Abstract Body (Do not enter title and authors here): Introduction: Conventional cardiovascular disease (CVD) risk scores are inaccurate for prostate cancer patients (PC); the Guha-Stabellini machine learning (GS-ML) score shows promise but lacks external validation.
Hypothesis: The GS-ML score is superior to conventional CVD scores in patients with PC.
Aim(s): To conduct external validation of the GS-ML score in patients diagnosed with PC.
Methods: The validation used holdout data from Seidman Cancer Center’s CAISIS platform (internal validation cohort) and RADICAL PC1 (external validation cohort; a prospective study of men diagnosed with PC within 1 year or starting ADT within 1 month of enrollment), matching covariates with the GS-ML score. With limited follow-up, short-term CVD (2,000 days from PC diagnosis) was the outcome. Performance was assessed via area under the receiver operating characteristic curve (AUC) with assistance from Youden statistic cutoff, comparing metrics with ACC/AHA pooled cohort equations (PCE), SCORE-2, and AHA-PREVENT scores. Atherosclerotic cardiovascular disease (ASCVD) included non-fatal ischemic stroke and myocardial infarction. CVD included ASCVD and heart failure.
Results: We included 2,495 patients from RADICAL PC1 and 1,506 from the internal validation cohort (Table 1). In the internal validation cohort, the CVD AUCs were 0.58, 0.49, 0.65, 0,73, and 0.75 for PCE, SCORE2, PREVENT simple, PREVENT enhanced, and GS-ML score, respectively. In RADICAL PC1, the AUCs for CVD were 0.60, 0.43, 0.67, and 0.64 for PCE, SCORE2, PREVENT simple and GS-ML score, respectively. For ASCVD, internal validation cohorts’ AUCs were 0.63, 0.62, 0.66, 0.73, and 0.67 for PCE, SCORE2, PREVENT simple, PREVENT enhanced, and Guha-Stabellini, respectively. The RADICAL PC1 ASCVD AUCs were 0.64, 0.46, 0.65, and 0.61, for PCE, SCORE2, PREVENT simple and GS-ML scores, respectively,
Conclusion(s): The GS-ML score was validated for PC patients, showing superior performance to PCE and SCORE2, and similar performance to AHA-PREVENT in predicting CVD. The European Cardio-Oncology guidelines should reconsider using SCORE2 for PC patients. Further improvement and validation with nationally representative datasets are needed to corroborate these findings and enhance the generalizability of the GS-ML score.