An analysis of thyroid function diagnosis using Bayesian-type and SOM-type neural networks.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Thyroid function diagnosis is a critical classification task in healthcare.
• Multivariate analysis has been previously applied to human thyroid data.
• Neural networks offer advanced pattern recognition capabilities for medical data.

Purpose of the Study:

• To reanalyze human thyroid data using two novel neural network approaches.
• To compare the diagnostic performance of neural networks against traditional multivariate analysis.
• To evaluate the effectiveness of Self-Organizing Maps (SOM) and Bayesian Regularized Neural Networks (BRNN) for thyroid function classification.

Main Methods:

• Application of Self-Organizing Map (SOM) for patient clustering and visualization based on laboratory tests.
• Utilizing a missing value SOM variant for handling incomplete datasets.
• Employing Bayesian Regularized Neural Network (BRNN) with Automatic Relevance Determination (ARD) for disease classification.
• Direct calculation of classification rates for comparison.

Main Results:

• SOM successfully identified three distinct clusters corresponding to hyperthyroid, hypothyroid, and normal thyroid function.
• Patient position within the SOM map provided detailed diagnostic information.
• BRNN demonstrated superior prediction accuracy for thyroid disease classification compared to multivariate analysis.
• The Automatic Relevance Determination (ARD) method within BRNN was confirmed as effective.

Conclusions:

• Neural network approaches, specifically SOM and BRNN, offer powerful tools for thyroid function diagnosis.
• SOM provides intuitive visualization and clustering of patient data.
• BRNN enhances classification accuracy, with ARD optimizing feature selection for improved diagnostic performance.