Machine learning-driven conditional survival prediction model for atypical teratoid/rhabdoid tumor.

PURPOSE: Atypical teratoid/rhabdoid tumor (AT/RT) is a rare and highly malignant pediatric brain tumor with dismal prognosis. We aimed to characterize dynamic survival patterns using conditional survival (CS) and annual hazard rate (AHR) analyses, identify key prognostic factors through machine learning (ML)-based feature selection, and develop an interpretable CS-nomogram for individualized prognostication. METHODS: Data of AT/RT patients diagnosed between 2000 and 2022 were obtained from the SEER database. CS and AHR analyses were performed to assess temporal survival dynamics. Four complementary algorithms-LASSO regression, Boruta algorithm, stepwise regression, and best subset regression-were used for feature selection. The final CS-nomogram was validated using ROC curves, calibration plots, and decision curve analysis (DCA). Model interpretability was evaluated with Shapley additive explanations (SHAP), and a web-based calculator was created for clinical application. RESULTS: A total of 382 patients were analyzed. The 5-year survival probability at diagnosis was 34.11%,but increased to 95.14% among those surviving beyond 4 years, while AHR declined from 44.76% in the first year to < 5% after 5 years. Tumor extension, radiotherapy, and chemotherapy were consistently identified as key predictors. The CS-nomogram showed excellent discrimination (1-, 3-, and 5-year AUCs: 0.860, 0.796, and 0.762 in training; 0.857, 0.764, and 0.744 in validation), and SHAP analysis confirmed radiotherapy, chemotherapy, and tumor extension as major contributors. CONCLUSION: CS analysis revealed a marked improvement in long-term survival among AT/RT survivors. The ML-based CS-nomogram provides a robust, interpretable, and clinically applicable tool for dynamic, individualized prognostication and personalized follow-up planning in this rare malignancy.