Pairwise diversity ranking of polychotomous features for ensemble physiological signal classifiers

Area of Science:

• Biomedical Engineering
• Machine Learning
• Pattern Recognition

Background:

• Fusion classifiers with diverse components outperform those with less diversity.
• Component diversity is crucial for designing effective fusion classifiers in clinical diagnostics and pattern recognition.
• Existing methods lack a unified approach to determine optimal component diversity.

Purpose of the Study:

• Introduce a novel pairwise diversity-based ranking strategy for selecting diverse ensemble components.
• Develop unified classifier-fusion and data-fusion systems applicable to polychotomous classifiers and datasets.
• Evaluate the performance of the proposed strategy in classifying multichannel event-related potentials.

Main Methods:

• A pairwise diversity-based ranking strategy is proposed to select diverse subsets of ensemble components.
• Classifier-fusion and data-fusion systems are formulated using the diversity-based selection strategy.
• The strategy is applied to classify multichannel event-related potentials.

Main Results:

• Classification accuracy increases with component ensemble diversity for both classifier and data fusion.
• Data fusion demonstrates superior performance compared to classifier fusion for event-related potential classification.
• The diversity-based selection strategy effectively identifies optimal data components for enhanced performance.

Conclusions:

• The proposed diversity-based ranking strategy is effective in improving fusion classifier performance.
• Data fusion, guided by diversity metrics, offers a promising approach for physiological signal classification.
• The unified strategy can be applied to various pattern recognition problems involving diverse data types.