Depressive Symptoms and Cognitive Function in Late-Stage Prostate Cancer: Transformer-Based Survival Modeling From the CHARLS 2011-2015 Big Data Cohort.

BACKGROUND: To examine whether depressive symptoms and multi-domain cognition predict 4-year all-cause mortality among men with survey-defined late-stage prostate cancer (PCa), and to evaluate an attention-based discrete-time survival model for short-horizon risk stratification in this urologic oncology setting. METHODS: Men ≥ 45 years in China Health and Retirement Longitudinal Study (CHARLS) 2011-2015 were assembled; incident PCa with ≥1 survey-available systemic/palliative indicator was used to approximate a late-stage phenotype (n = 21) and compared with cancer-free controls frequency-matched by age band and province (n = 600; N = 621). Exposures were the 10-item Center for Epidemiologic Studies Depression Scale (CES-D) and a global cognitive z-score. A tabular transformer modeled discrete-time hazards with isotonic calibration and was compared with discrete-time logistic survival and Cox models. Province-blocked internal-external validation and decision-curve analysis evaluated discrimination, calibration, transportability, and clinical utility. RESULTS: Over four years, 51 deaths occurred (late-stage PCa 33.3% vs controls 7.3%). The calibrated transformer achieved an area under the receiver operating characteristic curve (AUC) at 4 years of 0.828 (Integrated Brier Score (IBS) 0.057) and outperformed benchmarks (AUC ≤ 0.774). Leave-one-province-out testing showed median AUC at 4 years 0.832 with stable error. Higher CES-D and lower cognition were associated with higher discrete-time mortality hazard; estimates among late-stage cases were directionally similar but imprecise. CONCLUSIONS: Depressive symptoms and cognition may provide urology-relevant prognostic signal for near-term mortality among men with late-stage PCa. An attention-based discrete-time survival model demonstrated good discrimination and calibration in CHARLS; external validation in clinically staged advanced PCa cohorts is needed before clinical use.