Interpretable machine learning to identify health system levers for survival outcomes of patients with prostate cancer.

Abstract

1595 Background: Global disparities in prostate cancer outcomes are among the most pronounced across all cancer types. While incidence continues to rise globally, mortality reductions are more pronounced in countries with advanced healthcare systems. Understanding how health system factors relate to prostate cancer outcomes is critical for targeted policy interventions, particularly as the cancer burden is projected to rise substantially. We used explainable machine learning models to examine how these factors are associated with prostate cancer outcomes across countries. Methods: We compiled national mortality-to-incidence ratios (MIRs) for prostate cancer from GLOBOCAN 2022 and integrated them with system-level indicators from the WHO, the World Bank, UN agencies, and the Directory of Radiotherapy Centres (DIRAC). Variables included GDP per capita, the Universal Health Coverage (UHC) index, radiotherapy center density, physician and nursing workforce capacity, pathology services, and health expenditure composition. We implemented a CatBoost ensemble with repeated cross-validation and integrated SHAP (SHapley Additive exPlanations) for feature attribution to quantify country-level determinants of prostate cancer MIR while accounting for nonlinear, context-dependent relationships. Results: The model demonstrated strong predictive performance (R² = 0.796, RMSE = 0.078, correlation = 0.892). SHAP analysis identified radiotherapy center density, UHC index, and GDP per capita as the leading system drivers of MIR. Countries with robust radiotherapy capacity and comprehensive UHC consistently achieved lower MIR, whereas higher generic health spending alone showed a weaker correlation with outcomes. Yemen exhibited the highest MIR (0.720), driven by deficits in GDP, gender inequality, radiotherapy infrastructure, and UHC. The United States achieved the lowest MIR (0.107), driven by extensive radiotherapy density, high health expenditure, and robust coverage systems. Country-specific SHAP decompositions revealed heterogeneity: in higher-income settings, radiotherapy and health workforce predominantly lowered MIR, whereas in lower-SDI countries, lack of insurance and infrastructure remained major barriers. Conclusions: SHAP-empowered machine learning accurately predicts country-level prostate cancer MIR and provides actionable policy guidance. Investments in radiotherapy infrastructure and universal health coverage are likely to yield greater reductions in mortality than undifferentiated budget increases. This reproducible framework enables data-driven resource allocation and supports a precision-aligned approach to global cancer control, emphasizing equity, access, and the optimization of system-specific interventions.