Development of a machine learning-based prognostic prediction model and a web-based tool for pediatric hepatoblastoma: a Surveillance, Epidemiology, and End Results (SEER) database study.

backgroundPrognostic evaluation of pediatric hepatoblastoma (HB) remains challenging due to the low accuracy of traditional risk stratification models. This study aims to leverage the U.S. Surveillance, Epidemiology, and End Results (SEER) database to develop and compare several machine learning-based survival models and create an online prediction tool for individualized survival probability estimation in children with HB.

methodsA total of 614 pediatric HB patients diagnosed from 2000 to 2021 in the SEER database were identified and randomly split (7:3 ratio) into a training set (n = 429) and an internal hold-out validation cohort (n = 185). This split was selected to retain sufficient data for model development while preserving an adequate validation cohort for performance assessment. Least Absolute Shrinkage and Selection Operator (LASSO) regression analysis was used for feature selection, and three prognostic prediction models (Cox regression, random survival forest (RSF), and a deep learning-based survival (DeepSurv) neural network) were developed using the training data. Model performance was evaluated by Harrell's concordance index (C-index), time-dependent area under the receiver operating characteristic (ROC) curve (AUC) at 1, 3, and 5 years, and decision curve analysis (DCA). We also used SHapley Additive exPlanations (SHAP) analysis to interpret the RSF model. Based on RSF-derived risk scores, patients were classified into high-, intermediate-, and low-risk groups for survival analysis. Ultimately, a web-based tool was developed to enable real-time prediction of 1-, 3-, and 5-year survival probabilities using individual patient characteristics.

resultsIn the validation cohort, the RSF model achieved the highest C-index (0.745), outperforming the DeepSurv (0.720) and Cox regression models (0.709). The RSF also yielded favorable 1-year, 3-year, and 5-year AUC values, and DCA indicated a greater net benefit across clinically relevant threshold probabilities. SHAP analysis highlighted distant-stage disease, surgical treatment status, and tumor extent as key factors influencing survival predictions. Kaplan-Meier curves stratified by the RSF risk groups showed significantly different survival outcomes among high-, intermediate-, and low-risk patients (log-rank P < 0.001 for all comparisons). The finalized online tool allows users to input patient characteristics and obtain estimated 1-year, 3-year, and 5-year survival probabilities, which may assist individualized prognostic assessment as an adjunctive reference.

conclusionThe RSF-based prognostic model showed favorable predictive performance compared with the Cox and DeepSurv models in this retrospective SEER-based cohort of pediatric HB. With interpretable model outputs and an accessible web-based interface, the model may serve as an adjunctive tool for individualized risk assessment. However, external validation and prospective evaluation are required before routine clinical implementation.