Classifier fusion approaches for diagnostic cancer models

Area of Science:

• Machine Learning
• Medical Data Analysis
• Ensemble Methods

Background:

• Classifier ensembles enhance predictive accuracy and feature space coverage.
• Recent trends favor statistically sound integration schemes over heuristic combinations for quantifiable results.
• Ensemble methods are crucial for improving generalization capability and understanding error bounds.

Purpose of the Study:

• To evaluate an ensemble of 8 classifiers using 15 fusion strategies on two medical datasets.
• To analyze base classifier correlations using 11 established metrics.
• To provide a foundation for selecting an improved hyper-classifier.

Main Methods:

• Utilized an ensemble of 8 base classifiers.
• Applied 15 distinct data fusion strategies.
• Assessed classifier correlation with 11 common metrics.
• Tested performance on two distinct medical problems.

Main Results:

• Identified varying performance across 15 fusion strategies.
• Quantified base classifier correlations, revealing dependencies and diversities.
• Established criteria for selecting a superior hyper-classifier based on ensemble behavior.

Conclusions:

• The choice of fusion strategy significantly impacts ensemble performance in medical applications.
• Understanding base classifier correlation is key to designing effective ensemble models.
• This research provides a framework for developing robust, high-performance hyper-classifiers for medical data.