An Integrated Population Pharmacokinetic and Machine Learning Model for Predicting Tacrolimus Exposure in Adult Patients with Nephrotic Syndrome

Yehui Zhou,1,2,* Zhou Zhou,2,* Shiting Chen,1,2,* Lili Zhu,3,* Yuting Yun,1,2 Yuan Yuan,4 Chen Chen,2 Jianjun Zou,2 Jun Zhao5,6 1School of Basic Medicine and Clinical Pharmacy, China Pharmaceutical University, Nanjing, Jiangsu, People’s Republic of China; 2Department of Pharmacy, Nanjing First Hospital, Nanjing Medical University, Nanjing, Jiangsu, People’s Republic of China; 3Pharmacy Laboratory Centre, School of Pharmacy, Xinjiang Second Medical College, Karamay, Xinjiang, People’s Republic of China; 4Department of Pharmacy, First Affiliated Hospital of Xinjiang Medical University, Urumqi, Xinjiang, People’s Republic of China; 5Xinjiang Key Laboratory of Clinical Drug Research, The First Affiliated Hospital of Xinjiang Medical University, Urumqi, Xinjiang, People’s Republic of China; 6Ministry of Education Engineering Research Center of Xinjiang and Central Asian Medicine Resources, Urumqi, Xinjiang, People’s Republic of China*These authors contributed equally to this workCorrespondence: Jun Zhao, Xinjiang Key Laboratory of Clinical Drug Research, The First Affiliated Hospital of Xinjiang Medical University, Urumqi, 830000, People’s Republic of China, Email [email protected] Jianjun Zou, Department of Pharmacy, Nanjing First Hospital, Nanjing Medical University, No. 68 Changle Road, Nanjing, 210006, People’s Republic of China, Email [email protected]: Tacrolimus (TAC) is a calcineurin inhibitor widely used for immunosuppressive therapy. It has a narrow therapeutic window and substantial interindividual variability, making therapeutic drug monitoring (TDM) essential. Traditional population pharmacokinetic (PPK) models have limited capacity to capture complex nonlinear relationships and multifactorial interactions. This study aimed to integrate PPK and machine learning (ML) to predict TAC plasma concentration.Patients and Methods: A total of 182 consecutive patients with nephrotic syndrome (NS) who received oral TAC therapy and underwent TDM and CYP3A5 genotyping at the First Affiliated Hospital of Xinjiang Medical University between January 2018 and December 2023 were enrolled. Individual PK parameters derived from the PPK model were incorporated as new features into ML models. Feature selection was performed using recursive feature elimination with cross-validation (RFECV), and the union of features selected by the three models with the highest R2 was used to construct the final feature set. Nine ML models were then developed, and the three best-performing models were combined into a weighted ensemble. Model performance was assessed by R2, mean absolute error (MAE), and root mean square error (RMSE). SHAP (SHapley Additive exPlanations) was used to interpret feature contributions.Results: Incorporating PK parameters resulted in a consistent improvement in predictive performance compared with the model without PK parameters (R2: 0.633 vs. 0.602). The weighted ensemble of CatBoost, AdaBoost, and GraBoost (5:3:2) achieved the best performance (R2 = 0.633, MAE = 1.081, RMSE = 1.377), whereas the optimal weighting without PK parameters was 3:6:1 (R2 = 0.602, MAE = 1.139, RMSE = 1.432). SHAP analysis indicated that PK parameters was the most influential feature, exceeding conventional biochemical indicators.Conclusion: Incorporating PK parameters enhanced model predictive performance, providing quantitative support for individualized TAC dosing and facilitating safer and more precise clinical applications.Keywords: tacrolimus, population pharmacokinetics, machine learning, RFECV, plasma concentration prediction