Non-negative matrix factorization and sparse representation for sleep signal classification

Real-life signals such as biomedical signals are non-stationary and random in their pattern, and cannot be characterized by any specific waveform or spectral content. Processing of these natural signals involves consideration of certain significant attributes such as their non-stationary behavior over time, scaling behavior, translation invariance. Due to their random behavior, the existing discriminative methods often fail to provide a reasonable quantification performance, thereby resulting in poor classification rates. In order to address this issue, there exists a need for defining a suitable theoretical framework for biomedical signals. We have proposed, a robust Time-Frequency Nonnegative Matrix Factorization (TF-NMF) framework that uses sparse representation for quantification of sleep signals. This scheme incorporates a novel feature extraction algorithm. For signals that are nonstationary in nature, the degree of sparsity is lower compared to the stationary signals. This results into poor classification accuracy. However our proposed approach has proven that using NMF as input to the sparse representation for classification will improve the discrimination performance. Overall, maximum cross-validation performance of 87:9% was obtained, using the leave-one-out (LOO) approach for sleep abnormality detection using EMG signals. Although the computational complexity of the proposed algorithm might be higher compared to the other similar methods, this TF-NMF based method shows great potential for quantification and localization of time varying signals.