Machine learning combined with population pharmacokinetics: a hybrid model for predicting the plasma concentration of linezolid in critically ill pediatric patients

Objective Linezolid is a crucial agent for treating drug-resistant Gram-positive bacterial infections in critically ill pediatric patients. However, its pharmacokinetics exhibit high inter-individual variability, making standard dosing regimens susceptible to underexposure or overexposure. This study aimed to explore the integration of population pharmacokinetics (PopPK) and machine learning (ML) algorithms to build a best-performing model for predicting individual linezolid plasma concentrations in critically ill children, thereby guiding personalized dosing. Methods Based on data from a retrospective cohort of 145 critically ill pediatric patients (213 samples), 32 features (including PK parameters) and the target variable (linezolid plasma concentration) were included. A PopPK model was established, and Monte Carlo simulations were conducted to optimize dosing regimens. After systematically evaluating 7 ML algorithms, the best predictive model was identified. Its interpretation was then performed with SHapley Additive exPlanations (SHAP). Results A one-compartment model with first-order elimination adequately described the pharmacokinetics of linezolid. Monte Carlo simulations for minimum inhibitory concentration (MIC) values between 0.5 and 2 mg/L, most renal function levels achieved a probability of target attainment (PTA) > 90% through dose adjustments. When MIC reached 4 mg/L, a PTA > 90% was achieved only in patients with severe renal impairment using 20 mg/kg q8h, yet with high safety risk, indicating the need for alternative antimicrobial agents. The LightGBM algorithm was best-performing among the seven algorithms evaluated, with testing set R 2 = 0.777 and external validation set R 2 = 0.686. However, these R 2 values were optimistic, biased estimates inflated by the EBE-derived CL and V, which have been estimated using all available concentrations per patient, rather than conventional predictive performance. SHAP analysis revealed that clearance, time after dose, daily dose, weight, and serum creatinine were the top five most important features. Conclusion This exploratory study integrated PK parameters into ML models for predicting linezolid plasma concentrations in critically ill pediatric patients, providing preliminary insights for future research as a proof of concept. However, the reported metrics could not support clinical deployment without leakage-free re-estimation using time-constrained or sequential validation.