Machine learning-assisted tacrolimus dose optimization in childhood- onset systemic lupus erythematosus through population pharmacokinetic modeling

Abstract

Objective

This study aimed to improve treatment effectiveness in childhood-onset systemic lupus erythematosus (cSLE) by developing machine learning algorithms integrated with pharmacokinetic parameters to predict individualized tacrolimus dosing for optimized therapy.

Methods

Data from 480 trough tacrolimus concentrations in 86 cSLE patients over five years were analyzed. A nonlinear mixed-effects model was constructed to characterize the pharmacokinetics of tacrolimus. We screened 27 machine learning and deep learning models using 29 clinical variables to select the best-performing individualized dose prediction model. To further enhance prediction accuracy, the pharmacokinetic parameters were subsequently embedded within the optimized model. Model performance was assessed through the utilization of goodness-of-fit plots and diagnostic parameters such as objective function values and Shapley Additive exPlanations (SHAP) values.

Results

The pharmacokinetic profile of tacrolimus was best accurately characterized by a one-compartment model, which incorporated first-order kinetics for both absorption and elimination processes. The typical estimates for apparent clearance (CL/F) and volume of distribution (V/F) were 3.52 L/h/70 kg and 124.84 L/70 kg, respectively. Among the 27 machine learning models, the XGBoost algorithm demonstrated the best prediction accuracy for tacrolimus dose (R² = 0.74, mean absolute error [MAE] = 0.016, mean square error [MSE] = 0.0005). After incorporating 11 key variables, including pharmacokinetic parameters (CL/F and V/F), into the model and performing hyperparameter turning, prediction accuracy significantly enhanced (R² = 0.80, MAE = 0.013, MSE = 0.0004). Two cSLE patients who received model-predicted doses achieved disease activity indexes ≤4 after treatment.

Conclusion

This study successfully developed an accurate machine learning-based model for predicting individualized tacrolimus dosing in cSLE patients. The integration of pharmacokinetic parameters significantly enhanced the model's accuracy, improving dosing precision and reducing overexposure, thus enhancing therapeutic efficacy.