Designing an automated sleep scoring system.
(Work in progress)
Checkout the code here
Sleep occupies about one-third of human life which makes it necessary to understand its needs and functions [1]. The most important function, among others, is the consolidation of memory and the weakening of synaptic connections. Sleep occurs in repeated cycles of the three most important stages, Wake, REM, and Non-REM. All three stages occur with the brain activity of varying frequency and amplitude. While Wake and REM stages are dominated by high-frequency and low amplitude waves, NREM sleep is mostly dominated by low-frequency, high amplitude waves also called slow waves.
The current standard method for sleep classification is by manual visual inspection, which is prone to several errors. Moreover, we are only limited to three standard stages, REM, NREM, and WAKE. Although recent studies have provided evidence that sleep can be subclassified into many more substages that are hidden from the scorer in visual scoring.
Here we have demonstrated an unsupervised method for sleep scoring based on the features extracted from the LFP (Local Field Potential) data recorded from the hippocampus and thalamus or neocortex of the mice’ brain. In this method, we need not have prior knowledge on the number of sleep stages we want to classify our data in. This method can help us detect sleep stages that would have been impossible to detect during manual scoring.
Using the unsupervised mcRBM (mean covariance Restricted Boltzmann Machine) model [4] we can find the latent states in the raw sleep dataset which can define the hidden properties in the data. Each epoch in the raw dataset can then be associated with one of the latent states. The mcRBM method has been proven very effective in finding the latent sleep stages as evidenced by this study [3] using EEG and EMG data.
Initially we started with the power values in various frequency ranges and the EMG-like signal or accelerometer signal. We took delta, beta, and theta power values and ultimately took the ratio of each pair as done in the original study.
The model we used for estimating the distinct sleep stages is the mean-covariance Restricted Boltzmann Machine (mcRBM). It is a stochastic, unsupervised method to identify all the substages of sleep that were hidden to standard analysis. This method aids in inferring a set of latent states or representations describing different modes in the sleep cycle and reoccur a number of times in the entire duration. These latent states form clusters based on their transition probabilities, each of which can be represented as a sleep stage.
Being an unsupervised approach we need not give the number of sleep stages we want to classify our data in. This helps in identifying new sleep stages which would be highly conducive in sleep research.