Thyroid Disease Prediction Using Selective Features and Machine Learning Techniques.

Area of Science:

• Endocrinology and Medical Informatics
• Computational Biology and Machine Learning

Background:

• Thyroid disease prediction is crucial, but current methods often use binary classification, small datasets, and lack validation.
• Existing research prioritizes model optimization over feature engineering, limiting diagnostic accuracy and scope.

Purpose of the Study:

• To investigate advanced feature engineering techniques for improved thyroid disease prediction.
• To develop a robust model capable of classifying multiple thyroid conditions beyond binary outcomes.

Main Methods:

• Employed feature selection methods: forward feature selection, backward feature elimination, and bidirectional feature elimination.
• Utilized machine learning-based feature selection with extra tree classifiers.
• Integrated selected features with machine learning and deep learning models, including random forest.

Main Results:

• Extra tree classifier-based feature selection, when combined with a random forest classifier, achieved 0.99 accuracy and an optimal F1 score.
• The proposed approach successfully predicted Hashimoto's thyroiditis, increased binding protein, autoimmune thyroiditis, and non-thyroidal illness syndrome (NTIS).
• K-fold cross-validation and comparison with existing studies confirmed the superior performance and efficiency of the developed method.

Conclusions:

• Machine learning models, particularly random forest with optimized features, offer a superior approach for accurate and computationally efficient thyroid disease detection.
• Feature engineering is a critical, yet underexplored, component for advancing the accuracy and scope of thyroid disease diagnostic tools.