Development and External Validation of a Machine Learning Model to Predict Clearance in Patients with Epilepsy and Migraine Receiving Intravenous Topiramate with or without Enzyme-Inducing Comedications

Abstract

Objectives: Topiramate (TPM) is indicated for seizures and migraine prophylaxis and is used off-label for a growing list of neuropsychiatric conditions. Currently, only oral TPM is available, and patients who cannot take TPM by mouth face potential treatment interruptions. We conducted a phase I study of intravenous (IV) TPM safety and pharmacokinetics in patients with seizures and migraines [ 1]. We used the data from the PK study to externally validate a machine learning model trained to predict patient clearance. The study aims were to evaluate the performance of a machine learning model trained using simulated data from a population oral TPM pharmacokinetic model to predict IV TPM clearance in a real-world population under different administration scenarios. Methods: Patient characteristics (i.e., sex, weight, age, comedications, serum creatinine (SCr), and creatinine clearance (CrCL)) and TPM concentration-time data were simulated from a published PopPK model of children and adults receiving oral TPM (n = 2000) [ 2]. The median (range) administered dose was 100 mg (25- 1000 mg). After assessing simulation accuracy using reference covariates, we used the simulated data (training dataset) to train five ML algorithms, including L1- regularization (LASSO), K-Nearest Neighbors (KNN), Gradient-Boosted Machine (GBM), Random Forest (RF), and eXtreme Gradient Boosting Machine (XGBoost) to predict TPM clearance (CL in L/Hr). To prevent overfitting, a 10-fold cross-validation was performed during model training. A test dataset with concentration-time data from a phase I pharmacokinetic study of IV TPM (n = 20) and corresponding demographic information was created. Individual clearances were estimated using noncompartmental analysis (NCA). The predictors for IV?TPM CL included in the ML model were age, sex, weight, comedications, CrCL, and dose. Relative Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R-squared values (R2) were computed for each ML algorithm. We also calculated the percentage of predictions within ± 20 % of the actual value (20% accuracy) for each algorithm. Results: XGBoost achieved the highest predictive performance (MAE = 0.49, RMSE = 0.73; R2 = 0.57, 20% accuracy = 70%), followed by RF (MAE = 0.6, RMSE = 0.82; R2 = 0.6, 20% accuracy = 60%). Feature importance analysis revealed comedications, CrCL, and weight as the most important predictors for IV TPM CL. Conclusions: We show the utility of ML methods for reliably predicting drug clearance using a priori information from published PopPK models. These results highlight the potential of machine learning models to enable faster prediction of clinical pharmacokinetic parameters, such as clearance, and support model-informed dosing strategies.