Title:Diffuse to fuse EEG spectra -- intrinsic geometry of sleep dynamics for classification

Abstract: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.