Development and External Validation of a Machine Learning-Based Risk Classification Score for Prostate...

Patients with prostate cancer (PC) face elevated cardiovascular (CV) risk. Conventional CV risk scores are often poorly calibrated for PC and require laboratory tests that are not routinely available in PC clinics, limiting oncology use.

The development cohort (DC) included men with PC (age ≥18 years) diagnosed in 2010-2014. Validation cohort 1 (V1) included men age 45 years and older within 1 year of PC diagnosis or near initiation of androgen deprivation therapy. Validation cohort 2 (V2) included men with PC diagnosed in 2006-2019. Outcomes were cardiovascular disease (CVD), atherosclerotic CVD (ASCVD), and heart failure (HF). Predictors were selected with XGBoost and LASSO, modeled with Fine-Gray and penalized Cox regression, and scaled to additive point scores (low/intermediate/high risk). Performance was assessed using time-dependent AUC (TDAUC), which quantifies the model's ability to correctly rank patients by risk at each time point, and compared with conventional scores.

The DC included 1,815 patients, and V1 and V2 included 4,022 and 1,729 patients, respectively (median follow-up 10.3 years). At 10 years, cumulative incidence was 15.0% for composite CVD, 9.4% for ASCVD, and 10.1% for HF. The composite CVD score (age, race, smoking history, high-risk Gleason score, and household members) achieved a 10-year TDAUC of 0.71 (95% CI, 0.64 to 0.78) in the DC and 0.59 (95% CI, 0.52 to 0.65) in V2, with a 2-year TDAUC of 0.66 (95% CI, 0.56 to 0.75) in V1. In the DC, ASCVD and HF scores achieved 10-year TDAUCs of 0.66 (95% CI, 0.58 to 0.75) and 0.70 (95% CI, 0.62 to 0.79), respectively, and retained discrimination in validation.

To our knowledge, these are the first PC-specific CV risk scores derived from oncology-available variables, supporting pragmatic CV risk stratification in PC clinics and motivating further validation and implementation studies.