Sergio Sánchez-Herrero*1,2, Marina Cuquerella-Gilabert1, Jenifer Serna1, Almudena Rueda-Ferreiro1, Hinojal Zazo3,4, José M. Lanao3,4
(1) Simulation Department, Empresarios Agrupados Internacional S.A., Madrid, Spain; (2) Computer Science, Multimedia and Telecommunication Department, Universitat Oberta de Catalunya. Barcelona (Spain) (3) Pharmaceutical Sciences Department, University of Salamanca, Salamanca, Spain: (4) Institute of Biomedical Research of Salamanca (IBSAL), Salamanca, Spain
Objectives:
The efficacy of drug therapy in neonates critically depends on selecting appropriate dosage regimens. Physiologically based pharmacokinetic (PBPK) modelling emerges as a valuable tool for assessing the clinical efficacy and safety of current adult treatments in paediatric populations [1]. However, challenges such as significant pharmacokinetic variability, a narrow therapeutic index, and limited sample sizes underscore the need for precise dosing strategies to ensure adequate drug exposure [2]. In addressing these complexities, machine learning (ML) methods offer promising support for clinical research due to their data-driven nature, robustness, and adaptability in managing intricate systems. Consequently, the integration of hybrid approaches emerges as a potential solution to enhance the accuracy of model-informed precision dosing (MIPD) strategies. The objective of this study was to assess the potential of various hybrid (PBPK-ML) continuous learning approaches in enhancing PBPK modelling, including semi-mechanistic or physiological models, and subsequent simulation outcomes. Our primary focus was on evaluating the effectiveness of Artificial Neural Networks (ANNs) and Gradient Boosting Algorithms for Supervised Learning (XGBOOST) algorithms in predicting gentamicin plasma concentrations over time.
Methods: A Gentamicin PBPK model [1] was employed within PhysPK through Monte Carlo simulations. These simulations produced a spectrum of potential outcomes, accounting for various variables or scenarios based on a priori likelihoods. The random numbers generated followed a log-normal probability density function. Neonatal population was divided into three sub-populations: Term, Pre-term 1 (38-33 weeks) and Pre-term 2 (32-24 weeks). Virtual populations were utilized as a complement to a real dataset (99 neonates with gestational ages between 24 and 39 weeks) for training algorithms used (ANNs and XGBOOST), in Python. Performance metrics such as Average Fold Error (AFE), Absolute Average Fold Error (AAFE), and Percent Prediction Error (PPE) were employed for data validation. The virtual gentamicin plasma concentrations generated were then compared with a retrospective analysis of 49 neonates with gestational ages between 24 and 39 weeks and younger than one-week postnatal age.
Results:
External validations corroborated the model’s alignment with theoretical assumptions and its accurate prediction in 2 and 24 hours. Performance metrics model’s efficacy for PBPK model for each sub-population respectively were AFE=0.96/0.95/0.88, AAFE=1.25/1.22/1.32 and PPE=0.4/-0.51/-5.8 for 2 hours and AFE=0.86/0.91/0.74, AAFE=1.52/1.33/1.72 and PPE=-0.99/-2.72/0.44 for 24 hours. Performance metrics with the algorithm ANNs for PBPK-ML model were the best, though without significant accuracy improvement in comparison with the PBPK model (AFE=1.18/1.21/0.9, AAFE=1.23/1.23/1.3 and PPE=20.58/27.72/-5.8 for 2 hours and AFE=0.99/0.97/1.15, AAFE=1.2/1.23/1.4 and PPE=1.76 /0.04/51 for 24 hours and for each sub-population respectively).
Conclusions:
Despite not getting significant improvements, hybrid PBPK and ML models could be a way of research to provide better recommendations for an optimal dosing regimen in neonates. This approach could represent an innovative way to predict dosage regimens. Other algorithms, methods or ML approaches will be analysed in the future.
References:
[1] Zazo Gómez, Hinojal, et al. (2022). Physiologically-based pharmacokinetic modelling and dosing evaluation of gentamicin in neonates using PhysPK. Frontiers in Pharmacology, 13 (977372).
[2] Wicha, Sebastian G., et al. “From therapeutic drug monitoring to model‐informed precision dosing for antibiotics.” Clinical Pharmacology & Therapeutics 109.4 (2021): 928-941.
Reference: PAGE 32 (2024) Abstr 11213 [www.page-meeting.org/?abstract=11213]
Poster: Methodology - New Modelling Approaches