Hybrid Dimensionality Reduction of Multi-sets Using Nature Inspired Algorithms and Discriminant Canonical Correlation Analysis for Automatic Sleep Stage Classification

Abstract

Automatic sleep stage classification using various signals is an important tool to assess sleep disorders and sleep quality.However, the numerous variables from multi sources lead to the classification problems.To improve the problem, this paper proposes a new hybrid dimensionality reduction by combining the feature selection based on the nature-inspired algorithms (NIAs) and multi-sets transformation technique by Canonical Correlation Analysis (CCA)/ Discriminant Canonical Correlation analysis (DCCA).The NIAs is first adopted to generate the updated population positions in each dataset, and then the CCA/DCCA is used to fuse the selected subsets.The proposed algorithm performance is demonstrated on sleep-wake detection and multi-class sleep stage classification.Furthermore, the proposed method called Dissimilarity Binary Grey Wolf Optimization -Discriminant Canonical Correlation Analysis (DisBGWO-DCCA), modified by using the differential evolution (DE) technique based on the similarity of Jaccard coefficient, provides the best classification accuracies with 97.75% of the sleep-wake detection and 95.85% of the multi-class sleep stage classification when comparing with other single dimensionality reduction approaches [1][2].Moreover, the proposed algorithm achieves the computational cost better than the conventional NIAs as shown later in experiment.Our experiment is also operated on both healthy subjects and sleep disorder patients with efficient sleep stage classification.