Explainable Machine Learning Framework for Thyroid Cancer Recurrence Prediction

Abstract

Accurate prediction of thyroid cancer recurrence is essential for improving long-term patient management and supporting evidence-based clinical decision-making. Although machine learning has demonstrated promising predictive performance, limited model interpretability remains a major barrier to its clinical adoption. This study aims to develop an Explainable Machine Learning framework for thyroid cancer recurrence prediction by integrating Extreme Gradient Boosting (XGBoost) with SHapley Additive exPlanations (SHAP) using clinicopathological features. A publicly available dataset containing 383 patient records was preprocessed through label encoding, correlation analysis, Chi-Square-based feature selection, and Min-Max normalization. Logistic Regression, Decision Tree, Random Forest, and XGBoost were comparatively evaluated using 10-fold stratified cross-validation with Accuracy, Precision, Recall, F1-score, and ROC-AUC as evaluation metrics. The best-performing model was subsequently interpreted using global and local SHAP analyses. XGBoost achieved the highest performance, with an accuracy of 95.8% ± 4.4%, precision of 93.4% ± 8.3%, recall of 91.4% ± 9.9%, F1-score of 92.2% ± 8.3%, and ROC-AUC of 98.6% ± 2.5%, outperforming the other models. SHAP analysis identified Response, Risk, and N Stage as the most influential clinicopathological factors affecting recurrence prediction. This study contributes by developing a unified Explainable Machine Learning framework that integrates comparative model evaluation, XGBoost prediction, and global and local SHAP interpretation within a single workflow. The proposed framework provides accurate and clinically interpretable recurrence prediction, supporting trustworthy risk assessment and personalized decision-making in thyroid cancer management.