Therapeutic drug monitoring of zonisamide in children with epilepsy: development and validation of a precise binary prediction model.

OBJECTIVE: This study aimed to investigate the key factors influencing the failure to achieve steady-state trough concentrations of zonisamide in pediatric patients with epilepsy. Furthermore, we sought to construct and validate a binary machine learning prediction model for zonisamide concentration attainment to provide an objective reference for therapeutic drug monitoring in clinical practice. METHODS: A retrospective analysis was conducted on 958 pediatric patients with epilepsy who received zonisamide treatment at the Department of Neurology, Kunming Children's Hospital, from May 2022 to January 2026. Multi-dimensional feature screening was performed on the training set using spearman correlation analysis, variance inflation factor multicollinearity diagnosis, univariate analysis, lasso regression, boruta algorithm, and 500-iteration bootstrap resampling. Subsequently, 17 machine learning classification models were developed. Hyperparameter optimization was achieved through grid search combined with 10-fold cross-validation. Model performance was comprehensively evaluated in terms of discrimination, calibration, and clinical net benefit using 500-iteration bootstrap resampling. Finally, SHAPley additive exPlanations analysis and a visualized decision tree were employed to interpret the feature contribution and intrinsic decision logic of the optimal model. RESULTS: Following multi-stage screening, six core predictive variables were identified: dosage, uric acid, age, total bilirubin, vitamin D, aspartate aminotransferase, and gender. A comparative analysis of the 17 machine learning models revealed that the classification and regression tree model exhibited the best comprehensive predictive performance, demonstrating superior stability and robustness under resampling validation. The area under the curve values for the internal test set and the time-split internal validation set were 0.82 and 0.87, respectively. Calibration curves and decision curve analysis confirmed ideal calibration and a wide range of clinical net benefit. SHAPley additive exPlanations analysis and decision tree results confirmed that dosage is the primary determinant of the probability of achieving target ZNS concentrations. Sex, liver and kidney function, and nutrition-related indices collectively modulate the processes of drug absorption, metabolism, and clearance, thereby influencing the risk of failing to achieve target concentrations. CONCLUSION: The machine learning prediction model constructed using the CART algorithm can accurately and stably predict whether steady-state trough concentrations of zonisamide reach the target therapeutic range in pediatric patients with epilepsy.