Stratifying Traumatic Brain Injury Mortality and Readmission Risk: An Unsupervised Machine Learning Approach

Abstract

Background:

Traumatic brain injury (TBI) presents several challenges due to its heterogeneity in care trajectories. Many who survive the acute phase remain at risk for long-term disability, hospital readmission, and mortality. Predicting these outcomes, particularly for those with severe injuries requiring intensive care, involves a complex interplay of clinical and patient-specific factors. This study used unsupervised machine learning (ML) to identify patient subgroups with clinically meaningful mortality and readmission risk profiles.

Methods:

A retrospective analysis was conducted using electronic health records of 263 adult patients admitted to the intensive care unit at a Level 1 trauma center diagnosed with TBI between 2016 and 2023, identified through diagnostic codes (73.7% male, median age 41 years [interquartile range: 26–60]). In total, 187 patients (71.1%) had severe TBI (Glasgow Coma Scale [GCS] 3–8), 19 (7.2%) had moderate TBI (GCS 9–12), and 57 (21.7%) had mild TBI (GCS 13–15). Clinical, demographic, and diagnoses from imaging data were reduced using Uniform Manifold Approximation and Projection. Agglomerative clustering was performed to generate subgroups. Cluster differences were evaluated using univariate analyses, and time-to-event analysis was used to stratify mortality and readmission risk across clusters.

Results:

Three clusters were identified based on their clinical features (silhouette score: 0.58). Cluster A included younger, lower-comorbidity patients with severe TBI, with excellent survival yet the highest readmission, especially after 90 days (p < 0.01), likely influenced by diffuse axonal injury. Cluster B consisted of middle-aged patients also with severe TBI, yet the highest risk of mortality (p < 0.001), driven by the prevalence of subarachnoid hemorrhage with concurrent intra- and extra-axial injuries. Cluster C comprised older patients with mild injuries, primarily concussions, but with a high comorbidity burden and significant readmissions, especially within 90 days (p < 0.05).

Conclusion:

Unsupervised ML can be a promising method to stratify TBI patients and gain more insights into post-TBI mortality and readmission risk for personalized patient planning. Future research should investigate the causes of readmission and validate these clusters with larger, multicenter datasets.