Joint ensemble empirical mode decomposition and tunable Q factor wavelet transform based sleep stage classifications

Abstract

Since the medical professionals require to segment the electroencephalograms (EEGs) into the pieces, study each piece of the EEGs and perform the manual annotation on these pieces, the sleeping classification requires a huge workload from the medical professionals. To reduce the workload, this paper investigates the automatic classification of the sleep stages so that the diagnosis of the sleep disorders can be speeded up. In particular, this paper proposes a joint ensemble empirical mode decomposition (EEMD) and tunable Q factor wavelet transform (TQWT) based method for performing the sleep stage classifications. Here, only the single channel EEGs are employed. For a given EEG, first the EEMD is employed to decompose the EEG into the intrinsic mode functions (IMFs). Then, the first two IMFs are selected. Next, the original EEG and these two selected IMFs are further decomposed using the TQWT. Second, the energies of the obtained subbands are computed. The four largest energy subbands are retained. Third, five statistical features are extracted from each selected wavelet component. Finally, the bagging classifier is employed for performing the classification. Here, the total number of the trees inside the classifier is determined via a statistical based criterion such as the out of bag (OOB) error. To demonstrate the effectiveness of our proposed method, the EEGs are taken from both the Sleep EDF database and the Sleep EDF expanded database. It is found that our proposed method can yield the classification accuracies at 90.11%, 91.29%, 94.71%, 94.72% and 98.16% for the 6 sleep stage classification, the 5 sleep stage classification, the 4 sleep stage classification, the 3 sleep stage classification and the 2 sleep stage classification, respectively, based on the sleep EDF database as well as 88.14%, 89.37%, 90.66%, 92.89% and 97.37% for the 6 sleep stage classification, the 5 sleep stage classification, the 4 sleep stage classification, the 3 sleep stage classification and the 2 sleep stage classification, respectively, based on the sleep EDF expanded database. The obtained computer numerical simulation results show that our proposed method can achieve the good classification performance. In addition, as only the single channel EEGs are employed for performing the classification, the hardware implementation cost is low. Moreover, as only four subbands are selected while the rest subbands are discarded, the required computational power of our proposed method is low. Hence, the diagnosis of the sleep disorders can be significantly speeded up.