Interpretable machine learning models for survival prediction in prostate cancer bone metastases

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Prostate cancer bone metastasis (PCBM) presents a significant survival challenge.
• Current clinical models for PCBM survival prediction lack precision.
• Accurate prognostic tools are crucial for managing PCBM patients.

Purpose of the Study:

• To develop and validate machine learning models for enhanced PCBM survival prediction.
• To identify key prognostic factors influencing PCBM patient survival.
• To assess the clinical applicability of developed predictive models.

Main Methods:

• Utilized SEER database (2010-2019) for PCBM patient data.
• Employed univariate and multivariate Cox regression for feature identification.
• Developed and validated XGBoost models with five-fold cross-validation.
• Assessed model performance using AUC and accuracy; feature importance via SHAP values.
• Conducted decision curve and Kaplan-Meier analyses.

Main Results:

• XGBoost models demonstrated robust performance with AUCs of 0.76 (1-year), 0.83 (3-year), and 0.91 (5-year).
• Significant prognostic factors included T stage, grade, age, PSA, and Gleason score.
• Chemotherapy and radiotherapy were associated with improved survival; surgery was not significant in multivariate analysis.
• Single marital status and lower income correlated with higher mortality risk.

Conclusions:

• XGBoost models offer high accuracy and interpretability for PCBM survival prediction.
• Identified prognostic factors and treatment impacts can guide personalized treatment strategies.
• A web-based tool was developed for clinical integration of the predictive model.