A Hybrid Machine Learning Approach for Thyroid Cancer Recurrence Prediction: Stacked Ensembles, Voting Models, and Feature Reduction

Thyroid cancer marks the most frequent endocrine cancer while having high recurrence potential, necessitating accurate future disease predictions. The proposed research develops a combined machine learning system that integrates stacked ensembles and ensemble voting models alongside feature reduction strategies to improve thyroid cancer recurrence predictions. A clinical and demographic dataset was analyzed using multiple ML models, including XGBoost, LightGBM, CatBoost, Random Forest, Support Vector Machines (SVM), and Logistic Regression. The models were evaluated based on key metrics such as accuracy, AUC-ROC, precision, recall, and F1-score. Logistic Regression showed 88.77% predictive ability, while LightGBM achieved 88.25% according to individual model results. The ensemble voting model, which aggregates multiple classifiers' probabilistic outputs, demonstrated improved predictive stability, while the stacked ensemble model surpassed all other individual predictors, attaining the highest accuracy (90.91%) and an AUC-ROC value of 97.27%. To improve model interpretability, SHapley Additive exPlanations (SHAP) was employed, identifying cancer stage, risk classification, and patient age as the most influential predictors. The findings demonstrate that ensemble learning techniques, including stacked and voting-based models, not only enhance predictive accuracy but also improve model robustness and reliability. The research integrates explainable AI components to connect sophisticated black-box ML models with clinical treatment decisions, thus enabling healthcare providers to make individualized patient-risk assessments. The proposed framework based on AI functions as the groundwork to create clinical decision support systems that improve prognosis evaluation and individualized treatment development for thyroid cancer patients.