Prediction model of tacrolimus blood concentration based on population pharmacokinetics information and machine learning
- VernacularTitle:基于群体药动学信息与机器学习的他克莫司血药浓度预测模型研究
- Author:
Zhaolin CHEN
1
;
Pengli ZHU
1
;
Weijian NI
1
;
Lei ZHANG
1
;
Liqin TANG
2
Author Information
1. Dept. of Pharmacy，The First Affiliated Hospital of University of Science and Technology of China （Anhui Provincial Hospital），Hefei 230001，China;Anhui Provincial Key Laboratory of Precision Pharmaceutical Preparations and Clinical Pharmacy，Hefei 230001，China
2. Dept. of Pharmacy，The First Affiliated Hospital of University of Science and Technology of China （Anhui Provincial Hospital），Hefei 230001，China
- Publication Type:Journal Article
- Keywords:
renal transplantation;
tacrolimus;
population pharmacokinetics;
machine learning;
therapeutic drug monitoring
- From:
China Pharmacy
2026;37(14):1898-1904
- CountryChina
- Language:Chinese
-
Abstract:
OBJECTIVE To construct a prediction model for tacrolimus trough concentration based on population pharmacokinetics （PPK） information and machine learning， evaluate its predictive performance in clinical therapeutic drug monitoring （TDM） data， and provide evidence-based support for individualized dose adjustment of tacrolimus after renal transplantation. METHODS A retrospective cohort study design was adopted， and consecutive TDM records of kidney transplant recipients treated in the First Affiliated Hospital of University of Science and Technology of China （Anhui Provincial Hospital） from January to June 2025 were included to construct the final analysis set. Based on 29 input variables including previous TDM concentration， previous concentration/dose （ C / D ） ratio， and individual predicted exposure （IPRE）， 13 commonly used regression algorithm models and 1 weighted ensemble model were systematically compared. The dataset was split into training set and test set in an 80∶20 ratio by patient. Five-fold cross-validation based on patient grouping was performed on the training set， and the model performance was evaluated on the obtained evaluation points using the coefficient of determination （ R 2 ）， mean absolute error （MAE）， root mean square error （RMSE）， Pearson correlation coefficient （Corr） and ±20%/±30% accuracy. The SHapley Additive exPlanations （SHAP） method was used for model interpretability analysis to identify core predictive features. RESULTS A total of 809 consecutive TDM records from 83 kidney transplant recipients were included， with 66 cases in the training set and 17 cases in the test set. The SVR model showed the best performance， with R 2 of 0.649 8， MAE of 1.126 0 ng/mL， RMSE of 1.478 6 ng/mL， Corr of 0.819 1， ±20% accuracy of 59.68%， and ±30% accuracy of 79.03%. The Weighted Ensemble model and Extra Trees model ranked second and third， respectively. SHAP interpretability analysis showed that the top 6 features contributing to the prediction were IPRE， previous TDM concentration， previous C/D ratio， dose-normalized concentration， hematocrit （HCT）， and concentration-dose coefficient （ θ cd ）， among which IPRE was the core predictive feature （SHAP mean value 0.915 4）. CONCLUSIONS The machine learning model incorporating PPK information demonstrates good predictive ability for tacrolimus concentration in the current TDM cohort. SVR is the optimal model， IPRE is the core predictive feature of the model， and a few core PK-derived features （IPRE， dose-normalized concentration， θ cd ） together with clinical features （previous TDM concentration， HCT， etc.） constitute the main information sources for model prediction.
