GMDH-based feature ranking and selection for improved classification of medical data

Area of Science:

• Computational biology
• Machine learning in medicine
• Data mining for healthcare

Background:

• Medical datasets often have numerous disease markers but limited records.
• High dimensionality can hinder classifier performance in medical diagnosis.

Purpose of the Study:

• To develop and evaluate a novel feature ranking and selection method for medical diagnosis.
• To improve the performance and implementation of diagnostic classifiers by reducing data dimensionality.

Main Methods:

• Utilized a Group Method of Data Handling (GMDH) based abductive network training algorithm for feature ranking.
• Employed a progressive feature inclusion strategy to identify optimal subsets and prevent overfitting.
• Applied Receiver Operating Characteristics (ROC) analysis for performance evaluation.

Main Results:

• Achieved significant dimensionality reduction: 22% for breast cancer and 54% for heart disease data.
• Demonstrated improved overall classification performance without significant loss in the area under the ROC curve.
• GMDH-based feature selection proved effective with neural network classifiers.

Conclusions:

• Complete feature ranking and selection using GMDH offers substantial data dimensionality reduction in medical applications.
• This approach enhances classifier performance and implementation efficiency for medical diagnosis.
• The method is robust and effective, even with complex datasets and neural network models.