MACHINE LEARNING FOR PERSONALIZED DRUG DOSING IN CHILDREN: PHARMACOKINETIC MODELING AND CLINICAL VALIDATION

Abstract

Background: Pediatric pharmacokinetics is highly variable due to age-dependent physiological changes, making standard weight-based dosing inadequate for many children. Objective: To develop and validate machine learning models for individualized pharmacokinetic prediction and dose optimization in children. Methods: ML models were developed for five high-risk pediatric medications (vancomycin, tacrolimus, methotrexate, phenobarbital, aminoglycosides) using therapeutic drug monitoring data from 8,234 pediatric patients. Bayesian ML models integrating patient covariates with drug level measurements were compared against conventional population pharmacokinetic models. Results: ML-Bayesian hybrid models achieved 23.4% lower prediction error compared to conventional approaches (mean absolute prediction error 15.3% vs. 19.9%). Clinical implementation reduced vancomycin nephrotoxicity rates from 12.4% to 7.1%. Conclusion: ML-enhanced Bayesian pharmacokinetic modeling substantially improves individualized dosing precision in children, with demonstrated reduction in serious adverse drug events.