Abstract 14810: ASCVD Risk Score vs Machine Learning-Based Algorithm in the Prediction of ASCVD Events in Women With Breast Cancer
Nickolas Stabellini, Roger S Blumenthal, Marcio S Bittencourt, Seamus P Whelton, Darryl Leong, Justin Moore, Jennifer Cullen, Priyanshu Nain, John Shanahan, Susan F Dent, Alberto Montero, Avirup Guha- Physiology (medical)
- Cardiology and Cardiovascular Medicine
Introduction: Cancer patients may face an elevated risk of developing atherosclerotic cardiovascular disease (ASCVD). Consequently, a conventional pooled cohort equation may not accurately predict cardiovascular (CV) outcomes in this population.
Hypothesis: A cancer-specific algorithm is superior to conventional ASCVD risk scores in women with breast cancer (BC).
Methods: Women ≥18 years, diagnosed with BC between 2005-2012 at a hybrid academic-community practice (Northeast Ohio, US) were included. A Machine Learning (ML) XGBoost algorithm (with survival modelling), developed using a training subset of the cohort (60% train + 20% test), ranked 40 covariates (including demographic, treatment-related and social determinants of health information) for ASCVD prediction using SHAP (SHapley Additive exPlanations) values. The top 10 ML predictors were transformed in a predictive equation using logistic regression models. This equation was tested in the cohort validation subset (20%), and subsequently compared to the ACC/AHA ASCVD risk score via time-dependent receiver operating characteristic curve.
Results: BC women (n=5,687, Table 1) had a median age of 60 (interquartile range 50-71) years, 17% had advanced stage disease (TNM III-IV), 44.2% received chemotherapy, 60% endocrine therapy, 29.2% radiotherapy, and 73.3% surgery, respectively. Of those, 16.7% had a 10-year ASCVD. The ACC/AHA risk score had an area under the curve (AUC)=0.76 and underestimated ASCVD in 4.9% (mean risk=21.3% [95% CI 19.9-22.6] vs. mean predicted risk=16.4% [95% CI 15.2-17.5]). The equation pooled from the top-10 predictors of the ML algorithm (C-index=0.81 [95%CI 0.80-0.82]) achieved an AUC=0.84.
Conclusion: Conventional ASCVD risk scores tend to underestimate the risk in women with BC. A cancer-specific model exhibited excellent performance and generated a user-friendly equation for predicting ASCVD risk in this population. Further external validation is needed.