BACKGROUND: Integrating machine learning (ML) with population pharmacokinetic (PPK) modeling may improve therapeutic drug monitoring predictions. METHODS: Tacrolimus trough concentrations from lung transplant patients were split into a training data set (1151 concentrations in 80 patients) and a testing data set (224 concentrations in 20 patients). A PPK model was developed using NONMEM, followed by the development of 10 ML models to fit individual pharmacokinetic parameters from the PPK model with Bayesian forecasting. The best performing ML model was selected as the final model. Both the final PPK and ML models were compared for prediction performance. A web-based dashboard was established with R-shiny to recommend dosing regimens based on patient data. RESULTS: In the PPK model, postoperative days, hematocrit, aspartate aminotransferase, tacrolimus daily dose, coadministered voriconazole or posaconazole, and the CYP3A5*3 genotype were identified significant covariates on the clearance. The Cubist ML model outperformed the PPK model, showing lower root mean squared error for tacrolimus concentrations in the testing data set. An online web-based precision dosing dashboard was created, accessible at (https://tac-dose-ml.shinyapps.io/shiny/). CONCLUSIONS: Integrating ML with PPK modeling could yield superior tacrolimus concentration predictions for lung transplant patients, offering an efficient alternative to Bayesian forecasting. The online dashboard provides rapid, individualized dosing recommendations.
Cai et al. (Fri,) studied this question.