Clinical manifestations of traumatic brain injury (TBI) are heterogeneous, varying widely from person to person, and often classified as mild, moderate, or severe based on the Glasgow Coma Scale (GCS) score (1). TBI is associated with a broad range of signs and symptoms (2,3). There is a lack of robust, objective criteria and more mild TBI diagnosis relies largely on self-reported symptoms. These mild TBIs have been called ‘invisible injuries’ given the lack of objective diagnostic criteria and may be unrecognized in certain clinical settings, particularly in multi-trauma cases (4). Variable symptom presentation, response to treatment, and recovery can lead to challenges in managing TBI.

In addition to GCS scoring, three cardinal signs or symptoms of TBI have historically contributed to qualifying the severity of injury including: loss of consciousness, alteration of consciousness/mental state (i.e., feeling disoriented or confused), and post-traumatic amnesia (5). Categorizing TBI patients in this manner fails to account for the complex heterogeneity among individuals and, to date, clinical treatment trials based on this classification system have failed to optimally translate to effective treatment and recovery in the real world (6,7). Patient classification in traditional severity categories may inhibit the discovery of effective therapies that improve outcomes based on more granular clinical profiles (7,8). For these reasons, more refined evidence-based approaches to classification and treatment are needed.

The limitations of the current classification system allow space to find alternative methods that have the potential to stratify TBI patient subpopulations for better targeted treatment. Unsupervised machine learning has been proven as a promising method for discovering patient phenotypes, improving upon the classification and identification of patient subpopulations for several diseases (9,10), including TBI (11–13). The goal of this study was to determine if unsupervised machine learning could identify TBI patient phenotypes with unique clinical and outcome profiles in national survey data.

There were 1,364 respondents in the study who reported a head injury in the past 12 months. The sample of respondents with a head injury was composed of 212 non-cases (respondents who had an affirmative response to a head injury but reported no symptoms) and 1,152 respondents with symptomatic head injury that were then classified as a TBI (an affirmative response to a head injury and at least one symptom).

The median number of symptoms reported was four (data not shown). Among respondents who self-reported symptoms, headache was the most common symptom (85%, n = 980), followed by confusion (55%, n = 631), and balance problems (43%, n = 501) (Figure 1). Loss of consciousness (LOC) was reported by 19% (n = 214) of respondents with a TBI. Respondents who self-reported only headache (n = 163) were the most common, followed by respondents who experienced headache and confusion together (n = 40) and headache and change in temperament together (n = 30). There were 12 respondents who self-reported all 12 signs and symptoms.

The PAM algorithm grouped respondents into five clusters (TBI phenotypes A-E, Table 1 and Supplemental Figure 1S). Phenotype C had the highest prevalence of symptoms (i.e., >50% of respondents in this cluster self-reported 11 out of the 12 symptoms) and included two of the cardinal symptoms (i.e., LOC and difficulty remembering what happened just before/after injury). Additionally, this phenotype had the highest median number of symptoms (n = 8, Table 1). Phenotype C respondents had a higher prevalence of medical evaluation (58.1%) for their head injury, a higher prevalence of symptoms that were not currently resolved or resolved in 8+ days (60.5%), greater impact on social and work functioning (Moderately/Quite A Bit/Extremely = 43.2% and 46.9%, respectively), and a lower prevalence of symptoms that resolved in one day (4.5%) compared to other phenotypes (Table 2). Although demographic characteristics were not involved in creating cluster phenotypes, Phenotype C had a higher prevalence of respondents who were a race/ethnicity other than non-Hispanic white (45.6%), had an education of high school or less (52.4%), and were renting a residence (45.5%) when compared to other phenotypes (Table 3).

Phenotype E was composed of all the non-cases (i.e., respondents that reported a head injury with no symptoms) (Table 1). Compared to other phenotypes, Phenotype E had a lower prevalence of those who were evaluated (11.2%) for their head injury and had lower impact on their social and work functioning (Moderately/Quite A Bit/Extremely = 1.2% for both) (Table 2), and a higher prevalence of respondents who were aged 55+ (39.9%) and had a Bachelor’s degree or more for education (36.2%) (Table 3).

Phenotype D was the largest cluster and had the second highest median number of symptoms (n = 5, Table 1). Phenotype D was most similar to the average respondent in the sample included in the cluster analysis (please refer to the ‘Total’ column in Tables 2 and 3 to see the overall TBI cluster sample): Phenotype D respondents had the second highest prevalence of medical evaluation (30.7%) for their head injury, symptoms that were not currently resolved or resolved in 8+ days (38.1%), and impact on social and work functioning (Moderately/Quite A Bit/Extremely = 23.1% and 23.6%, respectively) (Table 2). Additionally, similar to the total sample, Phenotype D respondents were 51.8% male, 44.9% were between 18–34 years of age, 69.1% non-Hispanic white, 64.9% had some college education or more, and 56.6% owned a residence (Table 3).

Phenotypes A and B had few symptoms (where >50% of respondents in a cluster self-reported a particular sign or symptom), which included headache alone (Phenotype A) or confusion plus headache (Phenotype B) (Table 1). Correspondingly, Phenotypes A and B had a high prevalence of respondents who had one day of symptoms (48.1% and 35.4%, respectively) and a lower prevalence of respondents who were evaluated (12.4% and 16.7%, respectively) for their head injury, had symptoms that were not currently resolved or resolved in 8+ days (12.7% and 24.0%, respectively), and social (Moderately/Quite A Bit/Extremely = 1.4% and 2.5%, respectively) and work impairment (Moderately/Quite A Bit/Extremely = 3.4% and 5.3%, respectively) (Table 2). However, these phenotypes differed by demographics. Phenotype A had a higher prevalence of respondents who were younger (18–35 years old, 49.8%), while Phenotype B had the highest prevalence of males (65.5%), respondents who were non-Hispanic white (79.1%), and respondents who owned a residence (75.0%) (Table 3).

Table 4 displays the results of the logistic regressions, ordered according to clinical severity (most to least severe). The magnitude of effect (odds ratio) was in the expected direction (e.g., more severe phenotypes demonstrated worse outcomes). Compared to Phenotype A, Phenotype C was associated with higher odds of medical evaluation (OR = 9.8, 95% CI = 5.8–16.6), symptoms that were not currently resolved or resolved in 8+ days (OR = 10.6, 95% CI = 6.2–18.1), higher impacts on social (OR = 54.7, 95% CI = 22.4–133.4) and work (OR = 25.4, 95% CI = 11.2–57.2) functioning, and lower odds of symptoms that resolved in one day (indicator of less severity) (OR = 0.1, 95% CI = 0.03–0.1). Displaying a similar pattern to Phenotype C, Phenotype D was also associated with higher odds of medical evaluation (OR = 3.1, 95% CI = 1.9–5.2) and symptoms that were not currently resolved or resolved in 8+ days (OR = 4.2, 95% CI = 2.6–7.0), higher impacts on social (OR = 21.6, 95% CI = 8.8–52.7) and work (OR = 8.8, 95% CI = 3.9–19.9) functioning, and lower odds of symptoms that resolved in one day (OR = 0.3, 95% CI = 0.2–0.4). Phenotype B was associated with higher odds of symptoms that were not currently resolved or resolved in 8+ days (OR = 2.2, 95% CI = 1.1–4.1) and lower odds of symptoms that resolved in one day (OR = 0.6, 95% CI = 0.4–0.9).

In this study, we identified five TBI patient phenotypes using unsupervised machine learning, based solely on self-reported signs and symptoms from national survey data. Each TBI phenotype demonstrated unique clinical characteristics that corresponded to specific differences in the severity of outcomes and demographic profiles. This work directly contributes to a growing foundation of research establishing feasibility of precision diagnostics in TBI. It also supports personalized care with better prognostication of outcomes as compared to traditional classification schemes.

One phenotype (Phenotype C) in this study was characterized by a high number of symptoms and worse outcomes. Additionally, this phenotype also had the highest percentages of respondents with two cardinal symptoms (i.e., LOC and difficulty remembering what happened just before/after injury). Similar to our data, past research has shown that these cardinal symptoms are associated with worse outcomes (e.g., longer functional recovery, increased symptom duration, cognitive impairment, etc.) among TBI patients (20–22). This phenotype was also disproportionately populated by people who were non-Hispanic white, had less education, and rented a residence. This association suggests, as previous literature (23–25) has reported, that TBI may disproportionality impact marginalized groups. Previous studies examining social determinants of health (non-medical factors that influence health outcomes) (26) demonstrate that certain factors (e.g., disability, race, insurance status) can negatively impact an individual’s recovery and outcome following TBI (27,28). Recognizing and seeking treatment when a suspected brain injury is sustained are important, and research supports that patients who receive clinical care sooner recover faster (29,30). Even in those who seek care, not all patients with head injury are evaluated for TBI (31,32) and may not be diagnosed or may go untreated (33–35). Future studies could assess and address structural inequalities related to phenotypic diagnosis and subsequent treatment for patients with TBI.

Another distinct cluster demonstrated a phenotype (Phenotype A) where respondents self-reported headache as their only symptom where > 50% of respondents self-reported a particular sign or symptom in that cluster. This finding is consistent with other TBI studies (3,36,37) that also demonstrate a high prevalence of headache among patients. This phenotype likely represents less severe TBIs based on its lower association with worse outcomes (i.e., medical evaluation, symptoms that were not currently resolved or resolved in 8+ days, impact on social and work functioning) and reflects a unique group of patients who may benefit from a streamlined, simple treatment regimen that largely targets headache reduction. However, further research is warranted to ensure resolution of symptoms and no long-term sequelae in this group.

While clinical severity (e.g., GCS or the presence or absence of the cardinal TBI signs/symptoms) has been demonstrated as strong predictors of worse outcomes after TBI (20,22,38), estimates of these measures are often inaccurate due to recall bias, difficulty in interpretation or collection of these data, and have poor concordance or conflicting associations between them (39,40). Additionally, previous literature (including clinical trials) that categorize individuals with TBI using these standard severity scales have demonstrated limited utility in terms of stratifying TBI patients for effective treatments (6,7). For example, past studies (11,41) that have examined GCS versus machine learning have demonstrated that TBI severity using GCS classification alone may not be optimal or granular enough to capture the complexity of TBI. In comparison to using GCS or to traditional statistics, machine learning can identify non-linear relationships that can reveal meaningful patterns and insights, and create novel representations of clinical profiles that may be better suited to addressing multifaceted and complex health conditions and target treatments (42,43).

Strengths of this work include an innovative machine learning algorithm for classification applying a data-driven approach to identify TBI phenotypes that have a distinct clinical profile. Another strength lies in the use of a large dataset derived from a random-digit-dial (RDD) telephone survey. However, this study does have limitations. Respondents who self-reported a head injury that was then classified as a TBI may not have been clinically evaluated or diagnosed and results should be interpreted with this understanding. However, the data collected as part of this study were intended for surveillance, and limiting the inclusion of individuals who sought care or received a TBI diagnosis would also likely restrict the data to those with 1) a more severe presentation of TBI or 2) ease of accessibility to care (i.e., potentially excluding individuals who are socially disadvantaged) (14). Additionally, self-report data are subject to issues such as recall bias, social desirability bias, and under or over-reporting. Another limitation is that the input data for the clusters were limited to individual signs and symptoms. Other studies (11–13,41) examining TBI phenotypes or endotypes have included biological biomarkers, clinical findings, neuroimaging, laboratory tests, or therapies to create clusters. These data were not available in our study, and future research can consider using machine learning in comprehensive datasets that link biomarkers with symptom profiles and other characteristics to even better inform phenotypes. Prospective controlled studies of phenotype-directed diagnosis and treatment are potentially important next steps for research. Finally, the pilot NCSS survey, from which this study’s data is drawn, was designed to support TBI case ascertainment and measure development. Estimates were weighted with the intent of exploring differences in TBI prevalence derived from different databases, using different case definitions. Weighted estimates from this pilot work were not specifically intended to produce nationally representative estimates of TBI.

Classifying TBI phenotypes using machine learning, as demonstrated here by five distinct phenotypes, may inform next steps in research to focus on their utility in clinical diagnosis and symptom-based treatment for faster patient recovery through a more personalized approach.