Unsupervised Machine Learning Reveals Clinically Relevant Phenotypes in the Prospective Transforming Research and Clinical Knowledge in Geriatric Traumatic Brain Injury (TRACK-GERI) Study

Introduction: Older adults are particularly vulnerable to traumatic brain injury (TBI) due to distinct pathophysiology and variable baseline health. Geriatric TBI outcomes are heterogeneous with little known about the interaction between baseline health and injury-related factors. The objective of the present study was to apply an unbiased, data-driven approach to identify clinically relevant geriatric TBI phenotypes.

Methods: TRACK-GERI is a two-site prospective cohort study of adults age ≥ 65y presenting acutely to the emergency department with head trauma who undergo head CT to rule out intracranial trauma. Baseline variables (N=107) across domains of sociodemographics, acute injury and neuroimaging features, baseline functional and cognitive measures, comorbidities, and medications were assessed. Dimensionality reduction and variable prioritization were achieved through Principal Component Analysis. K-Means clustering was performed. 12-month Glasgow Outcome Scale-Extended (GOS-E) and mortality were compared between clusters using linear and Cox regression models, respectively.

Results: We identified three clusters in this study. Cluster A (N=139) consisted of patients with mild TBI (median GCS [IQR]: 15 [15,15]) and robust baseline health; Cluster B (N=67) comprised patients with mild TBI (median GCS [IQR]: 15 [15,15]) and poor baseline health; Cluster C (N=28) included patients with more severe TBI (median GCS [IQR]: 14 [10,14]). Functional Activities Questionnaire, Groningen Frailty Indicator, and Charlson Comorbidity Index, but not age, were the most important explanatory features informing clustering.

While Cluster A and Cluster B had similar acute injury and neuroimaging features, Cluster B had worse baseline function/cognition across nine metrics and twice as many comorbidities. Cluster A achieved significantly better 12-month GOS-E than Cluster B (P < 0.0001) and Cluster C (P < 0.0001) and had lower 12-month mortality (B vs. A, HR 7.3, 95% CI 2-26.6, P=0.003; C vs. A, HR 24.7, 95% CI 7-86.7, P< 0.0001).

Conclusion : A data-driven approach identified clinically relevant geriatric TBI baseline phenotypes that are significantly associated with 12-month functional outcome and mortality. External validation and predictive modeling approaches are now warranted to inform patient and family counseling and precision rehabilitation.