Interpretable machine learning for presurgical differentiation of Hürthle cell carcinoma and adenoma: a SHAP-augmented approach.

Background: The preoperative differentiation between benign Hürthle cell adenoma (HCA) and malignant Hürthle cell carcinoma (HCC) remains clinically challenging. This study aimed to develop an interpretable machine learning framework to improve diagnostic accuracy and assist clinical decision-making. Methods: We retrospectively enrolled 554 patients (280 HCA, 274 HCC) from a single center. Fifteen clinical, serological, and ultrasonographic variables were incorporated. Following rigorous feature selection via LASSO and Random Forest algorithms, four advanced machine learning models and logistic regression (LR) were trained and evaluated using 10-fold cross-validation. The SHapley Additive exPlanations (SHAP) and Individual Conditional Expectation (ICE) trajectories were utilized to interpret the optimal model globally and locally. Results: The XGBoost model demonstrated superior discriminative performance, achieving an Area Under the Curve (AUC) of 0.911, significantly outperforming LR (AUC = 0.746). Decision curve analysis confirmed its higher clinical net benefit. SHAP analysis demystified the algorithmic "black box," identifying vascularity grade, absent halo sign, and elevated serum thyroglobulin as top malignant predictors. Furthermore, patient-specific ICE trajectories successfully simulated counterfactual clinical reasoning for misclassified cases. Conclusion: The SHAP-augmented XGBoost framework provides highly accurate, transparent, and personalized presurgical risk stratification for Hürthle cell neoplasms, potentially reducing unnecessary diagnostic thyroidectomies.