Data-driven survival modeling for breast cancer prognostics: A comparative study with machine learning and...

Area of Science:

• Oncology
• Biostatistics
• Data Science in Healthcare

Background:

• Investigating data-driven survival modeling for prognostic assessment in breast cancer.
• Comparing machine learning (ML) and conventional survival analysis techniques for identifying breast cancer survival predictors.

Purpose of the Study:

• To evaluate and compare the predictive capabilities of ML and survival analysis models for breast cancer survival.
• To identify consistent key predictors of breast cancer survival outcomes across different modeling approaches.

Main Methods:

• Employed survival-specific models (Cox Proportional Hazards, Random Survival Forests, DeepSurv) and ML models (Random Forests, XGBoost, SVM, LightGBM).
• Utilized Shapley Additive Explanation (SHAP) for interpretability and identification of key predictors.
• Analyzed data from 4,024 women diagnosed with breast cancer (2006-2010) from the SEER program.

Main Results:

• Survival-specific models (Cox, RSF) demonstrated accurate survival probability predictions (lowest Integrated Brier Score).
• ML models showed fair discriminatory ability but did not account for censoring.
• Key predictors identified include lymph node involvement, tumor grade, progesterone status, and age.

Conclusions:

• Survival-specific models are more suitable for accurate survival predictions due to their ability to handle time-to-event data and censoring.
• ML models enhance interpretability in identifying key variables, complementing survival models.
• Integrating ML and survival models offers valuable, personalized insights for clinical decision-making and improved patient care.