Establishment of the predictive model for delayed metabolism of methotrexate in children with acute lymphoblastic leukemia: a retrospective study (Preprint)

BACKGROUND Methotrexate (MTX) is an effective chemotherapy for acute lymphoblastic leukemia (ALL) in children. However, prolonged exposure to high concentrations of MTX is associated with substantial adverse effects, including individual metabolic variability. Therefore, serum concentrations of MTX should be closely monitored in severe ALL patients to provide timely intervention. OBJECTIVE To achieve timely intervention for metabolic delay and personalized treatment by detecting serum methotrexate (MTX) concentration, we used multicenter MTX prescription data to predict the delayed MTX metabolism in children with Acute lymphoid leukemia(ALL) using machine learning (ML). METHODS We collected MTX chemotherapy data of 1729 cases. Univariate analysis and least absolute shrinkage and selection operator (LASSO) regression analysis were used to feature screening. Random forest classifier (RFC), eXtreme gradient boosting (XGBoost), adaptive boosting (AdaBoost), and light gradient boosting machine (LightGBM) ML algorithms were applied to construct prediction models. A series of indicators included the area under the receiver operating characteristic curve (AUROC) and the area under the PR curve (AUPR). Shapley Additive exPlanations (SHAP) were used to interpret the prediction models. RESULTS Eleven predictors were screened using univariate analysis and LASSO regression analysis, including age, weight, creatinine, uric acid total bilirubin albumin, white blood cell count, hemoglobin, prothrombin time, type of cell morphological classification, and omeprazole comedication. Among the test set, XGBoost had the AUROC of 0.898 (95% CI, 0.860–0.936) and AUPR of 0.752, RFC had the AUROC of 0.841 (95% CI, 0.789–0.893) and AUPR of 0.365, AdaBoost had the AUROC of 0.890 (95% CI, 0.848–0.932) and AUPR of 0.729, LightGBM had the AUROC of 0.900 (95% CI, 0.866–0.935) and AUPR of 0.722. CONCLUSIONS The XGBoost algorithm owned the best performance and the model showed accurate discriminatory ability in predicting the delayed metabolism of the MTX.