Leveraging Machine Learning Models to Explore Disparities in Prostate Cancer Diagnosis, Treatment, and Survival

Introduction Prostate cancer (CaP) is the second most common cancer in men globally and a leading cause of cancer-related mortality, particularly among older men. In the United States, disparities in incidence, stage at diagnosis, and outcomes persist across racial, socioeconomic, and geographic lines. Men in medically underserved regions like Appalachia experience higher mortality due to limited access to timely screening and treatment. Given the complex interplay of clinical, behavioral, and sociodemographic factors, machine learning (ML) offers promise in identifying survival predictors that traditional models may overlook. This study applies ML models to assess the impact of age at diagnosis, treatment modality, and other sociodemographic and clinical factors on CaP survival outcomes using data from the Kentucky Cancer Registry (KCR). Methods We retrospectively analyzed 37 893 CaP cases diagnosed from 2010 to 2022 using KCR data linked to mortality records. Kaplan-Meier (KM), Random Survival Forest (RSF), and Elastic Net regression were used to estimate survival, assess variable importance, and evaluate predictive performance. Missing data were handled via multiple imputation, and leave-one-out cross-validation minimized overfitting. Models were compared using out-of-bag error, continuous ranked probability scores (CRPS), and interpretability. Results ML models identified age at diagnosis, treatment modality, and smoking status as the top survival predictors. RSF showed that age was approximately 2.5 times more influential than treatment type. Patients diagnosed before age 60 were more likely to undergo surgery and had lower mortality, while older men more often received non-surgical therapies and experienced worse outcomes. Insurance status, tumor grade, lymph node involvement, and marital status also affected survival, with pronounced disparities among uninsured and government-insured patients. Conclusions ML models highlighted age, smoking, and treatment type as key predictors of CaP survival. Findings support early screening, equitable treatment access, and behavioral health integration to reduce disparities in high-risk areas such as Kentucky and Appalachia.