Przewidywanie stężenia zonisamidu u dzieci z epilepsją: zastosowanie sztucznej inteligencji

OBJECTIVE: To construct and validate prediction models for zonisamide (ZNS) concentration in pediatric patients with epilepsy based on 12 machine learning algorithms, to screen for the optimal algorithm, and to provide a scientific basis for the formulation of individualized ZNS dosing regimens in children. METHODS: Clinical data of patients who underwent ZNS therapeutic drug monitoring at Kunming Children's Hospital from May 2022 to January 2026 were retrospectively collected and randomly divided into a training set and a test set at a ratio of 7:3. Key predictive variables were determined through a multi-stage feature screening strategy (covering correlation analysis, collinearity diagnosis, univariate analysis, Lasso regression, and random forest algorithm). Based on the selected variables, 12 machine learning regression models were constructed to predict ZNS concentration, and grid search combined with 5-fold cross-validation was employed for parameter optimization and performance evaluation. The coefficient of determination (R), mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE) were used as primary evaluation metrics. Finally, the SHAP method was adopted to interpret the feature contribution and decision logic of the optimal model. RESULTS: The modeling cohort of this study enrolled 532 pediatric patients who received zonisamide treatment at the Department of Neurology, Kunming Children's Hospital from May 2022 to May 2024. These patients were randomly divided into a training set (375 cases) and an internal validation set (157 cases) at a ratio of 7:3. Additionally, 436 pediatric patients from the same department of the same hospital from June 2024 to January 2026 were included as an external validation cohort. Comparisons of general clinical data, laboratory indicators, and medication-related data among the groups showed no statistically significant differences (all p > 0.05), indicating that the baseline data were balanced and comparable. The median (interquartile range) ZNS concentrations in the training set, internal validation set, and external validation cohort were 9.65 (8.11, 11.98) μg/mL, 9.93 (8.23, 12.06) μg/mL, and 10.77 (7.67, 13.69) μg/mL, respectively. Through multi-stage feature screening, gender, dosage, age, red blood cell count, concomitant medication status, total protein, uric acid, and platelets were identified as key factors influencing ZNS concentration. Among the 12 constructed machine learning models, the Random Forest (RF) algorithm demonstrated the optimal performance: in the training set, R was 0.97%, RMSE was 0.83%, MAE was 0.57%, and Err20 was 7.40%; in the internal validation set, R was 0.78%, RMSE was 1.99%, MAE was 1.48%, and Err20 was 31.20%; and in the external validation set, R was 0.89%, RMSE was 1.31%, MAE was 0.86%, and Err20 was 16.50%. SHAP method combined with representative decision tree analysis revealed that dosage and gender were the primary factors influencing ZNS concentration, followed by total protein and uric acid; among them, dosage showed a positive contribution, gender exhibited a bidirectional effect, and laboratory indicators mostly showed nonlinear associations. Decision tree analysis indicated that the model used gender as the primary splitting feature and incorporated multi-indicator interactions; its decision logic aligned with pharmacokinetic theory, providing strong support for the interpretability of the model's clinical application. CONCLUSION: This study successfully constructed and validated a ZNS concentration prediction model for pediatric patients with epilepsy based on the random forest algorithm. The model demonstrated high precision, strong stability, and good generalization ability. Gender, dose, uric acid, and total protein are core variables influencing ZNS concentration. The findings can provide a reference for the formulation of individualized ZNS dosing regimens in children.