A Stochastic Coding Method of EEG Signals for Sleep Stage Classification

The recording of fired action potentials (spikes) of brain neurons, also known as spike trains, is considered to be the primary mode of information transmission in the nervous system. Electroencephalography (EEG) is the most direct sampling method for spike trains. However, due to the inherent biological properties of neurons such as spike randomness, timing dynamic, and noisy containment, there are challenges in EEG-related physiological identification tasks (such as sleep staging, epilepsy detection, etc.). Traditional feature engineering of EEG has a tendency toward deterministic statistical analysis and inference, which often ignores the biological properties of neurons. In this paper, we propose an innovative non-deterministic coding method of EEG signals for improving the performance of sleep stage classification tasks. By local normalization, probabilistic sampling, and window projection on the EEG signals, we discretize the continuous signals and feed them into a subsequent classification model. The coding method is tested on the public datasets and typical deep learning models for EEG. Our proposal achieved competitive sleep staging results. The precision of 0.95, 0.84, 0.92, 0.98, and 0.85 were obtained in the Wake, N1, N2, N3, and REM stages, respectively. Our research shows that the non-deterministic coding of EEG has potential for further application in biomedical devices.

Automated summary

This paper presents an innovative non-deterministic coding method for EEG signals and achieves competitive results in sleep stage classification tasks.