Compare the performance of multiple machine learning models in predicting tacrolimus concentration for infant...

This study aims to establish multiple ML models and compare their performance in predicting tacrolimus concentration for infant patients who received LDLT within 3 months after transplantation.

Retrospectively collected basic information and relevant biochemical indicators of included infant patients. CMIA was used to determine tacrolimus C₀ . PCR was used to determine the donors' and recipients' CYP3A5 genotypes. Multivariate stepwise regression analysis and stepwise elimination covariates were used for covariates selection. Thirteen machine learning algorithms were applied for the development of prediction models. APE, the ratio of the APE ≤3 ng ml^-1 and ideal rate (the proportion of the predicted value with a relative error of 30% or less) were used to evaluate the predictive performance of the model.

A total of 163 infant patients were included in this study. In the case of the optimal combination of covariates, the Ridge model had the lowest APE, 2.01 (0.85, 3.35 ng ml^-1 ). The highest ratio of the APE ≤3 ng ml^-1 was the LAR model (71.77%). And the Ridge model showed the highest ideal rate (55.05%). For the Ridge model, GRWR was the most important predictor.

Compared with other ML models, the Ridge model had good predictive performance and potential clinical application.