Learning sparse dictionary for long-term biomedical signal classification and clustering

Long-term observational biomedical signals are often used for many medical diagnoses including sleep disorder and epilepsy. Effective management and usage of this type of data through classification or clustering problem is the key to real-world applications. This work focuses on learning a se of selected sparse basis functions, called a double-sparse dictionary, directly from specific data, in order to produce a collection of discriminative features with low variability. Our approach is to combine wavelet transform with sparse principal component analysis (SPCA), namely wavelet sparse PCA (WSPCA), and apply it to a signal segment matrix. The application of this proposed method is demonstrated by classification and clustering problems of long-term EEG signals, and the results are compared to other PCA-based sparse methods. The nearly perfect classification accuracy (i.e., 99.7%) is obtained by using WSPCA for the data we consider. Although PCA leads to the best performance among all methods we considered, WSPCA does not lose classification accuracy significantly and it is more suitable for long-term signal classification due to the time-domain signal dimension reduction by wavelets.