Diffuse to fuse EEG spectra -- intrinsic geometry of sleep dynamics for classification
share
Background: Based on the clinical needs, we introduce a novel sleep stage visualization and prediction algorithm based on two electroencephalogram signals. New Method: The nonlinear-type time-frequency analysis and diffusion geometry are combined to extract and visualize intrinsic sleep dynamics features. The hidden Markov model is trained to predict the sleep stage. Results: The extracted features reconstruct the nonlinear geometric structure of the sleep dynamics. The prediction algorithm is validated on a publicly available benchmark database, Physionet Sleep-EDF SC^* and ST^*, and a private database of sleep apnea subjects with the leave-one-subject-out cross validation. The overall accuracy and macro F1 achieve 82.66% and 74.95% in Sleep-EDF SC^*, 76.7% and 71.7% in Sleep-EDF ST^*, and 63.2% and 49.7% in the private database. Comparison with Existing Methods: The performance is compatible or slightly better than the state-of-the-art results. Conclusion: In addition to visualizing the geometric structure of different sleep stages, the prediction result suggests its potential in practical applications.
READ FULL TEXT