Multi-Model Machine Learning for Survival Predictions for Castration-Resistant Prostate Cancer

Area of Science:

• Oncology
• Medical Informatics
• Biostatistics

Background:

• Accurate survival prediction is crucial for optimizing treatment in castration-resistant prostate cancer (CRPC).
• Traditional models struggle with complex data interactions and limited variable inclusion.
• Machine learning (ML) offers potential for improved prognostic accuracy in CRPC.

Purpose of the Study:

• To develop and compare ML models for predicting cancer-specific mortality (CSM) and overall mortality (OM) in CRPC patients.
• To evaluate the performance and interpretability of various ML algorithms for survival prediction.
• To identify key predictors of survival outcomes in CRPC.

Main Methods:

• Retrospective analysis of 46 variables from 801 CRPC patients.
• Development of ML models: Random Survival Forests (RSF), XGBoost, LightGBM, and logistic regression.
• Performance evaluation using C-index, AUC, accuracy, precision, recall, F1-score, and SHAP for interpretability.

Main Results:

• RSF achieved the highest C-index for CSM (0.772) and OM (0.771) in the test set.
• RSF excelled in 2-year survival prediction, while XGBoost was superior for 3-year survival prediction (F1-score).
• Key predictors identified: time to first-line CRPC treatment, hemoglobin, and alkaline phosphatase levels.

Conclusions:

• RSF provides robust time-to-event prediction, and XGBoost offers complementary value for 3-year survival classification in CRPC.
• Developed ML models offer accurate and interpretable prognostic tools for personalized treatment strategies.
• External validation and integration of new therapies are needed for broader clinical application.