Predicting Plasma Concentration–Time Profiles of Complex Drugs with Double Peaks Using Machine Learning: A Case Study with Veralipride

Research utilizing artificial intelligence for population pharmacokinetic (PPK) analysis has advanced recently. We previously validated the utility of machine learning (ML) models to predict the plasma concentration (Cp)–time profiles of remifentanil. However, Cp prediction of complex drugs is essential for drug safety assessment in clinical settings. This study aimed to construct a predictive ML model using clinical trial data of veralipride, whose Cp–time profiles following oral administration are characterized by double peaks. We analyzed the veralipride Cp–time profiles of 12 participants at 189 time points. The ML algorithms random forest (RF) and gradient boosting (GB) were used. PPK analysis was also performed using established a two-compartment model with two absorption sites, incorporating either additive or additive plus multiplicative error models. In our ML model, we used the Cp data at 0.5, 3, and 24 h (final time point) post veralipride administration from one participant and the data at all time points from other participants. Leave-one-subject-out cross-validation (LOSO-CV) was conducted to evaluate the predictive models. We observed that the first and second peaks occurred after approximately 0.5 and 3 h of administration, respectively. Pre-evaluation using all the data for model development showed that the PPK(additive) model visually reproduced the double-peak profiles; however, on performing LOSO-CV, only the ML models predicted them. Compared with the actual values associated with double peaks, PPK showed discrepancies of 44.0% (additive model) and 58.4% (additive + multiplicative model), whereas ML exhibited discrepancies of 34.7% (RF) and 33.0% (GB), indicating better reproduction of the observed double-peak profiles by ML. The application of ML to drugs with complex Cp–time profiles in clinical settings can improve the Cp prediction accuracy.