Objective  To explore the predictive value of two models based on machine learning algorithms in predicting the initial and subsequent doses of tacrolimus in kidney transplant recipients.

Methods  A retrospective analysis was conducted on the medical records of 1 013 Chinese kidney transplant recipients at the First Affiliated Hospital of Sun Yat-sen University from January 2015 to April 2019, focusing on the initial and subsequent doses in kidney transplant recipients. Thirty-three variables were collected for the initial dose, and twenty-six variables for the subsequent dose. A genetic algorithm combined with a random-restart hill-climbing algorithm was used to determine a small number of key clinical variables through majority voting, and variables with Lasso regression coefficients less than the optimal variable coefficient threshold were further eliminated. The selected clinical variables were input into a cascaded deep forest (CDF) and TabNet deep neural network for analysis and comparison based on structured tabular data, and the leave-one-subject-out method was used for validation.

Results  A total of 613 recipients were included in the training set, and 116 recipients were in the external validation set. In the initial dose algorithm of tacrolimus, the clinical variables ultimately included target concentration, time from surgery to target concentration, body weight, gender, type of surgery, time from surgery to first dose, WuZhi capsule, calcium channel blocker, creatinine, hemoglobin and CYP3A5. In the subsequent dose algorithm, the clinical variables ultimately included target concentration, time from surgery to target concentration, WuZhi capsule, creatinine, alanine aminotransferase, aspartate aminotransferase, previous dose, previous dose concentration and time from surgery to previous concentration. Based on the above variables, the TabNet model showed better predictive performance than the CDF model: in the initial dose prediction, the accuracy of the predicted dose within ±20% of the actual dose was 0.801, and the fitting index R² was 0.436; in the subsequent dose prediction, the corresponding accuracy and R² were 0.939 and 0.902, respectively. The results of feature contribution showed that CYP3A5 and target concentration contributed the most to the prediction of initial dose, while previous dose and its corresponding concentration had the greatest impact on subsequent dose prediction. In addition, the results of independent external validation were also satisfactory.

Conclusions  The optimized TabNet predictive model may provide important reference for drug dose prediction based on machine learning algorithms in clinical practice.