Machine learning prediction of cardiovascular events in men with prostate cancer receiving androgen deprivation therapy

Identifying prostate cancer (PCa) patients at high cardiovascular (CV) risk prior to androgen deprivation therapy (ADT) remains challenging, and no validated tool exists for major adverse CV event (MACE) risk stratification. We therefore aimed to develop and internally validate a machine learning–based model to predict long-term MACE risk in men with PCa receiving ADT. In this retrospective prognostic study, we analysed men with PCa treated with ADT, identified from the Taiwan Cancer Registry (2010–2021) linked to the National Health Insurance Research Database, with follow-up through 31 December 2022. Patients with less than three months of follow-up or MACE events during the first three months after ADT initiation were excluded. Fifteen sociodemographic, comorbidity, and ADT-related variables were evaluated. The cohort was randomly split 70/30 into derivation ( n = 27,257) and internal validation ( n = 11,681) sets. A random survival forest (RSF) model incorporating eight selected variables was developed to predict MACE (ischaemic stroke, myocardial infarction, or CV death), with non-CV death as the competing event. Over a mean follow-up of 4.26 years, 5,681 patients (14.6%) experienced a MACE. The final model identified eight predictors: age ≥ 75 years, coronary heart disease, atrial fibrillation, heart failure, prior myocardial infarction, peripheral arterial disease, hypertension, and ADT type (GnRH agonist or orchiectomy). Both the initial 15-variable and final 8-variable RSF models demonstrated satisfactory discrimination, with area under the curve (AUC) of 81.2% (95% CI: 80.5–81.9%) and 78.7% (95% CI: 78.0–79.4%), respectively. A positive correlation was observed between the number of risk factors and the incidence of MACE. We developed and internally validated an 8-variable machine learning model that stratifies long-term MACE risk in PCa patients who remain free of early CV events during the initial three months of ADT. This model may support cardiovascular risk reassessment during ADT, although prospective validation and demonstration of clinical utility remain essential before widespread implementation.