The thyroid gland plays a crucial role in regulating metabolism and overall physiological balance, and disorders of the thyroid can lead to serious health complications if not detected early. Despite the availability of biochemical tests, thyroid diseases remain widely underdiagnosed, particularly in resource-limited settings, due to inconsistent screening practices and limited diagnostic access. While machine learning methods have been explored to assist thyroid disease detection, many existing approaches emphasize predictive performance with limited interpretability, restricting clinical trust. To address this gap, this study proposes an interpretable stacking-based machine learning framework for thyroid disease detection. The framework integrates Random Forest, XGBoost, and Decision Tree classifiers through a regularized logistic-regression meta-learner, combined with multi-criteria consensus feature selection. Model transparency is ensured using SHAP and LIME, which provide both global and patient-level explanations aligned with endocrine physiology. All preprocessing steps are performed within stratified cross-validation to prevent data leakage.The proposed framework achieved an accuracy of 99.38% on a publicly available thyroid dataset, while maintaining clinically meaningful interpretability. The model is intended as a decision-support tool for early thyroid disease screening rather than a replacement for clinical diagnosis.