Abstract
Thyroid disorders represent a significant endocrine imbalance that adversely affects the body's
metabolic system. Critically, these physiological disorders often manifest as psychological
symptoms, such as depression, anxiety, and cognitive decline leading to potential misdiagnosis in
counseling and primary care settings. Traditional diagnosis is often labor-intensive, invasive, and
costly. Machine learning (ML) offers a non-invasive pathway for early, cost-effective screening to
support clinical decision-making. Method: Using the UCI thyroid illness dataset, we addressed
missing data via mean imputation and constant placeholders to preserve clinical information. We
optimized feature relevance using three distinct techniques: Hilbert-Schmidt Independence
Criterion (HSIC), Multi Spatially Uniform ReliefF (MultiSURF), and Minimum Redundancy
Maximum Relevance (mRMR). These selected features were integrated with eight optimized
classifiers, including SVM, Logistic Regression, KNN, Decision Tree, AdaBoost, Bagging,
Stacking, and Voting models. Results: Experimental results demonstrate that applying robust
feature selection significantly improved diagnostic performance. Among all combinations, mRMR
combined with the Bagging classifier achieved optimal results, reaching an accuracy of 98.40%,
precision of 97.85%, recall of 97.60%, F1-score of 97.31%, and an ROC-AUC of 99.57%.
Conclusion: The results indicate that the proposed framework significantly enhances predictive
stability and generalization. This intelligent system effectively predicts various thyroid conditions,
including primary hypothyroidism and concurrent non-thyroidal illness offering counselors and
clinicians a reliable auxiliary tool to identify physiological root causes of patient distress.