Abstracts

Rationale: Managing epilepsy in pregnancy requires balancing seizure control with fetal safety, yet current models often neglect antiseizure medication (ASM) pharmacokinetic variability and hormonal fluctuations. To address this gap, we developed machine learning models that integrate trough ASM concentrations and sex hormone levels to predict seizure risk in pregnant women with epilepsy.

Methods: This analysis was based on a prospective, multicenter cohort study conducted in China. We included female patients with epilepsy who met the diagnostic criteria of the International League Against Epilepsy, received monotherapy or polytherapy with lamotrigine (LTG), levetiracetam (LEV), or oxcarbazepine (OXC), and completed follow-up and therapeutic drug monitoring (TDM) throughout the peripregnancy period. Pharmacokinetic metrics and longitudinal sex hormone levels were assessed across pregnancy, including concentration-to-dose ratios (CDRs), ratios to predefined therapeutic target concentrations (RTCs), serum estradiol (E2) and progesterone (P). Tobit regression was used to impute estradiol values that exceeded the upper detection limit. Following data preprocessing and application of the SMOTE-Tomek resampling technique to address class imbalance, machine learning models were trained using RTCs and hormone variables. The models included logistic regression, random forest, and extreme gradient boosting (XGBoost). SHapley Additive exPlanations (SHAP) analysis was performed to interpret model predictions and assess feature importance.

Results: A total of 97 pregnant Han Chinese women with epilepsy were included in the study. In total, 433 plasma samples were collected for intensive TDM. A progressive decline in ASM concentrations was observed throughout pregnancy, with mean reductions in CDRs by the third trimester of 70.1% for LTG, 28.2% for LEV, and 36.4% for OXC. Receiver operating characteristic analysis identified critical RTC thresholds associated with increased seizure risk: < 0.51 for LTG, < 0.55 for LEV, and < 0.71 for OXC. Two predictive models were developed: Model 1 used continuous RTC values and achieved an AUC of 0.77 using XGBoost, while Model 2 applied binary RTC cut-offs and achieved a higher AUC of 0.84 using logistic regression. SHAP analysis identified RTCs below 0.55 for LEV, RTCs below 0.51 for LTG, E2 levels as the most influential predictors of seizure risk. Importantly, a significant interaction was observed between LTG-RTC and progesterone levels, suggesting that their effects on seizure risk are not independent and may be governed by additional, possibly synergistic, mechanisms.