Porównanie skuteczności predykcyjnej modeli uczenia maszynowego i farmakokinetycznych modeli populacyjnych w przewidywaniu klirensу kwasu walproinowego i jego stężeń: badanie symulacyjne w scenariuszach epilepsji
Comparative Predictive Performance of Machine Learning and Population Pharmacokinetic Models for Valproic Acid Clearance and Trough Concentrations: An In-silico Simulation Study in Epilepsy Scenarios
W skrócie
Badanie porównało dwie metody przewidywania dawek leku przeciwpadaczkowego - kwasu walproinowego: tradycyjne modele matematyczne oraz nowoczesne modele sztucznej inteligencji. Okazało się, że modele sztucznej inteligencji są lepsze, gdy mamy mało informacji o pacjencie, natomiast tradycyjne modele są lepsze, gdy mamy już dane o pacjencie i jego odpowiedzi na lek. Naukowcy zalecają dalsze badania w warunkach rzeczywistych, aby potwierdzić te wyniki.
Oryginalny abstract (angielski)
PURPOSE: While population pharmacokinetic (PopPK) models traditionally guide valproic acid (VPA) dosing, machine learning (ML) may better capture complex, nonlinear relationships. A direct comparison of their predictive performances remains poorly defined. This study compared the predictive performance of ML and PopPK models for VPA clearance and trough concentrations. METHODS: PopPK and ML models were developed and validated using two independent simulated datasets. Trough concentration and clearance were each evaluated under a priori and a posteriori conditions and were compared against matched PopPK references. RESULTS: For trough concentration, most ML models significantly outperformed PopPK population prediction (PRED) under a priori condition, while individual prediction (IPRED) significantly outperformed all ML models under a posteriori condition. For clearance, the best-performing ML model (kNN) significantly outperformed a population-typical prediction (CL PRED) under a priori condition, whereas the best-performing ML model (CatBoost) showed a small but statistically significant advantage over the empirical Bayes estimate for clearance (CL EBE) under a posteriori condition. CONCLUSIONS: Relative performance depended on the prediction target and available data. Under a priori condition, ML outperformed PopPK for both endpoints. Under a posteriori condition, PopPK remained superior for trough concentration, while ML performed comparably for clearance. Further real-world validation is needed.