# Explore intrinsic geometry of sleep dynamics and predict sleep stage by unsupervised learning techniques

## TL;DR

This paper introduces an unsupervised method using harmonic analysis and diffusion algorithms to explore sleep dynamics and automatically predict sleep stages from EEG data, achieving accuracy comparable to supervised methods.

## Contribution

The study presents a novel unsupervised framework combining diffusion-based algorithms and nonlinear spectral analysis for sleep stage prediction, avoiding reliance on labeled data.

## Key findings

- Achieved 82.57% accuracy on Sleep-EDF SC* dataset.
- Achieved 77.01% accuracy on Sleep-EDF ST* dataset.
- Performance is comparable to supervised learning methods.

## Abstract

We propose a novel unsupervised approach for sleep dynamics exploration and automatic annotation by combining modern harmonic analysis tools. Specifically, we apply diffusion-based algorithms, diffusion map (DM) and alternating diffusion (AD) algorithms, to reconstruct the intrinsic geometry of sleep dynamics by reorganizing the spectral information of an electroencephalogram (EEG) extracted from a nonlinear-type time frequency analysis tool, the synchrosqueezing transform (SST). The visualization is achieved by the nonlinear dimension reduction properties of DM and AD. Moreover, the reconstructed nonlinear geometric structure of the sleep dynamics allows us to achieve the automatic annotation purpose. The hidden Markov model is trained to predict the sleep stage. The prediction performance is validated on a publicly available benchmark database, Physionet Sleep-EDF [extended] SC* and ST*, with the leave-one-subject-out cross validation. The overall accuracy and macro F1 achieve 82:57% and 76% in Sleep-EDF SC* and 77.01% and 71:53% in Sleep-EDF ST*, which is compatible with the state-of-the-art results by supervised learning-based algorithms. The results suggest the potential of the proposed algorithm for clinical applications.