Causality-Guided Machine Learning for Retinoblastoma Survival Prediction: Development and Comparative Evaluation...

Area of Science:

• Oncology
• Biostatistics
• Machine Learning
• Causal Inference

Background:

• Retinoblastoma (RB) presents unique challenges for survival prediction due to small sample sizes and low event rates in rare pediatric cancers.
• Conventional association-driven feature selection methods can lead to unstable models, overadjustment, and limited generalizability in rare diseases.
• Existing survival prediction studies often neglect the importance of underlying causal structure in feature selection.

Purpose of the Study:

• To develop and validate a novel causality-guided machine learning model for retinoblastoma (RB) survival prediction.
• To jointly incorporate survival time and survival status as outcome variables within the predictive model.
• To address the challenges of feature selection in rare-disease data by integrating causal structure.

Main Methods:

• Analysis of 1015 retinoblastoma patients from the SEER database (1975-2020).
• A three-step causality-informed feature selection framework: univariate Cox proportional hazards analysis, causal structure learning (PC algorithm, DAG construction), and LASSO-based screening.
• Survival models (including random survival forest) were trained using the refined, causally informed feature set for performance evaluation.

Main Results:

• The causality-informed framework consistently identified four robust predictors: laterality, 'SEER historic stage A', 'RX Summ', and sequence number.
• This approach reduced the feature set size, enhanced model stability and interpretability, and demonstrated improved predictive performance compared to conventional methods.
• The random survival forest model achieved a C-index of 0.739, indicating reliable predictive performance with reduced overfitting in this low-event setting.

Conclusions:

• Incorporating causal structure into feature selection provides a more reliable and interpretable foundation for survival modeling in retinoblastoma.
• Causality-informed feature selection is critical for improving robustness in rare-disease prediction tasks, moving beyond mere algorithmic comparison.
• The developed framework offers a generalizable methodological template for survival prediction in other rare clinical settings prone to spurious associations.