Abstract
Recurrence prediction in well-differentiated thyroid cancer remains a clinical challenge, necessitating more accurate and interpretable predictive models. This study investigates the use of a supervised CatBoost classifier to predict recurrence in well-differentiated thyroid cancer patients, comparing its performance against other ensemble models and employing Shapley Additive Explanations (SHAP) to enhance interpretability. A dataset comprising 383 patients with diverse demographic, clinical, and pathological variables was utilized. Data preprocessing steps included handling values and encoding categorical features. The dataset was split into training and testing sets using a 70:30 ratio. Model performance was evaluated using accuracy and area under the receiver operating characteristic curve. A comparative analysis was conducted with other ensemble methods, such as Extra Trees, LightGBM, and XGBoost. SHAP analysis was employed to determine feature importance and assess model interpretability at both the global and local levels. The supervised CatBoost classifier demonstrated superior performance, achieving an accuracy of 97% and an area under the receiver operating characteristic curve of 0.99, outperforming competing models. SHAP analysis revealed that treatment response (SHAP value: 2.077), risk stratification (SHAP value: 0.859), and lymph node involvement (N) (SHAP value: 0.596) were the most influential predictors of recurrence. Local SHAP analyses provided insight into individual predictions, highlighting that misclassification often resulted from overemphasizing a single factor while overlooking other clinically relevant indicators. The supervised CatBoost classifier demonstrated high predictive performance and enhanced interpretability through SHAP analysis. These findings underscore the importance of incorporating multiple predictive factors to improve recurrence risk assessment. While the model shows promise in personalizing thyroid cancer management, further validation on larger, more diverse datasets is warranted to ensure robustness.