Abstract:
Thyroid disorders are among the most common hormonal dysfunctions across the globe, involving metabolic and hormone activity. Early and precise diagnosis is of great significance for optimal treatment and prevention of severe complications. Most of the conventional diagnostic methods depend on visual interpretation of laboratory results and therefore can be delayed or inaccurate. For overcoming these challenges, we have introduced a hybrid machine learning model ThyroidNet-RF that incorporates an interpretable Random Forest classifier and deep learning-based feature extraction for accurate and reliable thyroid disease diagnosis. It is based on a systematic series of steps comprising data cleaning, preprocessing, correlation analysis and feature selection with the Boruta algorithm. The class imbalance is addressed with the SMOTE-Tomek method during the balancing of the datasets. Several machine learning models including DT, KNN, SVM, ANN and XGBoost were developed as well as compared with ThyroidNet-RF model. The model showed highly competitive classification performance with an accuracy of 99.87%, precision of 1.00, recall of 0.98, F1-score of 0.99 and AUC of 1.00 when compared to other classifiers Results show that ThyroidNet-RF is able to improve the classification accuracy, reliability and interpretability efficiently, which makes it a useful decision-support tool for the clinicians. Results This research demonstrates the potential for assembling with deep learning in medical diagnosis and improved opportunities for early detection and individualized treatment decisions to enhance patient outcomes
