Mechanism-informed machine learning for individualized tacrolimus dose adjustment in the early post-kidney transplant...

Tacrolimus dosing in the early post-kidney transplant period is challenging due to a narrow therapeutic index and substantial interindividual pharmacokinetic (PK) variability. This study aimed to develop and validate mechanism-informed machine learning (ML) models to support individualized tacrolimus dosing during this critical period.

A total of 4311 tacrolimus trough concentrations (C_trough) within 7 days post-transplant were obtained from 1624 kidney transplant recipients. Two ML models, Gated Recurrent Unit (GRU) and eXtreme Gradient Boosting (XGBoost), were developed to predict C_trough and recommend doses to achieve target levels. Both models incorporated PK principles based on linear pharmacokinetics. Model performance was compared to a traditional Bayesian population PK (PopPK) model and purely data-driven ML models via internal cross-validation and external validation.

The mechanism-informed GRU model outperformed the Bayesian PopPK model in both internal validation (MSE = 7.81 vs. 9.27 ng²/mL², R² = 0.537 vs. 0.450) and external validation (MSE = 6.09 vs. 8.96 ng²/mL², R² = 0.397 vs. 0.211). The mechanism-informed XGBoost model performed comparably to the GRU model. The incorporation of PK principles enhanced model interpretability and generalizability without reducing accuracy. When clinically administered doses, determined by conventional therapeutic drug monitoring, fell within the GRU model's recommended range, subsequent C_trough reached the therapeutic target (8-12 ng/mL) in 51.3% of cases, compared to 37.0% overall (p < 0.01).

Mechanism-informed ML models offer a robust and interpretable approach for individualized tacrolimus dosing, with the potential to improve therapeutic target attainment by enabling accurate dose adjustments in the early post-transplant period.