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

Area of Science:

• Pediatric neuro-oncology
• Biostatistics
• Machine learning in medicine

Background:

• Atypical teratoid/rhabdoid tumor (AT/RT) is a rare, highly malignant pediatric brain tumor with a poor prognosis.
• Understanding dynamic survival patterns is crucial for improving patient outcomes and follow-up strategies.

Purpose of the Study:

• To characterize dynamic survival patterns in AT/RT using conditional survival (CS) and annual hazard rate (AHR) analyses.
• To identify key prognostic factors for AT/RT using machine learning (ML).
• To develop an interpretable CS-nomogram for individualized prognostication and clinical application.

Main Methods:

• Analysis of 382 AT/RT patients from the SEER database (2000-2022).
• Conditional survival (CS) and annual hazard rate (AHR) analyses for temporal survival dynamics.
• Machine learning (ML) feature selection (LASSO, Boruta, stepwise, best subset regression) and nomogram development.
• Validation using ROC curves, calibration plots, decision curve analysis (DCA), and SHAP for interpretability.

Main Results:

• 5-year survival probability at diagnosis was 34.11%, increasing to 95.14% for survivors beyond 4 years.
• Annual hazard rate (AHR) decreased significantly from 44.76% in year 1 to <5% after 5 years.
• Tumor extension, radiotherapy, and chemotherapy were identified as key predictors.
• The CS-nomogram demonstrated excellent discrimination in both training and validation cohorts (AUCs > 0.744).

Conclusions:

• Conditional survival analysis highlights a substantial improvement in long-term survival for AT/RT patients.
• The developed ML-based CS-nomogram is a robust, interpretable, and clinically applicable tool.
• This tool facilitates dynamic, individualized prognostication and personalized follow-up planning for AT/RT.