EEG-Based Sleep Quality Evaluation With Transfer Learning

Sleep quality evaluation has remarkable value in both scientific research and practical applications.Sufficient sleep is of great importance to human daily life,and the study of sleep mechanisms is an important part of brain science.An objective and effective measurement of sleep quality is quite valuable in transportation,medicine,health care,and neuroscience.For example,the tiredness of the drivers due to insufficient sleep imposes a severe threat to the public safety in the transportation industry.Aiming at exploring an efficient and convenient sleep quality evaluation method just like detecting drunkenness using alcometer,we propose a subject-independent approach with deep transfer learning to evaluate the last-night sleep quality using EEG data in this paper.We collect the EEG data of 10 subjects,who are required to sleep 4 hours,6 hours and 8 hours at night with increasing deep sleep time accordingly,and the corresponding categories of sleep quality are recognized as poor,normal and good,respectively.The conventional methods,when applied in the cross-subject scenarios,degrade dramatically due to the intrinsic cross-subject differences of EEG data and background noise variations during signal acquisition.Transfer learning methods are effective for building subject-independent classifiers with the capability of capturing the underlying common structure shared by different subjects while eliminating sleep quality unrelated noise.To reduce the intrinsic cross-subject differences of EEG data and background noise variations during signal acquisition,we adopt two classes of transfer learning methods to build subject-independent classifiers.One is to find a subspace in Reproducing Kernel Hilbert Space in which the EEG data distribution of different subjects are drawing closer when mapped into this subspace,and the other is to learn the common shared structure with the state-of-art deep autoencoder.The experimental results demonstrate that deep transfer learning model achieves the mean classification accuracy of 82.16% in comparison with the baseline SVM(65.74%)and outperforms other transfer learning methods.Our experimental results also indicate that the neural patterns of different sleep quality are discriminative and stable: the delta responses increase,the alpha responses decrease when sleep is partially deprived,and the neural patterns of 4-hour sleep and 6-hour sleep are more similar compared with 8-hour sleep.