Methotrexate (MTX) administered intravenously at doses of ≥500 mg/m² or >20 mg/kg is defined as high-dose methotrexate (HDMTX) (2019). HDMTX significantly increases the serum concentration of MTX, allowing it to penetrate the blood-brain barrier and reach poorly vascularized solid tumors, which makes it widely used in first-line chemotherapy for diseases such as Acute Lymphocytic Leukemia (ALL) and Non-Hodgkin Lymphoma (NHL) (Prodduturi and Bierman, 2012). HDMTX therapy can improve the overall prognosis of patients with hematological malignancies; however, the MTX doses used are potentially lethal. While these doses inhibit tumor growth, they can also damage proliferating normal tissues, leading to toxic events such as acute liver and kidney injury, bone marrow suppression, mucositis, gastrointestinal issues, fever, and infections (2019). Studies have shown that delayed MTX excretion further increases the risk of these toxicity events (Yu et al., 2024).

Currently, most related studies suggest that age, body surface area, and drug dosage are the primary influencing factors for MTX excretion delay (Chen et al., 2020; Zhang et al., 2022; Yu et al., 2024). Additionally, creatinine clearance, drug interactions, and genetic polymorphisms also impact the in vivo clearance process of MTX (Faganel Kotnik et al., 2011; Kawase et al., 2016; Tran and Herrington, 2016; Umerez et al., 2017; Hamada et al., 2018; Hui et al., 2019; Ishizaki et al., 2020; Ueda et al., 2020; Yang et al., 2020; Ferracini et al., 2021; Ebid et al., 2022; Mao et al., 2022). However, there are several discrepancies among different studies, and many of these studies rely on traditional statistical methods to screen risk factors and build predictive models. These methods often suffer from poor sensitivity and specificity, resulting in fluctuating prediction outcomes and significant reliance on the researcher’s data analysis expertise.

This study was based on real-world data and employed machine learning methods to construct a more stable and predictive risk model. The primary aims were to explore the predictive factors for delayed MTX excretion in pediatric patients, identify children at risk of MTX excretion delay as early as possible, and to provide a basis for precision treatment, thereby reducing the incidence of toxicity events.

HDMTX therapy is associated with significant antitumor efficacy; however, it also carries substantial toxicity risks (Mosleh et al., 2023). Given that children have less mature organ development and lower tolerance to drug toxicity compared to adults, they are at a higher risk of severe adverse events, including mortality. Delayed MTX excretion is typically defined based on whether plasma MTX concentrations exceed predefined threshold levels at specific time points after infusion. However, elevated MTX levels are usually detected only after delayed excretion has already occurred, limiting opportunities for early intervention. This study was a retrospective analysis of pediatric patients with ALL or NHL who received HDMTX therapy. Leveraging real-world data and machine learning techniques, we identified key predictors of delayed MTX excretion from demographic, dosing, treatment cycle, and laboratory variables. A risk prediction model was subsequently developed to identify high-risk patients before HDMTX administration. By enabling early risk stratification, this model supports the delivery of individualized, precision-based interventions to potentially reduce the incidence of delayed excretion and associated toxicities.

Most existing studies on delayed MTX excretion have traditionally relied on conventional statistical methods for feature selection and predictive model development. These traditional approaches, such as linear regression, require the fulfillment of numerous assumptions—linearity, independence, normality, homoscedasticity, among others—before they can be applied effectively. When dealing with smaller datasets, simple linear models are often the most reasonable choice due to their interpretability and straightforward implementation. However, these models’ expressiveness becomes notably limited as dataset sizes increase. In contrast, machine learning models can overcome many of the limitations imposed by traditional statistical models and incorporate techniques like regularization to mitigate overfitting (Christodoulou et al., 2019; Chen et al., 2023). Consequently, in this study, we opted to utilize machine learning methods for model construction. Machine learning models, especially more complex ones, typically exhibit superior predictive performance but are often criticized as “black boxes” due to their opaque decision-making processes. To address this interpretability issue, we employed SHAP, which leverages concepts from cooperative game theory to explain individual predictions made by machine learning models. By calculating each feature’s contribution to the prediction outcome, SHAP values provide a means to assign optimal credit allocation and local explanations. This approach allows us to better understand how the XGBoost model arrives at its decisions, thereby enhancing transparency and trustworthiness in clinical applications.

In this study, the feature selection analyses using LASSO, RF, and SVM-RFE consistently identified eight key predictors of delayed MTX excretion: Scr, total MTX dose, ALP, CK, Urea, GGT, HB, and height. These eight variables were then incorporated into four machine learning models—LR, NB, SVM, and XGBoost—for predictive modeling. Based on internal validation, the XGBoost model demonstrated superior performance compared to the other models. Therefore, we focused on interpreting the XGBoost model using SHAP to explore the contribution of each predictor in individual risk assessment.

The SHAP analysis conducted on the XGBoost model revealed that Scr had the highest contribution to the prediction of delayed MTX excretion. The dependence plot illustrated higher Scr values were consistently associated with increased model-predicted risk of delayed MTX excretion, which was consistent with the statistical associations observed in previous studies (Xu et al., 2014; Mei et al., 2018; Zhan et al., 2022). From a pathophysiological perspective, MTX and its metabolites are primarily eliminated through glomerular filtration and proximal tubular secretion into the primary urine, followed by excretion via urine (Mosleh et al., 2023). When renal function, particularly glomerular filtration rate (GFR), is compromised, the clearance of MTX decreases. Scr is a common indicator of glomerular filtration. Notably, SHAP reflects model-learned associations, not causation—while higher Scr correlates with higher predicted risk, this does not prove it directly causes delayed excretion.

The SHAP analysis revealed that total MTX dose had the second-highest contribution to predicting delayed MTX excretion, following Scr. Higher doses of MTX were associated with an increased model-predicted risk of delayed excretion. Additionally, ALP was identified as another significant predictor in the model. Lower ALP levels were associated with a model-predicted higher risk of delayed MTX excretion. This observation may reflect underlying anemia or malnutrition due to the primary disease. As malnutrition is known to affect the function of various tissues and organs, it could potentially influence drug metabolism efficiency. However, the model itself cannot establish a causal link.

CK, Urea, and GGT showed relatively moderate contributions in the SHAP feature importance ranking. The SHAP dependence plot for CK revealed a non-linear relationship between CK levels and the predicted risk of delayed MTX excretion: when CK < 50 U/L, lower values were associated with model-predicted risk scores of delayed excretion; in the intermediate range (50–100 U/L), changes in CK levels had minimal impact on the predicted risk; when CK > 100 U/L, higher values were linked to the predicted elevated risk of delayed MTX excretion. Low CK levels may be attributed to several factors, including the younger age of pediatric patients—whose baseline CK activity is naturally lower—or underlying disease-related malnutrition. However, this interpretation represents a clinical hypothesis independent of the model’s findings. The dependence plot for Urea demonstrated that: when Urea levels were below the normal reference range (1.79–6.43 mmol/L), lower values correlated with higher risk of delayed MTX excretion; within the normal range, variations in Urea had no significant effect on the model-predicted risk scores. The dependence plot of GGT indicated: within the normal reference range (8–61 U/L), changes in GGT levels did not significantly alter the predicted risk; when GGT exceeded the normal range, it may increase the risk of delayed MTX excretion. Elevated GGT levels are often indicative of hepatic or biliary dysfunction, such as impaired bile flow or liver damage (Lu and Association, 2022). Since MTX undergoes extensive hepatic metabolism, this may partially explain the association patterns identified by the model.

HB and height demonstrated relatively weaker contributions to the predictive performance of the XGBoost model. However, their SHAP dependence plots still revealed meaningful associations with the risk of delayed MTX excretion. The SHAP dependence plot for HB showed that lower HB levels were associated with a higher model-predicted risk of delayed MTX excretion. The dependence plot for height indicated a non-linear relationship: when height was below 150 cm, variations in height had minimal impact on the predicted risk; when height exceeded 150 cm, increased height was associated with higher predicted risk scores in the model. Generally, taller patients tend to have larger BSA and higher MTX clearance rates (Mei et al., 2018; Hui et al., 2019; Chen et al., 2020; Yang et al., 2020). However, they also receive higher total doses of MTX based on BSA-adjusted dosing protocols. The nonlinear patterns identified by the model may reflect a dynamic interplay between these two opposing effects: in children shorter than 150 cm, the increased clearance associated with greater height may offset the effects of higher drug doses, resulting in no significant net change in the risk of delayed excretion; in contrast, once height exceeds 150 cm, the increase in drug load may outpace the compensatory rise in clearance, leading to a greater likelihood of delayed excretion. This observation requires further investigation and validation in independent cohorts.

This was a large-sample, single-center retrospective cohort study to date investigating risk factors for delayed MTX excretion. A major strength of our work lies in the development of a web-based prediction tool based on the final XGBoost model, which allows clinicians to conveniently input patient-specific features and obtain real-time individualized risk predictions along with interpretable visualizations (e.g., waterfall and force plots). This user-friendly interface enhances the practical applicability of our findings in daily clinical practice. Despite these strengths, several limitations should be acknowledged. First, we did not include drug-drug interactions in our analysis, as the usage rates of potentially interacting medications were all below 10%. This low prevalence may reflect clinical efforts to actively avoid prescribing drugs that could potentially delay MTX excretion. Second, due to the retrospective nature of the study, genetic testing was not routinely performed at the time of treatment, and data on urine output and urine pH were frequently missing. As a result, polymorphisms in drug-metabolizing genes, urine volume, and urinary pH were not included in the analysis. Additionally, all patients in this cohort received standardized hydration, urinary alkalinization, and leucovorin rescue according to the Expert Consensus on High-Dose Methotrexate with Calcium Folinate Rescue for the Treatment of Malignant Tumors. As a result, variables such as hydration rate/volume, degree of alkalinization, and timing/dose of leucovorin were not included in the model. However, the omission of these factors may introduce unmeasured confounding bias. For instance, inadequate fluid intake and low urine output are well-established risk factors for MTX-induced nephrotoxicity, which directly delay drug excretion (Howard et al., 2016; Chen et al., 2020). If not accounted for, the model may underestimate risk in dehydrated patients while overestimating the importance of markers such as serum creatinine (Scr)—which itself can be influenced by volume status. Furthermore, urinary pH is critical for MTX solubility; when pH falls below 7.0, the risk of crystalline nephropathy increases substantially (Bezabeh et al., 2012; Kawaguchi et al., 2021). The current model did not capture this key physiological mechanism, potentially limiting its accuracy in predicting delays caused by intratubular precipitation. Concomitant use of interacting medications—such as proton pump inhibitors (PPIs), which inhibit renal tubular secretion of MTX, or other nephrotoxic agents—can significantly alter MTX pharmacokinetics (Howard et al., 2016). Quantifying the precise impact of these unmeasured confounders is challenging, but their absence may lead to an over-optimistic assessment of model performance. This could restrict the model’s generalizability across diverse clinical settings. The model may be most applicable to patients receiving optimal hydration, without significant drug interactions or additional renal insults. Future studies should aim to address these gaps by conducting multi-center prospective investigations, expanding the range of sample types, and incorporating a broader range of covariates to further refine and validate the predictive performance of the model and improve the model’s utility in supporting clinical decision-making.

In conclusion, machine learning models can serve as reliable tools for predicting the occurrence of delayed MTX excretion during HDMTX therapy in pediatric patients. Among the models evaluated, the XGBoost model, constructed using key predictors including Scr, total drug dose, ALP, CK, Urea, GGT, HB, and height, demonstrated the best predictive performance. This study provides a practical approach for early identification of high-risk patients for MTX excretion delay, thereby supporting personalized treatment decisions. By enabling timely interventions, the model has the potential to reduce the incidence of delayed excretion and associated toxicities, ultimately improving both the safety and adherence of chemotherapy in pediatric oncology.