Epidemiological Forecasting of Epilepsy After Traumatic Brain Injury
Abstract number :
1.153
Submission category :
16. Epidemiology
Year :
2024
Submission ID :
960
Source :
www.aesnet.org
Presentation date :
12/7/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Eamonn Kennedy, PhD – University of Utah
Shashank Vadlamani, MSc – University of Utah
Mustafa Ozmen, Dr – University of Utah
Mary Jo Pugh, PhD, RN – Salt Lake City VA and University of Utah
Rationale: In acquired epilepsies, spontaneous repeated seizures occur as a consequence of prior traumas including stroke, infection, disease, or traumatic brain injury (TBI). While early detection and treatment of epilepsy can improve health outcomes, it remains unclear who will and will not develop epilepsy following TBI. This study aimed to assess the feasibility of models for early prognostication of epilepsy within a US Veteran TBI cohort at 12, 18, and 24 months prior to first epilepsy diagnosis.
Methods: This observational study analyzed longitudinal data from Veterans Health Administration from 2003 – 2023 for N=146,545 post-9/11 Veterans with TBI of any severity, matched 1:1 with TBI negative controls. Diagnoses from health records, monthly pharmaceutical prescription data, and self-reported measures were used as longitudinal features to test a range of models for detecting epilepsy diagnosis after TBI. The models tested included logistic regression, random forest models, convolutional networks, and ensemble methods integrating multiple approaches. Models were tasked to correctly predict which Veterans do/do not go on to develop epilepsy within 12-24 months, with all patient data cutoff at least 12 months prior to first epilepsy diagnosis. Models were adjusted for sample imbalance. F scores and true/false positive rates were used to assess performance under 5-fold cross-validation. The explanatory power of each feature was assessed.
Results: We identified N= 146,545 post-9/11 Veterans with history of lifetime TBI of any severity. The majority of Veterans with TBI were male (91.9%), and the average age was 42.4 years. Overall, 5% of the TBI cohort developed a seizure disorder. The likelihood of future epilepsy diagnosis was estimated for each model. In logistic models, epilepsy risk increased with TBI severity, with odds ratios (OR) for mild (OR: 4.04 [3.70 - 4.42]), moderate/severe (OR: 8.18 [6.83 - 9.80]), and penetrating (OR: 15.7 (10.7 - 23.1) all showing significant associations. Other meaningful conditions included stroke (OR: 2.87 [1.67 - 4.94]) and brain tumor (OR: 6.03 [2.14 - 16.94]). In machine learning models, polypharmacy (p< 0.01), cumulative medication prescription (p< 0.01), and total comorbidity index (p< 0.01) were statistically significant epilepsy risk factors. We confirmed our hypothesis that Veterans with TBI who acquire epilepsy display distinct trajectories compared to those who do not acquire epilepsy. Accuracy and predictive power varied across models. Ensemble methods showed the best predictive performance overall, but required increased complexity.
Epidemiology