Development and validation of a multifactorial risk prediction model for breast cancer patients with co-occurring...

To develop and validate a multifactorial machine learning model predicting thyroid cancer (TC) co-occurrence risk in breast cancer (BC) patients.

This single-center retrospective matched case-control study analyzed 400 BC patients (200 with co-occurring TC, 200 matched BC-only controls) diagnosed between 2012-2025. Predictors included demographic, clinical, hormonal, and tumor biological variables. After feature selection via LASSO regression to handle multicollinearity, four machine learning algorithms (logistic regression, random forest, XGBoost, SVM) were developed and optimized using Bayesian hyperparameter tuning with 5-fold cross-validation. Model performance was evaluated on a 30% independent test set using AUC-ROC, calibration curves, and decision curve analysis.

Multivariate analysis identified independent risk factors for TC co-occurrence: radiotherapy history (aOR = 3.42, 95% CI: 2.14-5.46), elevated TSH level (aOR = 2.01 per µIU/mL, 95% CI: 1.65-2.45), ER-positive status (aOR = 2.47, 95% CI: 1.43-4.28), family history of TC (aOR = 3.05, 95% CI: 1.55-6.00), and younger age at BC diagnosis (aOR = 1.07 per year decrease, 95% CI: 1.04-1.10). The XGBoost model demonstrated superior discrimination (test AUC = 0.874, 95% CI: 0.836-0.934) compared to other algorithms, with 86.7% accuracy, 83.3% sensitivity, and 90.0% specificity. Notably, subgroup analysis revealed enhanced predictive performance in patients with a history of radiotherapy (AUC = 0.921). Decision curve analysis confirmed clinical utility across threshold probabilities (20-80%), showing a superior net benefit for personalized risk stratification.

The XGBoost-based model integrates radiotherapy exposure, hormonal profiles, and tumor biology to stratify TC risk in BC patients. It offers a clinically applicable tool for personalized surveillance, balancing early detection with resource optimization. External validation is warranted before implementation.