Opracowanie modelu predykcyjnego dla stężenia lewetyracetamu we krwi u dzieci z epilepsją na podstawie charakterystyki farmakokinetycznej: porównanie regresji liniowej i zaawansowanych algorytmów

OBJECTIVES: Levetiracetam (LEV) is mainly eliminated unchanged through the kidneys, and its steady-state serum concentration is associated with dose, body weight, renal function, and concomitant medications. However, marked concentration variability remains in children with epilepsy due to growth and developmental changes and interindividual pharmacokinetic differences. This study aims to compare the performance of multiple linear regression and complex machine learning algorithms in predicting steady-state trough concentrations of LEV in children with epilepsy, and evaluate the influence of increasing model complexity on prediction accuracy and stability. METHODS: Data were retrospectively collected from children with epilepsy who received LEV therapy and underwent therapeutic drug monitoring (TDM) at Fuzhou Children's Hospital of Fujian (Children's Specialized Campus of Fuzhou First General Hospital) between June 1, 2023 and October 30, 2024. A total of 696 children aged 1-18 years who had achieved steady-state LEV plasma concentrations and had complete clinical data were included. The patients were randomly divided into a training set of 556 patients and a test set of 140 patients at a ratio of 8꞉2. Candidate covariates included age, sex, body weight, total daily dose, dosage form, concomitant antiseizure medications, and laboratory indicators of hepatic and renal function and hematological status. Collinearity was assessed using variance inflation factors, and Pearson correlation analysis was performed to evaluate the relationships between the covariates and steady-state trough concentrations of LEV. Feature selection was conducted using Lasso regression and random forest, after which ordinary linear regression, ridge regression, Lasso regression, Bayesian linear regression, support vector regression, random forest regression, and XGBoost regression models were constructed. All models were trained, subjected to 10-fold cross-validation, and optimized through hyperparameter tuning in the training set. Their predictive performance was evaluated in the test set using the coefficient of determination (), mean absolute error (MAE), and root mean square error (RMSE). RESULTS: Correlation analysis showed that the daily weight-normalized dose was moderately positively correlated with LEV concentration (=0.399, <0.001) and was the most important linearly correlated variable, followed by the total daily dose (=0.287, <0.001). Clobazam dose, valproic acid concentration, carbamazepine dose, sex, and albumin were also correlated with LEV concentration to varying degrees. Both Lasso regression and random forest feature-importance analyses identified the daily weight-normalized dose as the predictor with the greatest contribution. Renal function-related indicators, albumin, serum potassium, total daily dose, and certain concomitant antiseizure medications also contributed to the variability in LEV concentrations. Based on the features selected by Lasso regression, ordinary linear regression demonstrated stable performance in the independent test set, with an of 0.473, an RMSE of 3.618 mg/L, and an MAE of 2.793 mg/L. Lasso regression, ridge regression, and Bayesian linear regression showed similar performance, with values of 0.472, 0.465, and 0.464, respectively. Among the nonlinear models, support vector regression achieved an of 0.458, which was close to the performance of the linear models. Random forest regression and XGBoost regression yielded values of 0.402 and 0.378, respectively, and did not demonstrate superior predictive performance. Age-stratified analysis showed that linear regression performed consistently across age groups overall. However, the predictive performance of all models declined in children aged 3 to 5 years, suggesting that variability in LEV concentrations in this age group may not be adequately explained by the currently available routine covariates. Goodness-of-fit and error analyses showed that predicted values of all models were generally distributed along the = reference line, and the median relative prediction error (PE%) was close to 0. Larger relative errors were mainly observed in the low-concentration range of <10 mg/L. Within the clinically relevant ranges of 10 to <20 mg/L and 20 to <30 mg/L, most samples had an absolute PE% of ≤30%. A slight regression-to-the-mean effect was observed only in individuals with extremely high or low concentrations, manifesting as slight overestimation at low concentrations and slight underestimation at high concentrations. The residuals of the linear regression model were generally distributed around 0, with no evident heteroscedasticity or highly influential observations dominating the model results. CONCLUSIONS: After incorporating dose, body weight, renal function, concomitant medications, and routine laboratory indicators, the principal predictable information regarding steady-state trough concentrations of LEV in children with epilepsy could be adequately captured by a multiple linear regression model. Its overall predictive performance, error stability, and clinical interpretability were not inferior to those of complex machine-learning models. The daily weight-normalized dose was the key determinant of steady-state trough LEV concentration, while renal function, albumin, serum potassium, and certain concomitant antiseizure medications further explained individual differences in LEV concentrations. Nevertheless, unexplained interindividual variability remained, potentially owing to factors such as medication adherence and developmental maturation. The model results may provide supportive information for the interpretation of TDM findings and dose adjustment, but their clinical application should be based on a comprehensive assessment of the child's clinical manifestations and therapeutic response.