Determinants and Machine Learning Prediction of Subtherapeutic Sodium Valproate Concentrations in Epilepsy Management...

Valproic acid is a classic antiepileptic drug; however, it is characterized by a narrow therapeutic window, limited safety margin, and marked individual variability. Therapeutic drug monitoring has become a key approach for individualized dosing of valproic acid. Nevertheless, in settings with limited medical resources, routine monitoring of valproic acid concentrations is often not feasible. This study aims to explore the factors influencing valproic acid concentrations and to identify machine learning algorithms with the most accurate classification performance, thereby supporting clinicians in the rational and individualized use of valproic acid in patients with epilepsy.

Patients receiving valproic acid were divided into a subtherapeutic group (< 50 mg/L) and therapeutic range group (50-100 mg/L). Least absolute shrinkage and selection operator (lasso) logistic regression was used to identify factors associated with subtherapeutic group. Multiple machine learning algorithms were also employed to construct binary classification models for predicting whether serum concentrations fall within the therapeutic range.

A total of 186 patients were ultimately included, comprising 110 patients in the therapeutic range group and 76 patients in the subtherapeutic group. Significant differences existed between the two groups in daily dosing frequency, total daily dose, route of administration, and alkaline phosphatase levels (P < 0.05). Lasso logistic regression analysis showed that daily dosage frequency (OR 0.163, P < 0.001) was the only clinically significant independent protective factor for achieving therapeutic blood concentration. The support vector machine (SVM) model achieved high values in the area under the curve (AUC), sensitivity, specificity, and accuracy for both the training and test sets, demonstrating strong generalization ability.

Increasing the daily dosing frequency of valproic acid is associated with a higher likelihood of achieving therapeutic serum concentrations. The results of this study suggest that a binary SVM model can be used to predict the risk of subtherapeutic levels.