Improved classification of medical data using abductive network committees trained on different feature subsets

Area of Science:

• Computational intelligence
• Machine learning for medical applications
• Data mining and pattern recognition

Background:

• Traditional committee methods often split training data, which can be suboptimal for high-dimensional datasets.
• High dimensionality (many features, few examples) presents challenges for accurate classification in medical diagnosis.
• Improving classification accuracy is crucial for reliable medical diagnosis and treatment planning.

Purpose of the Study:

• To introduce a novel method for training abductive network classifier committees using feature subsets.
• To enhance classification accuracy in medical diagnosis, particularly in high-dimensional data scenarios.
• To demonstrate the effectiveness of feature subsetting over data subsetting for committee training.

Main Methods:

• A novel approach for tentative feature ranking and forming subsets of uniform predictive quality.
• Utilizing the abductive network training algorithm for optimal predictor selection.
• Grouping features into mutually exclusive subsets of approximately equal predictive power for training committee members.
• Demonstration on breast cancer, heart disease, and diabetes datasets.

Main Results:

• Three-member committees trained on distinct feature subsets reduced classification errors by up to 20% compared to the best single model.
• The proposed feature subsetting approach achieved better diversity and quality for committee performance.
• Results indicate superiority over previous methods that relied on splitting the training set.

Conclusions:

• Training abductive committee members on feature subsets of equal predictive power is effective for improving medical diagnosis classification.
• Ensemble feature subset selection using GMDH-based learning algorithms is advantageous for high-dimensional data.
• This method offers a robust strategy for enhancing diagnostic accuracy in complex medical datasets.