Long-term survival prediction in older women with stage I-II breast cancer using decision tree-based machine learning

Introduction Predicting survival in older women with early breast cancer can guide personalized care and improve outcomes. Aging is an individualized process that influences tumor characteristics and survival, with cardiovascular disease (CVD) being the leading non-cancer cause of death due to cardiovascular risk factors. This study aimed to develop and validate machine learning (ML) models to predict all-cause, breast cancer-related, and CVD-related mortality in older women with stage I-II, hormone receptor-positive breast cancer at the U.S. population level. To address the heterogeneity associated with aging, we created separate models for two age groups (66–79 and ≥ 80 years). Materials and methods Using the 2006–2019 SEER–Medicare database, we identified women aged ≥66 years diagnosed with stage I–II breast cancer, representing early-stage invasive disease, who initiated adjuvant endocrine therapy (AET) between 2007 and 2009. The first date of AET use was defined as the index date. We assessed pre-existing comorbidities during the one year prior to the index date and followed patients for up to 10 years or until death. Outcomes included all-cause mortality, breast cancer-related mortality, and CVD-related mortality. We developed survival prediction models using the decision tree-based algorithms for the two age groups. Model performance was evaluated using the mean area under the receiver operating characteristic curve (AUROC), and model interpretability was enhanced using Shapley Additive Explanations. Results Among 10,104 women, all six models achieved a mean AUROC >0.7 using the random survival forest algorithm (RSF), indicating strong predictive performance. For all-cause mortality, key predictors in both age groups included age, screenings for suspected conditions (abnormal findings without diagnosis), and congestive heart failure. Tumor size, cancer stage, and secondary malignancies were most predictive of breast cancer-related mortality, while congestive heart failure, heart valve disorders, and other ill-defined heart diseases were critical for CVD-related mortality. Discussion We developed ML-based survival models across outcomes and age group using the decision tree-based algorithms to predict mortality in older women with stage I-II breast cancer. RSF demonstrated the best performance, with age, screenings for suspected conditions, and congestive heart failure consistently emerging as key predictors. Targeting these factors may enhance cardio-oncology care.