SAT-346 Identification of Demographic and Clinical Risk Factors for Breast Cancer Development in Thyroid Cancer Survivors using Machine Learning Algorithms

Abstract Disclosure: S.D. Jardines: None. C. Dominy: None. D. Lee: None. V. Midya: None. M. van Gerwen: None. Background: Thyroid cancer (TC) and breast cancer (BC) are the two most common malignancies diagnosed in females. Prior studies have reported up to a 67% risk of BC in female TC survivors compared to the general population. Certain risk factors have been identified, yet few larger population-based studies have explored this relationship. This study utilized interpretable machine learning (ML) algorithms to identify predictive features for BC diagnosis following TC in a large population database. Methods: TC patients were identified from the Surveillance, Epidemiology, and End Results (SEER) Program database, and stratified into cases (BC) and controls (NBC) based on a secondary BC diagnosis. Relevant demographic and clinical information such as age, sex, race/ethnicity, TNM stages, and histological type were included as potential predictors for BC diagnosis. We divided the dataset into training (60%) and testing (40%) sets and used forest-based Interpretable ML models (Random Forest and Extreme Gradient Boosting) to search for predictive features in the training set. The statistical significance of these features’ effects on the target variable was assessed using multivariable logistic regression in the test set. Results: Eligible patients (BC, n =1174 and NBC, n = 89,307) included females >= 20 years old, diagnosed with primary malignant TC, not preceded by any BC diagnoses, between 2010-2020. Hispanic ethnicity (OR = 0.876; CI [0.786, 0.976]) and unspecified N stage (NX) (OR = 0.88; CI [0.782, 0.991]) were identified as features significantly associated with decreased odds of BC following TC. In contrast, ages 40-54 (OR = 1.816; CI [1.523, 2.165]), 55-69 (OR = 2.562; CI [2.131, 3.08]), and >=70 (OR = 1.968; CI [1.648, 2.35]) were significantly associated with increased odds of BC. Other relevant features such as tumor size, papillary TC-follicular variant histology, and T3 stage, were identified as statistically significant by multivariable regression on the whole dataset, but not in test data subset, despite having high importance values assigned by the ML algorithms. Conclusion: Patients of > 40 years had an increased risk of developing BC after TC. Hispanic ethnicity and unspecified lymph node status, however, were associated with a decreased risk. These features may provide important prognostic information and aid providers in stratifying patients for increased surveillance strategies after TC diagnosis. Presentation: Saturday, July 12, 2025