Comparing Machine Learning and Population Pharmacokinetic Approaches for Drug Concentration Prediction in Dense and Sparse Data Settings

Population pharmacokinetic (PopPK) models, the gold standard for drug concentration prediction, are relatively complex and time-consuming; machine learning (ML) excels with high-dimensional data but remains exploratory. This study systematically evaluated seven ML algorithms-Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), Gated Recurrent Unit with Decay (GRUD), Phased Long Short-Term Memory (PLSTM), Neural Controlled Differential Equations (NCDE), and Neural Ordinary Differential Equation (NODE)-and their heterogeneous XGBoost-based ensembles against PopPK across three data settings: dense phase I trial data; simulated dense data: linear absorption and elimination (LL), Michaelis-Menten absorption with linear elimination (MMA), linear absorption with Michaelis-Menten elimination (MME); and sparse therapeutic drug-monitoring (TDM) data. On two dense trial data, BiLSTM showed low systematic bias (F30 = 48.44%, MdPE = -6.86%; F30 = 50.62%, MdPE = 1.08%), whereas NCDE (F30 = 50.00%, MdPE = -19.39%) and NODE (F30 = 62.96%, MdPE = 11.77%) achieved the highest F30 values, respectively, compared with PopPK (F30 = 37.50%, MdPE = 24.73%; F30 = 61.73%, MdPE = 14.88%). BiLSTM showed robust extrapolation across LL, MMA, and MME datasets (F30 ≥ 75%, |MdPE| ≤ 10%). XGBoost lacked extrapolation ability, limiting ensemble gains. On sparse single-point prior data, XGBoost (F30 = 23.61%, MdPE = -8.42%) was most robust, while PopPK failed to converge. On multi-point sparse prior data, XGBoost (F30 = 26.84%, MdPE = -5.88%) outperformed PopPK (F30 = 23.63%, MdPE = -13.88%). In single-point sparse posterior prediction, PopPK (F30 = 29.71%, MdPE = 2.32%) performed best, while several ensembles improved over their neural network components. Overall, BiLSTM showed strong performance in dense data, XGBoost performed best in the prior sparse scenarios, and PopPK retained the advantage after prior concentration information was incorporated. ML approaches achieved similar accuracy to PopPK with unique advantages, representing promising tools for drug concentration prediction.