their selection for this study was not systematic so that they may be considered to constitute a convenience rather than consecutive sample. Collaborating institutions received institutional review board approval for this project at their site.

Back to Top | Article Outline

Data collection

Data were collected by review of pertinent medical records, interview, and administration of test and questionnaires. Data abstracted from medical records included age, date of injury, GCS score at emergency department presentation, time from injury until commands were consistently followed (TFC), duration of posttraumatic amnesia (PTA), and hospital length of stay for both acute and rehabilitation hospitalizations. Interviews were conducted to collect data regarding race/ethnicity, gender, level of education, and history of alcohol and drug use.

While 36 questionnaires or tests were administered to facilitate development of the 12 dimensions, only 18 of these are required for calculation of dimension scores for each participant (see Table 1). The TBI Quality of Life (TBI-QOL) measurement system instrument is the source of 9 of the questionnaires. The TBI-QOL14 is a comprehensive patient reported outcomes measure for persons with TBI. The TBI-QOL consists of 22 item banks or subscales, and the 9 selected for use in this study are shown in Table 1. Areas measured are general concerns regarding cognitive functioning, self-esteem, independence, resilience, anxiety, emotional and behavioral dyscontrol, headache, pain interference, and upper extremity function. Other questionnaires that determine dimension scores are the Neurobehavioral Symptom Inventory,15 the Economic Quality of Life Scale,16 and the Family Assessment Device General Functioning Scale.17 The Neurobehavioral Symptom Inventory is a self-report scale of 22 symptoms that are characteristic of postconcussion syndrome. The Economic Quality of Life Scale assesses financial barriers and facilitators that affect community participation outcomes. The Family Assessment Device General Functioning Scale assesses overall health of family functioning.

Cognitive tests that contribute to dimension scores include measures of working memory (Wechsler Adult Intelligence Scale–IV Letter Number Sequencing),18 list learning (Rey Auditory Verbal Learning Test),19 speed of cognitive processing (Trails A from the Trail Making Test,20 Wechsler Adult Intelligence Scale–IV Coding),18 and verbal fluency (FAS test).21 Performance validity was assessed with the Word Memory Test.22 The Participation Assessment with Recombined Tools-Objective (PART-O)23 was used as a measure of community integration outcome. Additional description of all study measures may be found in the Sherer et al13 study.

Scores for the 18 measures that contribute to the 12 dimensions are typically expressed using a range of metrics (scaled scores, standard scores, T-scores, percentile ranks), and these measures were not all normed on the same normative sample. To generate scores that were on a common metric (mean = 0, standard deviation = 1) for all 12 dimensions, scores for dimensions defined by 1 measure were normalized and used as dimension scores and, for dimensions defined by 2 measures, principal components analysis was performed and the normalized first principal component score for each pair of measures was used as score for that dimension.

Back to Top | Article Outline

Data analysis

All the data were analyzed using statistical software R v3.1.3.24 Cluster analysis was performed to categorize all participating individuals into clusters using the scores from 12 dimensions. The K-means method was applied at each generation of clusters to merge individuals into larger clusters for minimizing the within-cluster sum of squares. The optimal number of clusters was determined and validated using the variance ratio criterion (VRC)25 and gap statistics.26 The VRC is the most commonly used partitional clustering strategy to measure the adequacy or similarity in which a given data set can be clustered. Gap statistics were used for estimating the optimal number of clusters and were calculated using the R package NbClust.27 The reliability of the primary K-means solution was assessed by comparing findings with a hierarchical cluster analysis performed using the Ward minimum variance method and to split-half analysis using the K-means method.

Summary statistics were expressed as means (standard deviations) for continuous variables and percentages (counts) for categorical variables. The mean of each dimension and its 95% confidence interval were calculated for each cluster. For continuous variables, analysis of variance models were fitted to investigate overall differences among the clusters and, for variables with overall significant differences, pairwise comparisons between each pair of clusters were conducted using Tukey post hoc tests. For categorical variables, overall significance for comparisons among the clusters was assessed using χ2 tests. Pairwise differences for variables with overall significance differences were assessed between pairs of clusters using χ2 post hoc tests. P values of .05 or less were considered statistically significant.

Back to Top | Article Outline

RESULTS

Data collection during the study period at the 3 sites resulted in a study sample of 504 participants meeting inclusion criteria and producing usable data. Characteristics of the entire study sample are provided in the Sherer et al13 study.

Back to Top | Article Outline

Determination and reliability of 5 clusters

We computed the VRCs for assessing the performance of 3 to 6 clusters, and the optimal number of clusters was determined to be 5 because of the smallest within-cluster variation and largest between-cluster variation. The optimal number of 5 clusters was also assessed using the gap statistics, with 5 clusters being the smallest number of clusters such that the critical value is zero or less. Cluster 1 comprised 126 participants, cluster 2 of 93 participants, cluster 3 of 125 participants, cluster 4 of 114 participants, and cluster 5 of 46 participants. Reliability analyses using the Ward minimum variance method and split-half analysis revealed agreement in cluster membership of 76% and 86%, respectively, indicating good to very good agreement. The possible clinical significance of the groupings can be interpreted on the basis of descriptive data provided later.

Back to Top | Article Outline

Description of the 5 clusters

Figure 1 shows the profiles on the 12 dimensions for the 5 groups. Descriptive data for the 5 clusters (groups) of persons with TBI are provided in Table 2. The 5 groups did not have overall differences on the proportion of males, proportions of GCS categories, age, or time since injury (all Ps > .05). However, overall differences were found for race/ethnicity (P < .001), years of education (P < .001), days to follow commands (TFC; P = .01), and duration of PTA (P < .001). Group 4 differed from groups 1 (P < .001) and 3 (P = .05) by having a greater proportion of black participants. For years of education, group 1 had a greater number of years than group 2, 4, or 5 (all Ps ≤ .01). For TFC, persons in group 2 took longer to recover ability to follow commands than persons in group 3 (P = .004), indicating greater injury severity. For PTA duration, persons in group 2 had longer PTA duration than those in group 1, 3, or 4 (all Ps ≤ .01), suggesting that persons in group 2 generally had more severe injuries than persons in other groups.

Table 3 shows the comparisons on the normalized dimension scores for the 5 groups. Overall significance was found for each of the 12 dimension scores. All pairwise contrasts are provided in the table. Significant pairwise differences among groups are too numerous to be fully described here, but we will summarize some key findings. For cognitive functioning, groups 1 and 3 generally showed the most intact functioning, with group 4 at an intermediate level and groups 2 and 5 being most impaired. Group 1 generally endorsed strengths in self-perception and environmental support (Independence and Self-esteem, Resilience, Economic and Family Support) and denied negative signs and symptoms (Cognitive Symptoms, Emotional Distress, Postconcussive Symptoms, Physical Symptoms, limitations in Physical Functioning) while showing performance validity. Groups 2 and 3 endorsed similar, intermediate levels of strengths, environmental support, and physical functioning. While both were still intermediate among the 5 groups, group 2 reported fewer cognitive, postconcussive, and physical symptoms and less emotional distress than group 3. Both groups had adequate performance validity. Groups 4 and 5 generally reported the highest levels of symptoms and the lowest levels of strengths among the 5 groups. Group 5 was the only group for which the average of the 3 “easy” tests for the Word Memory Test fell below the cutoff for valid performance. As a group, group 5 showed poorer performance validity than each of the other 4 groups.

Comparison of the 5 groups on community integration/participation outcome is shown in the final row of Table 3. Persons in groups 1 and 3 had the highest levels of community integration, whereas persons in groups 2, 4, and 5 had lower levels of participation.

Back to Top | Article Outline

DISCUSSION

Results of this investigation indicated that persons in the post-acute period following TBI can be clustered into 5 groups based on their scores on 12 dimensions of experience or functioning that include cognitive functioning, various symptoms including cognitive complaints and emotional distress, personal strengths, environmental supports, and performance validity. These 5 groups make “clinical sense” in this regard that they map on to types of clients typically seen in the rehabilitation setting. Group 1 can be thought of as an excellent recovery group. Persons in group 1 had good cognitive function; few concerns about cognitive, emotional, or physical functioning; good self-perceived strengths; and good environmental supports. As a group, they performed in the valid range on a performance validity measure. They had higher levels of preinjury education than some groups, suggesting that they may have functioned at a relatively high level before their injuries. Along with group 3, group 1 had more favorable participation/community integration outcomes than groups 2, 4, and 5.

Group 2 represents persons who remain with substantial cognitive impairment after their injuries and yet report fairly low levels of concern about cognitive function, emotional distress, and postconcussive symptoms. They have intermediate levels of positive self-perceptions and environmental support. Persons in group 2 sustained more severe injuries than persons in some other groups as indicated by a longer duration of PTA. Group 2 could be composed of persons who have made a relatively good adjustment to substantial cognitive impairment after TBI, persons who have some impairment of self-awareness after TBI, or some of both.

Group 3 is a group with strong cognitive recovery, comparable with group 1, but that still reports significant concern about cognitive functioning, emotional distress, postconcussive symptoms, and physical symptoms. Given these concerns, it is surprising that group 3 had favorable participation outcomes. Persons in group 3 may perceive TBI as a threat to the self so that they continue to report high levels of symptoms and distress, although others, based on strong cognitive recoveries and good participation outcome, may perceive that they are functioning well. This scenario could contribute to conflict with healthcare providers or family. Indeed, persons in group 3 report lower levels of environment support than those in group 1, although higher levels than those in groups 4 and 5.

Group 4 members show intermediate cognitive recovery with better functioning than groups 2 and 5 but poorer functioning than groups 1 and 3. Still this group reports the higher levels of concern about cognitive functioning, emotional distress, postconcussive symptoms, physical symptoms, and physical functioning than any group other than group 5. They also report low levels of personal strength and environmental support. However, they differ from group 5 in having an average validity indicator score in the valid range. Group 4 has a greater proportion of African Americans than some other groups.

Group 5 is notable for having a mean score for the Word Memory Test “easy test” that falls in the range consistent with invalid performance validity. The mean (standard deviation) percentage correct for the Immediate Recognition, Delayed Recognition, and Consistency scores for the Word Memory Test for group 5 was 66.3 (12.0), which is substantially below the cutoff for valid performance, whereas the lowest mean for the remaining groups was 90.5 (9.8), a score significantly above the cutoff. Consistent with the typical pattern for persons showing poor performance validity, persons in group 5 produced scores that, if valid, would indicate greater cognitive impairment (except for group 2), more severe cognitive and emotional symptoms, and poorer self-reported strengths, physical functioning, and environment supports (other than group 4) than other groups.

If additional persons with TBI are given the battery of tests and questionnaires needed to generate the 12 dimension scores, these persons can be classified on the basis of which group they are most similar to using straightforward mathematical calculations. In this way, the work to this point forms the basis for a classification scheme for persons with TBI who are in the post-acute period of recovery. While additional investigation is needed, the brief descriptions of the 5 groups provided earlier might form the beginnings of case conceptualizations for each of the groups that could be used to facilitate feedback to the person with injury, family/close others, healthcare providers, and others. The patterns of difficulties indicated by these scores could have implications for treatment planning. As an example, persons in group 2 would likely be candidates for cognitive rehabilitation therapies including compensatory strategies whereas persons in group 3 would be more likely to benefit from psychotherapy or other supportive interventions.

The current scheme that supports assignment of persons with TBI to one of only 5 groups is a clear oversimplification of the very heterogeneous manifestations of TBI and the diverse group of persons who sustain TBI. However, it is possible that even this simplified model could provide information to clinicians that would assist with case conceptualization and treatment planning. Additional work is needed to develop a more sophisticated scheme that permits a more fine-grained, individualized understanding of persons with TBI. We consider the current scheme to be an important initial step toward a more clinically relevant tool to assist clinicians with patient classification, case conceptualization, and treatment planning.

As with any investigation, the current study has limitations. Persons with TBI participated in this study an average of 6.3 years postinjury, although about 1 in 5 was within a year of injury. Thus, many of these participants would be unlikely to be involved in treatment and many might not have the opportunity to seek treatment even if it was desired. They may be more likely to be adjusted to or resigned to their current status than persons closer to injury who have more opportunity to receive interventions. Persons with very poor outcomes were excluded from study because of the requirement that participants have capacity to give their own consent and persons with very favorable outcomes may have been excluded, as most participants were sufficiently impaired early after injury to require inpatient rehabilitation services.

Back to Top | Article Outline

REFERENCES

1. Saatman KE, Duhaime AC, Bullock R, Maas AI, Valadka A, Manley GT. Classification of traumatic brain injury for targeted therapies. J Neurotrauma. 2008;25:719–738.
2. Teasdale G, Jennett B. Assessment of coma and impaired consciousness. A practical scale. Lancet. 1974;2(7872):81–84.
3. Bullock MR, Chesnut R, Ghajar J, et al Surgical management of acute subdural hematomas. Neurosurgery. 2006;58(suppl 3):S16–S24.
4. Brain Trauma Foundation. Guidelines for the management of severe traumatic brain injury. J Neurotrauma. 2007;24(suppl 1):S1–S106.
5. Sherer M, Madison CF, Hannay HJ. A review of outcome after moderate and severe closed-head injury with an introduction to life care planning. J Head Trauma Rehabil. 2000;15:767–782.
6. Sherer M, Novack T. Neuropsychological assessment for monitoring recovery after traumatic brain injury and making treatment recommendations in adults. In: Prigatano GP, Cullum M, Pliskin N, eds. Demonstrating the Utility and Cost Effectiveness in Clinical Neuropsychology. New York, NY: Psychology Press; 2003:39–60.
7. Bergquist TF, Yutsis M, Micklewright JL. Comprehensive assessment. In: Sherer M, Sander AM, eds. Handbook on the Neuropsychology of Traumatic Brain Injury. New York, NY: Springer; 2014:77–94.
8. Ben-Yishay Y. Reflections on the evolution of the therapeutic milieu concept. Neuropsychol Rehabil. 1996;6:327–343.
9. Ben-Yishay Y, Rattok J, Lakin P, Piasetsky E, Ross B, Silver S. Neuropsychological rehabilitation: quest for a holistic approach. Semin Neurol. 1985;5:252–259.
10. Prigatano GP, Fordyce DJ, Zeiner HK, Roueche JR, Pepping M, Wood BC. Neuropsychological rehabilitation after closed-head injury in young adults. J Neurol Neurosurg Psychiatry. 1984;47:505–513.
11. Prigatano GP, Fordyce DJ, Zeiner HK, Roueche JR, Pepping M, Wood BC. Neuropsychological Rehabilitation After Brain Injury. Baltimore, MD: John Hopkins University Press; 1986.
12. Dams-O'Connor K, Gordon WA. Integrating interventions after traumatic brain injury: a synergistic approach to neurorehabilitation. Brain Impairment. 2013;14:51–62.
13. Sherer M, Sander AM, Nick TG, et al TA. Key dimensions of impairment, self-report, and environmental supports in persons with TBI. Rehabil Psychol. 2015;60:138–146.
14. Tulsky DS, Kisala PA, Victorson D, et al TBI-QOL: development and calibration of item banks to measure patient reported outcomes following traumatic brain injury [published online ahead of print April 19, 2015]. J Head Trauma Rehabil. doi:10.1097/HTR.0000000000000131.
15. Cicerone KD, Kalmar K. Persistent postconcussion syndrome: the structure of subjective complaints after mild traumatic brain injury. J Head Trauma Rehabil. 1995;10(3):1–17.
16. Tulsky DS, Kisala PA, Lai J-S, Carlozi N, Hammel J, Heinemann AW. Developing an item bank to measure economic quality of life for individuals with disabilities. Arch Phys Med Rehabil. 2015;96:604–613.
17. Epstein NB, Baldwin L, Bishop DS. The McMaster family assessment device. J Marital Fam Ther. 1983;9:171–180.
18. Wechsler D. Wechsler Adult Intelligence Scale–Fourth Edition (WAIS-IV). San Antonio, TX: Pearson; 2008.
19. Rey A. L'examin Clinique en psychologie. Paris, France: Presses Universitaires de France; 1958.
20. Reitan RM, Wolfson D. The Halstead-Reitan Neuropsychological Test Battery: Theory and Clinical Interpretation. Tucson, AZ: Neuropsychology Press; 1985.
21. Gladsjo JA, Schuman CC, Evans JD, Peavy GM, Miller SW, Heaton RK. Norms for letter and category fluency: demographic corrections for age, education, and ethnicity. Assessment. 1999;6:147–178.
22. Green P, Allen L M, Astner K. The Word Memory Test: A User's Guide to the Oral and Computer-Administered Forms, U.S. Version 1.1. Durham, NC: Cognisyst; 1996.
23. Whiteneck GG1, Dijkers MP, Heinemann AW, et al Development of the participation assessment with recombined tools-objective for use after traumatic brain injury. Arch Phys Med Rehabil. 2011;92(4):542–551.
24. R Core Team. R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing. https://http://www.r-project.org. Accessed July 21, 2015.
25. Caliński T, Harabasz J. A dendrite method for cluster analysis. Commun Stat Theory Methods. 1974;3:1–27.
26. Tibshirani R, Walther G, Hastie T. Estimating the number of clusters in a data set via the gap statistic. J R Stat Soc Ser B Stat Methodol. 2001;63:411–423.
27. Charrad M, Ghazzali N, Boiteau V, Niknafs A. NbClust: an R package for determining the relevant number of clusters in a data set, J Stat Softw. 2014;61(6):1–36.

classification; environment; outcome; quality of life; traumatic brain injury; treatment

Copyright © 2017 Wolters Kluwer Health, Inc. All rights reserved.