Feature ranking and rank aggregation for automatic sleep stage classification: a comparative study.

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Machine Learning

Background:

• Sleep quality assessment relies heavily on polysomnography (PSG) for sleep stage identification.
• Manual sleep stage classification is labor-intensive, subjective, and expensive, necessitating automated methods.
• Feature extraction and selection are critical steps influencing the accuracy of automatic sleep stage classification algorithms.

Purpose of the Study:

• To evaluate and compare the performance of seven distinct feature selection methods and two feature rank aggregation techniques.
• To identify optimal feature extraction and selection strategies for enhancing automatic sleep stage classification accuracy.
• To assess feature selection methods based on classification accuracy, ranking stability, and similarity.

Main Methods:

• Utilized Pz-Oz EEG, horizontal EOG, and submental chin EMG recordings from 22 healthy participants.
• Extracted a comprehensive set of 49 features encompassing temporal, spectral, entropy-based, and nonlinear categories.
• Compared seven feature selection methods and two rank aggregation methods using established evaluation criteria.

Main Results:

• The Minimum Redundancy Maximum Relevance with Mutual Information Difference (MRMR-MID) method yielded the highest classification performance.
• The Fisher method provided the most stable feature ranking among the evaluated conventional methods.
• Feature rank aggregation methods (Borda, RRA) showed average performance, not significantly outperforming conventional techniques.

Conclusions:

• MRMR-MID and Fisher represent effective strategies for feature selection in sleep stage classification.
• Conventional feature ranking methods can be sufficient, with the choice depending on user-specific accuracy and computational needs.
• Further research may explore advanced aggregation techniques or hybrid approaches for improved performance.