A hybrid RF–XGBoost feature selection approach for multiclass thyroid disease prediction
Thyroid disorders are among the most common endocrine diseases worldwide, and their early and accurate diagnosis is important for preventing serious health complications. Despite the widespread use of machine learning techniques for thyroid disease prediction, comprehensive comparisons of knowledge-driven, statistical, model-based, and ensemble feature-selection strategies within a consistent experimental framework remain limited. A unified framework was developed for five-class thyroid disease classification to systematically compare five feature-selection approaches and one dimensionality-reduction approach, including Analytic Hierarchy Process (AHP), K-Nearest Neighbor (KNN)-based feature selection, Gradient Boosting (GB), Boruta, a hybrid RF–XGBoost feature-selection approach, and Principal Component Analysis (PCA). The proposed RF–XGBoost method uses importance scores from the Random Forest and XGBoost algorithms to build a reduced feature set. The algorithm was tested on a preprocessed subset of the publicly available UCI thyroid disease dataset. The AHP-based feature ranking incorporated expert knowledge obtained through consultation with an endocrinologist, enabling comparison between knowledge-driven and data-driven approaches. Four machine learning classifiers, including Random Forest (RF), Logistic Regression (LR), KNN, and XGBoost, were assessed based on their accuracy, precision, recall, F1-score, execution time, confusion matrix, ROC–AUC analysis, and 10-fold cross-validation. The feature selection method had a substantial impact on the classifier’s performance. The Boruta implementation in combination with the XGBoost algorithm provided the best results which amounted to the test accuracy of 0.9887. In turn, the proposed RF–XGBoost feature selection in combination with the XGBoost classifier reached a value of 0.9842. For the latter method, a 10-fold cross-validation resulted in an average accuracy of 0.9808 ± 0.0063 and macro-F1, precision, and recall of 0.9810 ± 0.0060, 0.9819 ± 0.0058, and 0.9808 ± 0.0064, respectively. The application of PCA for dimensionality reduction led to a substantially worse performance compared to the application of the feature selection methods identified above. In addition, the permutation-importance and SHAP analyses were conducted to evaluate the stability and interpretability of the feature importance rankings. Statistical comparisons showed that the performance of the proposed RF–XGBoost configuration was not significantly different from the results of the permutation-importance- and SHAP-based configurations. The proposed RF–XGBoost approach offers competitive and concise ensemble feature-ranking strategy for a five-class thyroid disease classification task while also demonstrating the importance of assessing the feature-selection procedures in a unified and leakage-free experimental setup.
Authors
- Mostafa Amiri (ORCID: https://orcid.org/0000-0003-3775-0393)
- Raheleh Ghouchan Nezhad Noor Nia (ORCID: https://orcid.org/0000-0001-5697-9041)
- Tahereh Saeidi Nia (ORCID: https://orcid.org/0009-0000-8804-294X)
- Zahra Saeidi Nia (ORCID: https://orcid.org/0009-0008-8211-1594)
Institutions
- Mashhad University of Medical Sciences (IR)
- Islamic Azad University, Mashhad (IR)
- Semnan University (IR)
Publication Details
- Journal
- BMC Medical Informatics and Decision Making
- Published
- 2026-11-14
- DOI
- https://doi.org/10.1186/s12911-026-03977-1
- Primary Topic
- Artificial Intelligence in Healthcare
- Type
- article
- Field-Weighted Citation Impact
- 0.00