Comparative Analysis of Machine Learning Models for Thyroid Disease Detection: Bridging Accuracy, Interpretability, and Clinical Integration

This research evaluates thyroid condition prediction using a Kaggle dataset that contains multi-feature data sets with imbalanced classes through five supervised machine learning models, including SVM, Decision Tree, and Random Forest, which achieved the best outcomes. Among all models assessed, the Random Forest together with Decision Tree achieved the highest level of accuracy at 95% and demonstrated the minimum misclassification rate of 5%. The performance analysis showed that Logistic Regression and SVM achieved 85% accuracy, alongside KNN, which had 82% accuracy. The TSH variable functioned as the critical element for the predictive models, comprising Logistic Regression, Decision Tree, and Random Forest, according to the results of the feature importance analysis. Random Forest was valued as the main model for prediction due to its high mean cross-validation accuracy, reaching 92%. The method emphasizes analyzing distinct diagnostics from each model design by measuring feature selection and classification performance. Analyzing model behavior and pertinent features became possible through the use of radar charts, heat maps, and PR curves on analyzed data. This research presents possibilities for future work that entails employing XGBoost models alongside acceptable oversampling techniques, as well as the implementation of XAI methods to enhance clinical significance and interpretability. Real-world experience during testing multiple datasets will enhance the practical reliability of these models when utilized for diagnostic processes.