Ensemble Feature Selection Achieves High Accuracy in Multiclass Thyroid Disease Classification

A unified machine learning framework comparing feature selection strategies for five-class thyroid disease classification demonstrated that combining Random Forest and XGBoost importance scores yields competitive predictive accuracy. In experimental evaluations on thyroid disease data, an ensemble feature-selection approach pairing Random Forest and XGBoost with an XGBoost classifier attained a test accuracy of 0.9842. The study systematically evaluated knowledge-driven, statistical, model-based, and ensemble techniques under a consistent experimental structure to determine how feature filtering affects multiclass diagnosis.

Thyroid disorders rank among the most prevalent endocrine diseases worldwide, making prompt and accurate diagnosis critical for mitigating severe health complications. While machine learning algorithms are frequently employed to predict and categorize thyroid conditions, comprehensive comparisons of different feature selection methodologies within a single standardized evaluation setup have historically remained limited. The researchers sought to address this gap by establishing an experimental baseline that directly contrasts diverse algorithmic feature reduction strategies.

To analyze diagnostic capabilities across five disease classes, the researchers established a unified framework testing five feature-selection methodologies and one dimensionality-reduction technique. The evaluated methods encompassed the Analytic Hierarchy Process, K-Nearest Neighbor-based feature selection, Gradient Boosting, the Boruta algorithm, the proposed hybrid Random Forest and XGBoost technique, and Principal Component Analysis. Testing was performed using a preprocessed subset of the publicly accessible UCI thyroid disease dataset.

The proposed hybrid strategy constructs a streamlined feature set by leveraging importance scores generated jointly by Random Forest and XGBoost models. To facilitate an explicit comparison between data-driven and clinical knowledge-driven approaches, the Analytic Hierarchy Process configuration integrated domain expertise obtained through formal consultation with an endocrinologist. The resulting feature subsets were then supplied to four separate machine learning classifiers: Random Forest, Logistic Regression, K-Nearest Neighbor, and XGBoost.

Classifier performance varied substantially depending on the feature-selection technique applied. Across the experimental configurations, the Boruta algorithm coupled with an XGBoost classifier produced the highest overall test accuracy, reaching 0.9887. The proposed hybrid Random Forest and XGBoost feature selector combined with the XGBoost classifier achieved a closely comparable test accuracy of 0.9842, demonstrating the viability of ensemble ranking strategies for multiclass clinical predictions.

During 10-fold cross-validation, the proposed hybrid feature-selection method combined with XGBoost recorded a mean accuracy of 0.9808 with a standard deviation of 0.0063. The model also demonstrated a macro-F1 score of 0.9810 ± 0.0060, a precision of 0.9819 ± 0.0058, and a recall of 0.9808 ± 0.0064 across validation folds. In contrast, applying Principal Component Analysis for dimensionality reduction yielded performance markedly inferior to the dedicated feature-selection algorithms.

Statistical evaluations indicated that the performance of the proposed hybrid Random Forest and XGBoost configuration did not differ significantly from configurations using permutation-importance or SHAP rankings. The authors concluded that the hybrid approach provides a concise and competitive ensemble feature-ranking solution for five-class thyroid disease classification while highlighting the necessity of assessing feature-selection pipelines in a unified, leakage-free environment.

Subject of Research: Machine learning feature selection for multiclass thyroid disease classification

Article Title: A hybrid RF–XGBoost feature selection approach for multiclass thyroid disease prediction

Article References: Saeidi Nia, T., Ghouchan Nezhad Noor Nia, R., Saeidi Nia, Z., & Amiri, M. (2026). A hybrid RF–XGBoost feature selection approach for multiclass thyroid disease prediction. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03977-1

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03977-1

Keywords: thyroid disease, feature selection, Random Forest, XGBoost, multiclass classification, Boruta, Analytic Hierarchy Process, Principal Component Analysis, K-Nearest Neighbor, Logistic Regression, SHAP analysis, permutation importance, UCI thyroid disease dataset, endocrine disorders, 10-fold cross-validation, ROC-AUC analysis, dimensionality reduction, ensemble feature ranking