Machine learning-based method for tacrolimus dose predictions in Chinese kidney transplant perioperative patients
Qun Fu,Yan Jing,Guozhen Liu Mr,Xuehui Jiang Mr,Hong Liu,Ying Kong,Xiongjun Hou,Lei Cao,Pei Deng,Pin Xiao,Jiansheng Xiao,Hongwei Peng,Xiaohua Wei
DOI: https://doi.org/10.1111/jcpt.13579
Abstract:What is known and objectives: Tacrolimus (TAC), a first-line immunosuppressant in solid-organ transplant, has a narrow therapeutic window and large inter-individual variability, which affects its use in clinical practice. Successful predictions using machine learning algorithms have been reported in several fields. However, a comparison of 10 machine learning model-based TAC pharmacogenetic and pharmacokinetic dosing algorithms for kidney transplant perioperative patients of Chinese descent has not been reported. The objective of this study was to screen and establish an appropriate machine learning method to predict the individualized dosages of TAC for perioperative kidney transplant patients. Methods: The records of 2551 patients were collected from three transplant centres, 80% of which were randomly selected as a 'derivation cohort' to develop the dose prediction algorithm, while the remaining 20% constituted a 'validation cohort' to validate the final algorithm selected. Important features were screened according to our previously established population pharmacokinetic model of tacrolimus. The performances of the algorithms were evaluated and compared using R-squared and the mean percentage in the remaining 20% of patients. Results and discussion: This study identified several factors influencing TAC dosage, including CYP3A5 rs776746, CYP3A4 rs4646437, haematocrit, Wuzhi capsules, TAC daily dose, age, height, weight, post-operative time, nifedipine and the medication history of the patient. According to our results, among the 10 machine learning models, the extra trees regressor (ETR) algorithm showed the best performance in the training set (R-squared: 1, mean percentage within 20%: 100%) and test set (R-squared: 0.85, mean percentage within 20%: 92.77%) of the derivation cohort. The ETR model successfully predicted the ideal TAC dosage in 97.73% of patients, especially in the intermediate dosage range (>5 mg/day to <8 mg/day), whereby the ideal TAC dosage could be successfully predicted in 99% of the patients. What is new and conclusion: The results indicated that the ETR algorithm, which was chosen to establish the dose prediction model, performed better than the other nine machine learning models. This study is the first to establish ETR algorithms to predict TAC dosage. This study will further promote the individualized medication of TAC in kidney transplant patients in the future, which has great significance in ensuring the safety and effectiveness of drug use.