Prediction of pediatric Wilms tumor recurrence using interpretable machine learning models: insights from a 20-year
Honggang Fang1,2,3,4, Yihang Yu1,2,3,4, Junjun Dong1,2,3,4
1Department of Urology, Children's Hospital of Chongqing Medical University Chongqing 401122, China.
Insights
Machine learning models accurately predict Wilms tumor (WT) recurrence using clinical data. The best model identifies key risk factors, enabling personalized follow-up for pediatric cancer patients.
Area of Science:
- Pediatric Oncology
- Machine Learning in Medicine
- Cancer Recurrence Prediction
Background:
- Wilms tumor (WT) has high survival rates, but ~15% of pediatric patients experience recurrence, facing poor outcomes.
- Current prediction methods lack precision, necessitating advanced tools for risk stratification and individualized follow-up.
- Significant long-term complications in recurrent WT patients underscore the need for improved post-operative management.
Purpose of the Study:
- To develop and validate machine learning (ML) models for predicting postoperative WT recurrence.
- To identify key clinical predictors associated with WT recurrence risk.
- To create a practical tool for risk stratification to guide individualized patient follow-up.
Main Methods:
- Utilized real-world clinical data from 351 pediatric WT patients undergoing radical surgery.
- Employed feature selection (Boruta, LASSO, subgroup analysis, univariate logistic regression) to identify seven key predictors.
- Constructed and evaluated nine ML models (including SVM, XGBoost, RF) using metrics like AUC, accuracy, and F1 score.
Main Results:
- Identified seven independent predictors: COG tumor stage, age, tumor rupture, histology, tumor thrombus, tumor volume, and Ki-67 index.
- The Support Vector Machine (SVM) model showed superior performance in the validation cohort (AUC=0.851, accuracy=0.830).
- SHAP analysis revealed unfavorable histology, advanced COG stage (IV-V), tumor thrombus, and high Ki-67 index as strong recurrence predictors.
Conclusions:
- An interpretable SVM-based ML model effectively predicts WT recurrence using seven key clinical factors.
- The Ki-67 index emerged as a particularly strong predictor of recurrence risk.
- This model serves as a valuable tool for risk stratification and personalized follow-up planning in pediatric WT patients.
Abstract:
Despite overall survival exceeding 80% in Wilms tumor (WT), approximately 15% of pediatric patients experience recurrence with poor post-relapse survival (~50%) and significant long-term complications, highlighting an unmet need for precise recurrence prediction. To address this, we developed and validated machine learning (ML) models for predicting postoperative WT recurrence using real-world clinical data. Among 476 pediatric WT patients who underwent radical surgery at our institution (June 2004-June 2024), 351 met inclusion criteria and were randomized into training (70%) and validation (30%) cohorts. Seven independent predictors - COG tumor stage, age, tumor rupture, histological subtype (COG classification), tumor thrombus, tumor volume, and Ki-67 index - were identified via feature selection intersection of Boruta, LASSO, subgroup analysis, and univariate logistic regression. Predictive models were constructed using nine ML algorithms (DT, LASSO, KNN, LightGBM, LR, MLP, RF, SVM, XGBoost), with performance evaluated using metrics such as AUC, accuracy, F1 score, specificity, positive predictive value (PPV), negative predictive value (NPV), and confusion matrix. Among the included patients, 51 (14.53%) experienced tumor recurrence (including 7 multi-site relapses), with median time to recurrence 6 months (IQR 4-16 months) and 80.4% occurring within the first postoperative year. In the validation cohort, the SVM model demonstrated the best performance (AUC = 0.851; accuracy = 0.830; specificity = 0.856; F1 = 0.550; PPV = 0.458; NPV = 0.939), and SHAP analysis highlighted unfavorable histology, COG stage IV-V, tumor thrombus, and elevated Ki-67 index as the strongest contributors to recurrence risk. This interpretable SVM-based model confirms seven key predictors - especially the Ki-67 index - and serves as a practical risk stratification tool to support individualized follow-up planning.
