Raltegravir Plasma Exposure: A Machine Learning-Based Model for its Prediction Using Limited Sampling Strategy

Abstract Recent studies have applied machine learning (ML)-based limited sampling strategies (LSS) to predict drug exposure (AUC), achieving low prediction error and performance comparable to or better than multiple linear regression and population pharmacokinetics LSS. This study aimed to develop and validate a machine learning-based limited sampling strategy capable of predicting raltegravir (RAL) exposure. Four machine learning algorithms (XGBoost, Random Forest, GLMNet, and SVM) were trained using pharmacokinetic profiles generated via Monte Carlo simulation from a population pharmacokinetic (POPPK) model. Data were divided into training (75%) and test (25%) sets. All possible combinations of sampling times, pairs and triplets, in steady-state, up to 12 h post-dose were evaluated. Model performance was assessed by the lowest root mean square error (RMSE) in the cross-validation, and the best performing model was evaluated in the test set and externally validated using simulated PK profiles from an independent POPPK model and patient data from a clinical study. XGBoost trained with concentrations at 0.5, 2, and 4 h showed the best predictive performance. The model achieved excellent accuracy in the test set (bias/RMSE: 0.8%/8.7%) and in the independent simulation (1.9%/14.3%). Performance decreased in real patient data (5.0%/24.1%), highlighting the need for caution when extrapolating predictions to populations whose characteristics differ from those represented in the training datasets. A machine learning model using only three sampling timepoints has been developed and validated in different datasets, enabling accurate estimation of RAL AUC₀-₁₂. This approach provides a tool for pharmacokinetic and PK/PD studies and reduces intensive sampling need in clinical settings. Graphical Abstract