While most children recover from pediatric "mild" traumatic brain injury (pmTBI), up to one-third experience persisting symptoms after concussion (PSaC) that interfere with school, social, and emotional functioning. Clinicians face the dual challenge of low PSaC rates and nonspecific symptom rating even among uninjured peers, making accurate prognosis especially challenging. We used machine learning in a large prospective pmTBI cohort (N = 321) to identify indicators of poor recovery at 4 months and 1-year postinjury. Participants completed comprehensive assessments within 11 days of injury that spanned multiple domains including demographics, injury-related factors, child and parent symptom ratings, cognitive tasks, objective performance, and symptom provocation on neurosensory tasks. Variable importance scores and 90% confidence intervals from 150 bootstraps were used to identify the best-performing assessments within each domain and then integrated into a combined model to assess overall prognostic value. Key findings suggest that retrospective self-report of symptom burden and vulnerability to symptom provocation were the most robust predictors of PSaC at both 4 months and 1-year postinjury, outperforming established risk scores. Other important measures included near point convergence, long-term memory, household size, and parental report of child symptom burden. Retrospective symptom burden and acute symptom provocation during simple tasks alongside established risk scores hold promise for improving early risk stratification and guiding individualized care. Additional research is needed to determine how to best integrate these measures into clinical workflows and validate their utility across diverse settings.
Keywords: machine learning; mild traumatic brain injury; outcome classification; pediatric; persisting symptoms after concussion.