Developing an Innovative Web-Based Machine Learning Tool for Predicting Delayed Methotrexate Elimination in Pediatric Patients with Osteosarcoma

Developing an Innovative Web-Based Machine Learning Tool for Predicting Delayed Methotrexate Elimination in Pediatric Patients with Osteosarcoma | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 26 March 2025 V1 Latest version Share on Developing an Innovative Web-Based Machine Learning Tool for Predicting Delayed Methotrexate Elimination in Pediatric Patients with Osteosarcoma Authors : Chang Liu 0009-0000-8335-0935 , Minqing Ji , Lichun Wu , Dongsheng Wang , and Fengyi Duan [email protected] Authors Info & Affiliations https://doi.org/10.22541/au.174298884.47399956/v1 188 views 99 downloads Contents Abstract Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract not-yet-known not-yet-known not-yet-known unknown Background： Pediatric osteosarcoma treatment with high-dose methotrexate (HD-MTX) risks delayed clearance due to immature organ function. Interpretable machine learning enables proactive prediction, improving monitoring and reducing toxicity risks effectively. Methods: This retrospective study (2020–2024) on pediatric patients with osteosarcoma treated with HD-MTX aimed to develop robust predictive models. Feature selection was conducted using LASSO regression, followed by 10-fold cross-validation and hyperparameter optimization to enhance model stability and generalizability. Predictive models, including LASSO, ridge, and logistic regression, were developed and rigorously compared to identify the best-performing model. Shapley additive explanations (SHAP) values were used to interpret predictor contributions and relative importance. Additionally, a Shiny-based interactive visualization tool was created for user-friendly clinical integration and data-driven decision-making. Results: This study analyzed 181 pediatric osteosarcoma cases treated with HD-MTX, with 51 experiencing delayed MTX elimination. The dataset was divided into training (117 cases) and test (64 cases) sets, maintaining proportional class distributions. LASSO regression identified seven key predictors through cross-validated error minimization. Three machine learning models (LASSO, logistic regression, ridge regression) were developed. LASSO outperformed the others, achieving an area under the curve (AUC) of 0.8442 across multiple metrics, including ROC, F1 score, and decision curve analysis. Calibration analysis confirmed superior predictive sensitivity for delayed MTX elimination. SHAP analysis ranked Methotrexate Concentration at 3 Hours (MTX3H) as the most critical feature. An interactive Shiny web application was developed to provide personalized predictions and insights into predictor contributions, supporting clinical integration and decision-making. The application is accessible at: https://sclslc.shinyapps.io/shiny_cls2_1model_dalex/. Conclusion： This study presents an interpretable machine-learning model for predicting delayed MTX elimination during high-dose MTX chemotherapy in children with osteosarcoma. Deployed as a web-based tool, the model enables personalized predictions, enhances patient monitoring, reduces toxicity risks, and supports evidence-based clinical decision-making. not-yet-known not-yet-known not-yet-known unknown Developing an Innovative Web-Based Machine Learning Tool for Predicting Delayed Methotrexate Elimination in Pediatric Patients with Osteosarcoma Chang Liu*1, Minqing Ji2, Lichun Wu1, Dongsheng Wang1, Fengyi Duan#2 1. Department of Clinical Laboratory, Sichuan Clinical Research Center for Cancer, Sichuan Cancer Hospital & Institute, Sichuan Cancer Center, University of Electronic Science and Technology of China, Chengdu, China. 2. Pediatric Oncology, Sichuan Clinical Research Center for Cancer, Sichuan Cancer Hospital & Institute, Sichuan Cancer Center, University of Electronic Science and Technology of China, Chengdu 610041, China. * first authors: Chang Liu. # corresponding authors: Fengyi Duan Pediatric Oncology, Sichuan Clinical Research Center for Cancer, Sichuan Cancer Hospital & Institute, Sichuan Cancer Center, University of Electronic Science and Technology of China, Chengdu 610041, China. Email: [email protected] . not-yet-known not-yet-known not-yet-known unknown Abstract Background ： Pediatric osteosarcoma treatment with high-dose methotrexate (HD-MTX) risks delayed clearance due to immature organ function. Interpretable machine learning enables proactive prediction, improving monitoring and reducing toxicity risks effectively. Methods: This retrospective study (2020–2024) on pediatric patients with osteosarcoma treated with HD-MTX aimed to develop robust predictive models. Feature selection was conducted using LASSO regression, followed by 10-fold cross-validation and hyperparameter optimization to enhance model stability and generalizability. Predictive models, including LASSO, ridge, and logistic regression, were developed and rigorously compared to identify the best-performing model. Shapley additive explanations (SHAP) values were used to interpret predictor contributions and relative importance. Additionally, a Shiny-based interactive visualization tool was created for user-friendly clinical integration and data-driven decision-making. Results: This study analyzed 181 pediatric osteosarcoma cases treated with HD-MTX, with 51 experiencing delayed MTX elimination. The dataset was divided into training (117 cases) and test (64 cases) sets, maintaining proportional class distributions. LASSO regression identified seven key predictors through cross-validated error minimization. Three machine learning models (LASSO, logistic regression, ridge regression) were developed. LASSO outperformed the others, achieving an area under the curve (AUC) of 0.8442 across multiple metrics, including ROC, F1 score, and decision curve analysis. Calibration analysis confirmed superior predictive sensitivity for delayed MTX elimination. SHAP analysis ranked Methotrexate Concentration at 3 Hours (MTX3H) as the most critical feature. An interactive Shiny web application was developed to provide personalized predictions and insights into predictor contributions, supporting clinical integration and decision-making. The application is accessible at: https://sclslc.shinyapps.io/shiny_cls2_1model_dalex/. Conclusion： This study presents an interpretable machine-learning model for predicting delayed MTX elimination during high-dose MTX chemotherapy in children with osteosarcoma. Deployed as a web-based tool, the model enables personalized predictions, enhances patient monitoring, reduces toxicity risks, and supports evidence-based clinical decision-making. Keywords ： Osteosarcoma, Machine learning, Methotrexate, Delayed metabolism Introduction Osteosarcoma is a malignant bone tumor that primarily occurs in children and adolescents, representing a significant clinical challenge. Its treatment often requires surgical resection combined with chemotherapy or supportive therapies [1-3] . Among the available options, high-dose methotrexate (HD-MTX) is widely regarded as the cornerstone of first-line chemotherapy. Authoritative clinical guidelines, including the National Comprehensive Cancer Network Clinical Practice Guidelines for Bone Tumors, the European Society for Medical Oncology Guidelines for the Diagnosis, Treatment, and Follow-up of Osteosarcoma , and the 2018 Chinese Clinical Evidence-Based Diagnosis and Treatment Guidelines for Osteosarcoma , recommend intravenous administration of HD-MTX at doses exceeding 8–12 g/m² [2–5] . This dosage is significantly higher than what is typically used for other conditions. However, the high dosage also introduces challenges, with delayed MTX clearance being a frequent clinical issue. HD-MTX is central to osteosarcoma chemotherapy, achieving high drug concentrations in the bloodstream and significantly improving outcomes. However, the maximum dose, approximately three times higher than that used for malignant hematological diseases, increases the risk of delayed drug elimination. This can cause severe hepatotoxicity, nephrotoxicity, and myelosuppression, especially in pediatric patients with underdeveloped liver and immune systems. Their limited ability to metabolize MTX heightens susceptibility to prolonged drug clearance, raising the likelihood of adverse effects and impacting long-term prognosis [5–9] . To address delayed methotrexate elimination, current clinical practice focuses on closely monitoring drug concentrations at specific intervals—24, 48, and 72 h after administration. Interventions such as calcium leucovorin rescue therapy and urine alkalization are applied as needed to facilitate drug elimination more efficiently and safely. While these measures are standard, predictions about delayed drug clearance cannot be made based on a patient’s baseline characteristics or clinical signs before treatment. This lack of predictive capability highlights the urgent need for tools that can identify patients at higher risk of complications earlier in the treatment process. Developing an accurate prediction model for delayed methotrexate clearance could lead to better-targeted interventions, ultimately improving patient outcomes. By using commonly collected laboratory parameters, a predictive model could provide physicians with a practical, non-invasive method for early risk assessment. This would enable more personalized care plans for children undergoing HD-MTX therapy, minimizing the likelihood of adverse reactions while preserving the drug’s therapeutic benefits. Machine learning (ML) algorithms, a branch of artificial intelligence, are powerful tools for addressing complex nonlinear relationships in high-dimensional data. They excel at capturing complex patterns in large datasets, enhancing data analysis, and have been widely applied in medicine for disease prediction [10–13] . Among these advancements, Shapley Additive Explanations (SHAP) has emerged as a state-of-the-art method for visualizing and interpreting the decision-making process of ML models. SHAP intuitively quantifies the contribution of each feature to the model’s predictions, providing a clearer and more interpretable understanding of decision-making while addressing the “black box” limitations of traditional models [14–16] . Osteosarcoma in children and adolescents undergoing HD-MTX chemotherapy poses risks of delayed drug clearance and related complications. This study aims to develop a machine learning-based model and web tool for early detection of delayed MTX clearance, supporting timely clinical interventions. not-yet-known not-yet-known not-yet-known unknown 2. Methods： 2.1 Study design and population This retrospective study assessed MTX dosing and associated laboratory parameters in pediatric and adolescent patients with osteosarcoma treated at Sichuan Cancer Hospital from 2020–2024. The inclusion criteria were: (1) age ≤ 18 years, (2) histologically confirmed diagnosis of osteosarcoma, and (3) administration of MTX-based chemotherapy during hospitalization. The exclusion criteria included (1) insufficient clinical data (≥ 20% missing values; see supplementary Figure.1) and (2) concurrent malignancies other than osteosarcoma. Delayed MTX elimination was defined as serum MTX concentrations exceeding the following thresholds: MTX24H ≥ 10.0 µmol/L, MTX48H ≥ 1.0 µmol/L, and MTX72H ≥ 0.1 µmol/L [17–22]. 2.2 Data collection and definition Demographic data, including age, gender, body surface area (BSA), height, and weight, along with clinical characteristics such as TNM stage and MTX dosage, were collected at baseline. Laboratory tests conducted before the initiation of MTX chemotherapy were analyzed. These tests included urinary, biochemical, hematological, and coagulation parameters. Urinary parameters included urine-specific gravity and pH. Serum biochemical markers included alpha-L-Fucosidase, total protein, albumin, globulin, prealbumin, albumin/globulin ratio, alanine aminotransferase, aspartate aminotransferase, total bilirubin, unconjugated bilirubin (UCB), conjugated bilirubin, total bile acids, chloride, magnesium, phosphorus, calcium (Ca), total carbon dioxide, urea, creatinine, uric acid, estimated glomerular filtration rate, cystatin C, and glucose. Hematological parameters encompassed white blood cell (WBC) count, lymphocyte count and percentage, neutrophil count and percentage, count and percentage, basophil count and percentage, and eosinophil count and percentage (EC). Additionally, the analysis included hemoglobin concentration (HB); red blood cell indices, such as red blood cell count, red blood cell distribution width (both coefficient of variation and standard deviation), hematocrit, mean corpuscular volume, mean corpuscular hemoglobin, and mean corpuscular hemoglobin concentration; as well as platelet count and platelet-related indices, including platelet volume, mean platelet volume (MPV), and platelet distribution width. Coagulation parameters were including activated partial thromboplastin time, thrombin time, prothrombin time, international normalized ratio, fibrinogen, and fibrin degradation products. Serum MTX concentrations were measured at 0 h, 3 h, 24 h, 48 h, and 72 h post-administration (MTX0H, MTX3H, MTX24H, MTX48H, MTX72H, respectively). Subsequently, supervised variable selection methods were employed to analyze and screen the collected parameters. Variables were selected based on statistical assessments of their correlation and significance in predicting delayed elimination of MTX, ensuring the inclusion of clinically and biologically relevant predictors in the model. 2.4 Model Building and Interpretation The process of model development is illustrated in Figure 1. The dataset was randomly divided into a training set (65% of cases) and a holdout test set (35%). To address the common issue of missing data in medical datasets, we used the multivariate imputation by chained equations (MICE) method to impute missing values. Specifically, the training set was imputed first using MICE, and the prediction matrix generated during this process was subsequently applied to the test set to maintain consistency across datasets. The MICE algorithm iteratively imputes missing values for each variable using others as predictors, continuing until convergence or the specified iterations. Predictive mean matching was used for continuous variables, and logistic regression for categorical ones [23, 24]. Class imbalance in medical datasets can bias models by overemphasizing the majority class. To address this, we combined SMOTE for minority class oversampling with random undersampling for the majority class, ensuring a balanced and representative training set. [25] To develop predictive models for the risk of delayed methotrexate excretion, we explored several algorithms, including LASSO regression, ridge regression, and logistic regression. Hyperparameter tuning for each algorithm was conducted using five rounds of 10-fold cross-validation. The final model was trained on the entire training set using the optimal combination of algorithm and resampling method, and its performance was evaluated on the holdout test set to ensure unbiased validation. [26] In clinical practice, the interpretability of machine learning models is crucial, as understanding the contributions of predictor variables is key to informed decision-making. To enhance interpretability, we employed Shapley additive explanations (SHAP), a method derived from game theory. Using the fastshap package, we calculated SHAP values to quantify the contribution of each variable to the model’s predictions. These values were visualized with SHAP plots, offering clear insights into how individual predictors influenced the model’s outputs. This approach not only helped us better understand the role of key predictors in identifying delayed methotrexate excretion but also ensured that the model’s predictions were both accurate and clinically actionable. [27] We developed a web-based application using the R package ”shiny” to make our predictive models accessible online. [28] 2.5 Statistical analysis To describe the characteristics of the data, categorical variables were summarized as frequencies and percentages, while continuous variables were presented as medians with interquartile ranges. To assess differences between groups, we applied the chi-square test for categorical variables and the Mann-Whitney U test for continuous variables, ensuring statistical methods matched to the type of data being analyzed. After defining the statistical approach, all data preprocessing, model training, and analyses were conducted using R software (version 4.3.1, R Foundation for Statistical Computing, Vienna, Austria). During this process, several R packages were employed to streamline the workflow and enhance reproducibility, including DataExplorer, mice, tidymodels, rsample, doParallel, dplyr, recipes, skimr, ROSE, rmda, and ggplot2. To interpret the results, we set the significance threshold at P value of < 0.05, ensuring statistical rigor in identifying meaningful differences or patterns. This significance level was consistently applied throughout the analysis. 3. Results 3.1 Cases characteristics In our dataset of 181cases, 51exhibited metabolic delay, while 130 did not. The dataset was proportionally split at a 6.5:3.5 ratio, resulting in a training set of 117 cases (34 with metabolic delay and 83 without). In the test set of 64 cases, 17 experienced metabolic delay while 47 did not. The comparison of baseline characteristics, including age, sex, height, body surface area (BSA), methotrexate dose, and TNM classification, is summarized in Table 1. The Mann-Whitney U test was for continuous variables employed to assess differences between groups. The analysis revealed statistically significant differences in age, weight, height, BSA, and methotrexate dose. Additionally, Fisher’s exact test indicated significant differences in TNM classification between the two groups, as detailed in Table 1. 3.2 Feature Selection and Data Preprocessing We applied LASSO (Least Absolute Shrinkage and Selection Operator) regression to the 52 candidate predictors identified during data preprocessing To ensure a robust and interpretable model. This method systematically shrinks regression coefficients, retaining the most predictive features while reducing the influence of less relevant variables. The LASSO coefficient paths across log-transformed λ values are shown in Figure 2A, where variables with coefficients approaching zero were considered less impactful. Several key predictors demonstrated significant contributions, with their importance becoming more pronounced as the regularization strength decreased. We selected the seven strongest predictors of delayed MTX elimination to optimize model performance based on the λ value that minimized cross-validation error. The cross-validation error plot displayed in Figure 2B reveals the optimal λ determined through the feature selection process. The final seven predictors chosen for model construction were MTX3H (MTX concentration at 3 h post-administration), WBC (white blood cell count), Ca (calcium levels), UCB, MPV, EC (eosinophil count), and HB (hemoglobin levels), chosen for their significant impact on the outcome variable. The correlations between these seven predictors are summarized in Figure 2C. Each circle represents the relationship between two variables, with size and color indicating the strength and direction of the correlation. Strong positive correlations are represented in deep red, while strong negative correlations appear in deep blue. Most predictor pairs demonstrated weak correlations, highlighting the independence of these features. For example, MTX3H showed modest positive correlations with WBC and EC. This low redundancy suggests that each predictor contributes unique information to the model, improving its generalization ability. The relative importance of predictors in the model was evaluated using the DALEX framework, which provides insight into the contribution of each predictor through explainable artificial intelligence techniques. As illustrated in Figure 2D, the importance of each variable was quantified by measuring its impact on the model’s performance when permuted or removed. Among the seven predictors, MTX3H was identified as the most influential feature, contributing significantly to the model’s accuracy and predictive reliability. MPV ranked as the second most critical predictor, followed by UCB and Ca, which demonstrated moderate contributions. In contrast, WBC and HB had relatively limited influence, and EC showed the smallest impact on the overall predictive performance. 3.3 Model evaluation and comparison In this study, delayed MTX elimination was defined as the positive class, while its absence was defined as the negative class. After selecting relevant variables, the dataset included serum levels of seven key indicators closely associated with the mechanism of delayed MTX elimination. Three machine learning models were developed based on these features: Lasso, logistic regression, and ridge regression. The Lasso model demonstrated superior performance in predicting delayed MTX elimination. As shown in the ROC curve analysis (Figure 3A), the Lasso model achieved an AUC of 0.8442 in both the training and internal validation cohorts, outperforming the logistic regression model (AUC = 0.8056) and the ridge regression model (AUC = 0.8237). A comprehensive comparison in Figure 3B further underscored the superiority of the Lasso model across multiple metrics, including balanced accuracy, F1 score, and the area under the precision-recall curve. Decision curve analysis (Figure 3C) highlighted the clinical utility of the Lasso model, which provided the highest net benefit over a broad range of threshold probabilities compared to logistic and ridge models. Additionally, calibration analysis (Figure 3D) revealed that the Lasso model exhibited a steeper slope in medium-to-high probability ranges, reflecting its enhanced sensitivity in predicting delayed MTX elimination events In summary, the Lasso model emerged as the most effective predictive tool among the three models. Its superior accuracy, robust net benefit, and clinically relevant calibration curve suggest its significant potential for practical application in clinical decision-making. To ensure a comprehensive understanding of the selected variables, we employed the SHAP algorithm to evaluate their predictive significance in the optimal LASSO model for MTX elimination. As illustrated in Figure 4A, MTX3H emerged as the most influential feature, followed by MPV, UCB, Ca, HB, WBC, and EC, reflecting their respective mean SHAP values. The color coding in this figure indicates the intensity of risk, with darker red representing higher risk values and lighter shades indicating lower risks. The influence of these features is further visualized in Figure 4B, where each dot represents an individual observation. The x-axis denotes the SHAP values, where red dots indicate higher feature values associated with elevated risk of delayed MTX elimination, while blue dots represent lower values. This hierarchical representation underscores the critical role these seven indicators play in understanding the mechanisms of MTX elimination in osteosarcoma, highlighting their potential as reliable biomarkers for clinical detection of delayed MTX elimination. We created an interactive web-based application using Shiny to make the final prediction model accessible for individual survival forecasting and interpretation. This platform enables users to generate personalized survival predictions while offering clear insights into the factors influencing each prediction. Furthermore, it provides a global perspective on the model’s overall behavior, enhancing both transparency and usability. The web application, as shown in Figure. 4C can be accessed at: https://sclslc.shinyapps.io/shiny_cls2_1model_dalex/. Discussion: Managing HD-MTX in pediatric osteosarcoma presents unique challenges distinct from its use in ALL, driven by profoundly higher dosing regimens and prolonged treatment durations [29] . This study developed and validated a machine learning-based predictive model to identify delayed MTX clearance in pediatric and adolescent patients with osteosarcoma undergoing HD-MTX chemotherapy. Among the three machine learning models evaluated, the LASSO regression model achieved the best performance metrics, including an AUC of 0.844, highlighting its potential as a clinically useful tool for risk stratification. This model, which integrates seven routinely collected clinical predictors, enables early identification of patients at high risk of delayed MTX elimination, facilitating more proactive and tailored therapeutic interventions. The model’s reliability is underscored by its consistent performance in both internal and cross-validation settings. Optimized by 10-fold cross-validation, The LASSO algorithm achieved an AUC of 0.844 in the holdout test set, with minimal discrepancy between training and validation metrics. This stability reflects the model’s resistance to overfitting, a critical concern in clinical prediction models, achieved through LASSO’s inherent regularization and our hybrid resampling strategy (SMOTE + under-sampling) to address the class imbalance. Furthermore, the model maintained high precision (F1 score: 0.79) and specificity (0.86) in the test cohort, minimizing false positives—a vital attribute for clinical tools designed to prevent unnecessary interventions. The rigorous data preprocessing pipeline further enhanced reliability. MICE preserved dataset integrity by minimizing bias from missing values, while correlation analysis confirmed low multicollinearity among selected predictors, reducing redundancy and ensuring independent contributions from each feature. A key strength of our study lies in the interpretability provided by SHAP, which offered intuitive insights into the contribution of each predictor variable to delayed methotrexate clearance. Among the predictors, MTX3H (serum MTX concentration at 3 h post-administration) was identified as the most influential feature (SHAP value 0.42 ± 0.15). This finding is consistent with methotrexate elimination kinetics, as higher MTX3H levels were strongly associated with delayed drug clearance in this study. Elevated concentrations at this early time point suggest renal tubular saturation and potential impairment in MTX excretion, leading to prolonged systemic exposure and increased toxicity risk. Recognizing MTX3H as a key determinant allows clinicians to prioritize its monitoring during HD-MTX therapy, enabling proactive interventions such as adjusting hydration or alkalization protocols to prevent adverse outcomes. Other baseline predictors, including MPV and UCB, contributed significantly to the predictive model. MPV, an indicator of platelet size variation, may reflect pre-infusion hematologic status, influencing susceptibility to systemic toxicity during HD-MTX therapy. Similarly, elevated UCB levels before infusion suggest pre-existing hepatocellular stress or dysfunction, potentially exacerbating MTX metabolism and clearance challenges. Ca, EC, HB, and WBC delineate the patient’s baseline systemic metabolic and immunological profiles, offering valuable insights into their readiness to tolerate HD-MTX therapy. These findings emphasize the importance of incorporating pre-infusion clinical metrics into predictive frameworks, enhancing risk stratification, and guiding personalized intervention strategies. Based on the advancements achieved in this study, two key areas warrant further exploration to deepen the clinical utility and academic value of the findings: 1) Expanding the clinical validation phase to ensure real-world usability, enabling iterative improvement through clinician feedback and broader clinical adoption, and 2) Utilizing liquid chromatography-mass spectrometry to monitor dynamic biomarkers in plasma, red blood cells, white blood cells, and platelets during MTX infusion. This approach aims to capture real-time pharmacokinetic and toxicodynamic changes, facilitating enhanced biomarker integration and model refinement to optimize prediction accuracy and strengthen clinical decision-making. From a clinical application perspective, our web-based prediction tool addresses two key unmet needs. First, it facilitates real-time risk assessment within the 3-h post-infusion window, unlike current protocols that require 24-, 48-, or 72-h monitoring [30, 31] . Second, the SHAP-driven interpretability framework bridges the ”AI gap” in clinical oncology by offering intuitive visualizations of competing risk factors—an element notably absent from previous MTX prediction models. Conflict of Interest statement The authors declare no competing interests. not-yet-known not-yet-known not-yet-known unknown Acknowledgements We gratefully acknowledge all the staff of the Information Section for their hard work and diligence in collecting cancer information, without which this research could not have been done. References Cole S, Gianferante DM, Zhu B, Mirabello L. Osteosarcoma: A Surveillance, Epidemiology, and End Results program-based analysis from 1975 to 2017. Cancer. 2022 Jun 1;128(11):2107-2118. Ni M. [Update and interpretation of 2021 National Comprehensive Cancer Network (NCCN) ”Clinical Practice Guidelines for Bone Tumors”]. Zhongguo Xiu Fu Chong Jian Wai Ke Za Zhi. 2021 Sep 15;35(9):1186-1191. Chinese. Casali PG, Bielack S, Abecassis N, Aro HT, Bauer S, Biagini R, Bonvalot S, Boukovinas I, Bovee J, Brennan B, Brodowicz T, Broto JM, Brugieres L, Buonadonna A, De Alava E, Deitos AP, Delmuro XG, Dileo P, Dhooge C, Eriksson M, Fagioli F, Fedenko A, Ferraresi V, Ferrari A, Ferrari S, Frezza AM, Gaspar N, Gasperoni S, Gelderblom H, Gil T, Grignani G, Gronchi A, Haas RL, Hassan B, Hecker-Nolting S, Hohenberger P, Issels R, Joensuu H, Jones RL, Judson I, Jutte P, Kaal S, Kager L, Kasper B, Kopeckova K, Krakorova DA, Ladenstein R, Le Cesne A, Lugowska I, Merimsky O, Montemurro M, Morland B, Pantaleo MA, Piana R, Picci P, Piperno-Neumann S, Pousa AL, Reichardt P, Robinson MH, Rutkowski P, Safwat AA, Schoffski P, Sleijfer S, Stacchiotti S, Strauss SJ, Sundby Hall K, Unk M, Van Coevorden F, van der Graaf WTA, Whelan J, Wardelmann E, Zaikova O, Blay JY, Esmo Guidelines Committee, P. Ern E. Bone sarcomas: ESMO-PaedCan-EURACAN Clinical Practice Guidelines for diagnosis, treatment and follow-up. Ann Oncol. 2018;29:79–95. Guo W, Niu X, Xiao J, Cai Z. Clinical evidence-based diagnosis and treatment guidelines for osteosarcoma. Chin J Bone Joint Surg. 2018;11:288–301. Howard SC, McCormick J, Pui CH, Buddington RK, Harvey RD. Preventing and Managing Toxicities of High-Dose Methotrexate. Oncologist. 2016 Dec;21(12):1471-1482. Levêque D, Santucci R, Gourieux B, Herbrecht R. Pharmacokinetic drug-drug interactions with methotrexate in oncology. Expert Rev Clin Pharmacol. 2011 Nov;4(6):743-50. Taylor ZL, Vang J, Lopez-Lopez E, Oosterom N, Mikkelsen T, Ramsey LB. Systematic Review of Pharmacogenetic Factors That Influence High-Dose Methotrexate Pharmacokinetics in Pediatric Malignancies. Cancers (Basel). 2021 Jun 7;13(11):2837. Zhan M, Chen Z, Ding C, Qu Q, Wang G, Liu S, Wen F. Risk prediction for delayed clearance of high-dose methotrexate in pediatric hematological malignancies by machine learning. Int J Hematol. 2021 Oct;114(4):483-493. Nakano T, Kobayashi R, Matsushima S, Hori D, Yanagi M, Suzuki D, Kobayashi K. Risk factors for delayed elimination of high-dose methotrexate in childhood acute lymphoblastic leukemia and lymphoma. Int J Hematol. 2021 May;113(5):744-750. Sengupta PP, Shrestha S. Machine Learning for Data-Driven Discovery: The Rise and Relevance. JACC Cardiovasc Imaging. 2019 Apr;12(4):690-692. doi: 10.1016/j.jcmg.2018.06.030 Li W, Huang G, Tang N, Lu P, Jiang L, Lv J, Qin Y, Lin Y, Xu F, Lei D. Effects of heavy metal exposure on hypertension: A machine learning modeling approach. Chemosphere. 2023 Oct;337:139435. Hu P, Wu Y, Yan T, Shu L, Liu F, Xiao B, Ye M, Wu M, Lv S, Zhu X. Deep learning-based quantification of total bleeding volume and its association with complications, disability, and death in patients with aneurysmal subarachnoid hemorrhage. J Neurosurg. 2024 Mar 29;141(2):343-354. Njei B, Osta E, Njei N, Al-Ajlouni YA, Lim JK. An explainable machine learning model for prediction of high-risk nonalcoholic steatohepatitis. Sci Rep. 2024 Apr 13;14(1):8589. Sevakula RK, Au-Yeung WM, Singh JP, Heist EK, Isselbacher EM, Armoundas AA. State-of-the-Art Machine Learning Techniques Aiming to Improve Patient Outcomes Pertaining to the Cardiovascular System. J Am Heart Assoc. 2020 Feb 18;9(4):e013924. Molenaar JM, Leung KY, van der Meer L, Klein PPF, Struijs JN, Kiefte-de Jong JC. Predicting population-level vulnerability among pregnant women using routinely collected data and the added relevance of self-reported data. Eur J Public Health. 2024 Dec 1;34(6):1210-1217. Jian C, Chen S, Wang Z, Zhou Y, Zhang Y, Li Z, Jian J, Wang T, Xiang T, Wang X, Jia Y, Wang H, Gong J. Predicting delayed methotrexate elimination in pediatric acute lymphoblastic leukemia patients: an innovative web-based machine learning tool developed through a multicenter, retrospective analysis. BMC Med Inform Decis Mak. 2023 Aug 3;23(1):148. Howard SC, McCormick J, Pui CH, Buddington RK, Harvey RD. Preventing and Managing Toxicities of High-Dose Methotrexate. Oncologist. 2016 Dec;21(12):1471-1482. Young EP, Cheng WS, Bernhardt MB, Wang LL, Rainusso N, Foster JH. Risk factors associated with delayed methotrexate clearance and increased toxicity in pediatric patients with osteosarcoma. Pediatr Blood Cancer. 2020 Apr;67(4):e28123. Wang Y, Zhang H, Chen Y, Xu H, Gao P, Liu M, et al. Predictive analysis of methotrexate elimination delay based on logistic regression model and ROC curve. Chin J Hosp Pharm. 2015;35(07):618–22. Schmidt D, Kristensen K, Schroeder H, Wehner PS, Rosthøj S, Heldrup J, Damsgaard L, Schmiegelow K, Mikkelsen TS. Plasma creatinine as predictor of delayed elimination of high-dose methotrexate in childhood acute lymphoblastic leukemia: A Danish population-based study. Pediatr Blood Cancer. 2019 Jun;66(6):e27637. Tsurusawa M, Gosho M, Mori T, Mitsui T, Sunami S, Kobayashi R, Fukano R, Tanaka F, Fujita N, Inada H, Koh K, Takimoto T, Saito A, Fujimoto J, Nakazawa A, Horibe K; lymphoma committee of the Japanese Pediatric Leukemia/lymphoma Study Group. Statistical analysis of relation between plasma methotrexate concentration and toxicity in high-dose methotrexate therapy of childhood nonHodgkin lymphoma. Pediatr Blood Cancer. 2015 Feb;62(2):279-284. Qian X, Li L, Wang Q, Li L, Xie J. Research progress of high dose methotrexate-calcium folinate to rescue children with acute lymphoblastic leukemia. Chin J Hosp Pharm. 2016;36(18):1613–8. Azur MJ, Stuart EA, Frangakis C, Leaf PJ. Multiple imputation by chained equations: what is it and how does it work? Int J Methods Psychiatr Res. 2011 Mar;20(1):40-9. Burns RA, Butterworth P, Kiely KM, Bielak AA, Luszcz MA, Mitchell P, Christensen H, Von Sanden C, Anstey KJ. Multiple imputation was an efficient method for harmonizing the Mini-Mental State Examination with missing item-level data. J Clin Epidemiol. 2011 Jul;64(7):787-93. Blagus R, Lusa L. SMOTE for high-dimensional class-imbalanced data. BMC Bioinformatics. 2013 Mar 22;14:106. Lee CC, Su SY, Sung SF. Machine learning-based survival analysis approaches for predicting the risk of pneumonia post-stroke discharge. Int J Med Inform. 2024 Jun;186:105422. Lundberg SM, Erion G, Chen H, DeGrave A, Prutkin JM, Nair B, Katz R, Himmelfarb J, Bansal N, Lee SI. From Local Explanations to Global Understanding with Explainable AI for Trees. Nat Mach Intell. 2020 Jan;2(1):56-67. Jia L, Yao W, Jiang Y, Li Y, Wang Z, Li H, Huang F, Li J, Chen T, Zhang H. Development of interactive biological web applications with R/Shiny. Brief Bioinform. 2022 Jan 17;23(1):bbab415. Widemann BC, Balis FM, Kempf-Bielack B, Bielack S, Pratt CB, Ferrari S, Bacci G, Craft AW, Adamson PC. High-dose methotrexate-induced nephrotoxicity in patients with osteosarcoma. Cancer. 2004 May 15;100(10):2222-32. Biermann JS, Chow W, Reed DR, Lucas D, Adkins DR, Agulnik M, Benjamin RS, Brigman B, Budd GT, Curry WT, Didwania A, Fabbri N, Hornicek FJ, Kuechle JB, Lindskog D, Mayerson J, McGarry SV, Million L, Morris CD, Movva S, O’Donnell RJ, Randall RL, Rose P, Santana VM, Satcher RL, Schwartz H, Siegel HJ, Thornton K, Villalobos V, Bergman MA, Scavone JL. NCCN Guidelines Insights: Bone Cancer, Version 2.2017. J Natl Compr Canc Netw. 2017 Feb;15(2):155-167. Niu XH. [Interpretation of 2020 NCCN Clinical Practice Guidelines in Oncology-Bone Cancer]. Zhonghua Wai Ke Za Zhi. 2020 Jun 1;58(6):430-434. Chinese. Legends Figure 1. Study flow chart Figure 2. (A) LASSO Coefficient Path: The x-axis shows the log(λ) values, and the y-axis represents variable coefficients. Variables persisting longer before coefficients shrink to zero are more influential. (B) Cross-validation Curve: Red dots indicate the mean cross-validated errors, and dashed lines represent Lambda.min (optimal accuracy) and Lambda.1se (simpler model). (C) Correlation Matrix: Circle size and color represent the strength and direction of feature correlations. (D) Feature Importance: The bar plot shows feature importance based on AUC loss after permutation, with higher bars indicating greater importance. Figure 3. Results of Model Evaluation on Test Data (A) ROC curves for Logistic Regression (red), LASSO (green), and Ridge Regression (blue), comparing predictive performance via AUC. (B) Performance metrics for each model, including accuracy, balanced accuracy, F1 score, MCC, PPV, NPV, precision, PR AUC, recall, ROC AUC, sensitivity, and specificity. (C) Decision Curve Analysis (DCA) showing the net benefit of Logistic (red), LASSO (green), and Ridge (blue) models across varying threshold probabilities. (D) Calibration plots assessing agreement between predicted probabilities and observed outcomes for each model. Figure 4. (A) Mean SHAP Values: Bar chart showing the relative importance of the seven predictors in the model. MTX3H has the highest impact, followed by MPV, UCB, and Ca, while WBC, HB, and EC have smaller contributions. (B) SHAP Summary Plot: Scatter plot illustrating how each feature influences predictions. Red represents higher feature values, and blue represents lower values. MTX3H and MPV show the strongest effects on the outcome. (C) Shiny App for Prediction: A user-friendly interface displaying prediction results and explanations for individual cases. The SHAP force plot highlights how each feature contributes to the predicted probability of delayed MTX elimination (DELY = Yes). Table 1 Clinical Data Baseline Table not-yet-known not-yet-known not-yet-known unknown SEX (%) 181 117 64 female 84 (46.4%) 52 (44.4%) 32 (50.0%) 0.53 male 97 (53.6%) 65 (55.6%) 32 (50.0%) AGE (year)(median [IQR]) 12.00 [9.00, 15.00] 12.00 [9.00, 15.00] 12.00 [7.75, 14.00] 0.12 HIGH (cm)(median [IQR]) 156.00 [133.00, 162.00] 156.00 [136.00, 164.00] 150.00 [130.00, 162.00] 0.14 WEIGHT (Kg)(median [IQR]) 37.50 [27.00, 48.00] 39.50 [28.00, 49.00] 35.00 [26.75, 47.25] 0.21 BSA(median [IQR]) 1.28 [1.02, 1.48] 1.32 [1.06, 1.49] 1.21 [1.01, 1.45] 0.22 Methotrexate(g)(median [IQR]) 10.00 [8.00, 12.00] 10.00 [8.00, 12.00] 8.75 [8.00, 11.25] 0.19 DELY (%) No 130 (71.8%) 83 (70.9%) 47 (73.4%) 0.86 Yes 51 (28.2%) 34 (29.1%) 17 (26.6%) TNM stage (%) 2 4 (2.2%) 3 (2.6%) 1 (1.6%) 0.86 3 113 (62.4%) 71 (60.7%) 42 (65.6%) 4 64 (35.4%) 43 (36.8%) 21 (32.8%) T stage (%) 1 4 (2.2%) 3 (2.6%) 1 (1.6%) 0.42 2 75 (41.4%) 47 (40.2%) 28 (43.8%) 3 89 (49.2%) 61 (52.1%) 28 (43.8%) 4 13 (7.2%) 6 (5.1%) 7 (10.9%) N stage (%) 0 147 (81.2%) 95 (81.2) 52 (81.2) 0.72 1 32 (17.7%) 20 (17.1) 12 (18.8) Not available 2 (1.1%) 2 (1.7%) 0 (0.0%) M stage (%) 0 128 (70.7%) 79 (67.5%) 49 (76.6%) 0.60 1 50 (27.6%) 35 (29.9%) 15 (23.4%) 2 1 (0.6%) 1 (0.9%) 0 (0.0%) Not available 2 (1.1%) 2 (1.7%) 0 (0.0%) MTX3H (umol/L)(median [IQR]) 266 [189.6, 348.5] 269.07[196.2, 349.89] 257.5 [182.515, 348.5] 0.61 WBC (10^9/L)(median [IQR]) 5.63 [4.01, 9.72] 5.58 [3.99, 8.95] 5.73[4.2325, 10.925] 0.55 Ca(mmol/L)(median [IQR]) 2.32[2.24, 2.4] 2.33[2.24, 2.4] 2.295 [2.2275, 2.42] 0.49 UCB (umol/L)(median [IQR]) 15.8 [10, 20.6] 15.5 [9.7, 20.5] 17.05[11.175, 20.725] 0.36 MPV (fL)(median [IQR]) 9.2 [8.2, 10.1] 9.3[8.2, 10.1] 9.15[8.375, 10.325] 0.86 EC (10^9/L)(median [IQR]) 0.03 [0.01, 0.13] 0.03 [0.01, 0.13] 0.025[0.01, 0.1425] 0.86 HB(g/L)(median [IQR]) 106[98, 120] 106 [98, 119] 105[97, 120.25] 0.83 not-yet-known not-yet-known not-yet-known unknown 1.Continuous variables are presented as median（Interquartile Range [IQR]） 2.Categorical variables are presented as n (%) 3.Continuous variables were compared using Mann-Whitney U test 4.Categorical variables were compared using Fisher’s exact test Information & Authors Information Version history V1 Version 1 26 March 2025 Copyright This work is licensed under a Non Exclusive No Reuse License. Keywords methotrexate osteosarcoma prediction modeling in cancer Authors Affiliations Chang Liu 0009-0000-8335-0935 Sichuan Cancer Hospital and Institute View all articles by this author Minqing Ji Sichuan Cancer Hospital and Institute View all articles by this author Lichun Wu Sichuan Cancer Hospital and Institute View all articles by this author Dongsheng Wang Sichuan Cancer Hospital and Institute View all articles by this author Fengyi Duan [email protected] Sichuan Cancer Hospital and Institute View all articles by this author Metrics & Citations Metrics Article Usage 188 views 99 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Chang Liu, Minqing Ji, Lichun Wu, et al. Developing an Innovative Web-Based Machine Learning Tool for Predicting Delayed Methotrexate Elimination in Pediatric Patients with Osteosarcoma. Authorea . 26 March 2025. DOI: https://doi.org/10.22541/au.174298884.47399956/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.174298884.47399956/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a002e6da4c39f047',t:'MTc3OTUyNzE1Ng=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();