Prediction of seizure risk after repetitive mild traumatic brain injury in childhood.

Objective: Despite the known negative physiological impact of repeated mild head trauma events, their multiplicative impact on long-term seizure risk remains unclear. The objective of this study was to evaluate how multiple mild traumatic brain injuries (mTBIs) impact long-term seizure risk by testing 3 distinct machine learning approaches. Baseline and injury-specific characteristics were incorporated to enhance prognostication of individual seizure risk.

Methods: Children with at least 1 mTBI event without prior evidence of seizure or antiepileptic drug treatment, from 2003 to 2021, were identified from a nationally sourced administrative claims database. The primary outcome of interest was a seizure event after mTBI, defined by qualifying principal diagnosis codes. Time-varying multivariable Cox regression was used to assess the impact of repeated mTBI.

Results: A total of 156,118 children (mean age 11.7 ± 4.7 years) were included, with a median follow-up duration of 22.6 months (IQR 9.2-45.4 months). Among patients who experienced seizure after mTBI, the median time to seizure was 306 days. Seizures among those with radiographic findings and/or loss of consciousness occurred earlier (median time to seizure 112.5 days [imaging findings only, IQR 5-526.25 days], 80 days [loss of consciousness only, IQR 7-652 days], 22 days [both, IQR 5-192 days]). Both mTBI without and with short-term loss of consciousness resulted in increasing seizure risk with repeated trauma (HR 1.196, 95% CI 1.082-1.322; HR 2.025, 95% CI 1.828-2.244; respectively). The random survival forest approach achieved fixed-time areas under the receiver operating characteristic curve of 0.780 and 0.777 at 30 and 90 days after mTBI, and children predicted at high risk by the final model experienced a significantly higher burden of early seizure after mTBI (46.7% within the first 30 days vs 17.7% and 19.9% of children at low and medium risk). A simplified model using the top 12 contributing features achieved 95% of the full model's performance in the validation set.

Conclusions: A novel machine learning model was developed and validated for personalized prediction of long-term seizure risk following multiple mTBIs. Model performance remained robust with a limited feature set, suggesting the feasibility of real-time incorporation into clinical workflows for individualized prognostication following each repeat mTBI event. In children predicted to be at high risk, early intervention should be considered.