Development of a web-based machine learning model for early prediction of delayed high-dose methotrexate clearance in pediatric osteosarcoma

Objective Pediatric osteosarcoma treatment with high-dose (HD) methotrexate (MTX) poses risks of delayed clearance due to immature organ function. Interpretable machine learning facilitates proactive prediction, enabling improved monitoring and reduced toxicity risks. Method A retrospective study (2020–2024) of 181 pediatric patients with osteosarcoma treated with HD-MTX was conducted to identify key predictors using least absolute shrinkage and selection operator (LASSO) regression and Boruta analyses. Training and test datasets were proportionally split. Ten-fold cross-validation and hyperparameter tuning were performed, followed by the development of models (LASSO, ridge, and logistic regression), which were subsequently compared. Model performance was comprehensively evaluated using metrics such as the receiver operating characteristic (ROC) and F 1 scores. Results Of the 181 patients, 51 experienced delayed MTX elimination. Nine predictors (MTX3H, IBil3H, Urea3H, Cr, APTT, PT, MPV, EC, and FIB) were identified. LASSO regression outperformed the other models, achieving an AUC of 0.8466, with robust performance across multiple metrics, including the ROC, F 1 score, and decision curve analysis. Conclusion This interpretable machine learning model effectively predicts delayed MTX elimination in pediatric patients with osteosarcoma, enhancing patient monitoring, minimizing toxicity risks, and supporting evidence-based clinical decisions. The application is publicly accessible at: https://sclslc.shinyapps.io/shiny_cls2_1model_dalex/