From time-series deep learning to physics-constrained population pharmacokinetics: AI-driven prediction of drug concentrations and exposure for model-informed dosing.

Therapeutic drug monitoring (TDM) starts from a single measured concentration, which population pharmacokinetics (PopPK) with Bayesian maximum a posteriori (MAPB) estimation converts into a revised dose. Since roughly 2021, two research lines have been converging. Time-series deep learning models read longitudinal electronic health record (EHR) data and predict concentrations directly, led by long short-term memory (LSTM) networks, neural ordinary differential equations (neural-ODEs), and attention-based architectures. Physics-informed neural networks (PINNs), scientific machine learning (SciML), and related mechanism-aware approaches embed pharmacokinetic knowledge into neural networks, chasing accuracy without giving up interpretability. We survey both strands from 2019 to 2026, covering foundational papers, representative applications, and the open-source ecosystem. We organize the mechanism-learning relationship into three levels: loose coupling feeds pharmacokinetic-derived features into a machine learning model; medium coupling builds mechanistic structure into the architecture; tight coupling writes the mechanism into the loss function as differentiable constraints. We compare the three across predictive accuracy, extrapolation, interpretability, data needs, computational cost, and clinical acceptability. A study-level evidence map and a critical appraisal of the primary statistics underpin a nuanced picture: mechanism constraints consistently improve extrapolation and identifiability and keep the physics consistent, but in-distribution accuracy gains over well-tuned PopPK + MAPB are modest, and several headline results rest on preprints or synthetic-data validation. We also locate each study on an evidence hierarchy separating concentration prediction from dose recommendation and clinical benefit. Open problems include data sparsity, identifiability, uncertainty quantification, multi-drug generalization, regulatory acceptance, and foundation models. We close with a roadmap toward mechanism-informed, clinically deployable precision-dosing systems. Throughout, pediatric antiepileptic drug (AED) TDM, above all lacosamide, lamotrigine, and oxcarbazepine, serves as the guiding clinical scenario; narrow therapeutic windows, developmentally heterogeneous sampling, and multi-drug comorbidity make it the clearest proving ground for the coupling framework, and we map the evidence gaps for these drugs.