Population-based prediction of cardiovascular mortality in gynecologic cancer survivors using machine learning (RISKHEART).

Abstract

5518 Background: General-population cardiovascular disease (CVD) risk scores underperform in cancer survivors by failing to capture treatment-related cardiotoxicity. Gynecologic cancer survivors constitute a distinct phenotype defined by older age and heterogeneous therapies. RISKHEART is an oncology-specific, survivorship-oriented model for individualized cardiovascular risk stratification. Methods: Women with primary gynecologic malignancies diagnosed between 2000 and 2020 were identified from the SEER program. Predictors included demographics (age, race, marital status) and SEER-available treatment variables (surgery, radiation, and chemotherapy [yes/no]). Model performance was evaluated using AUC, and explainability was assessed via SHAP. Results: The study included 104,384 women with gynecologic cancer. CVD-specific death occurred in 4,962 patients, with cumulative incidence rates of 2.44% at 5 years and 4.75% at 10 years. Affected patients were significantly older at diagnosis than those without CVD death (mean age, 75.0 vs 62.5 years; P < .001). CVD mortality differed by marital status and race (both P < .001), with widowed patients accounting for a higher proportion of CVD deaths (37.3% vs 17.7%). Endometrial cancer accounted for most CVD deaths (58.2%), whereas ovarian cancer was underrepresented among CVDs deaths relative to its prevalence in the overall cohort (15.1% vs 27.8%). In the independent testing cohort, the RISKHEART model demonstrated good discrimination for 10-year cardiovascular mortality (AUC, 0.786). For 5-year cardiovascular mortality, discrimination was slightly higher (AUC, 0.796), with sensitivity of 0.73 and specificity of 0.74. For the 5-year cardiovascular mortality model, discrimination remained elevated across 2000-2009 (AUC, 0.805) and 2010-2020 (AUC, 0.784). Explainability analyses identified age at diagnosis as the dominant predictor of cardiovascular death, with treatment-related variables contributing more prominently to short-term risk and sociodemographic factors increasing in relative importance over longer follow-up. Conclusions: In this large population-based cohort of gynecologic cancer survivors, the RISKHEART ML model demonstrated clinically meaningful discrimination for predicting CVD-specific mortality. This approach may support individualized CVD risk stratification and inform survivorship care.