The researchers developed a system to automatically and accurately classify sleep stages in mice using data from a single EEG (electroencephalogram) channel. This system can identify whether a mouse is awake, in non-REM (NREM) sleep, or in rapid eye movement (REM) sleep in real-time.
The researchers used 214 recordings, each 6 hours long, with sleep stages annotated by human experts. The system processes data in 10-second epochs.
The researchers created two systems: UTSN (Universal Time-Series Network) and UTSN-L (UTSN with Long Short-Term Memory).
UTSN uses a convolutional neural network (CNN) to process raw EEG data, FFT (Fast Fourier Transformation) data, and zeitgeber time (ZT), which relates to the circadian rhythm.
UTSN-L adds a Long Short-Term Memory (LSTM) network to consider past data, improving accuracy.
The systems were evaluated against conventional methods like logistic regression, random forest, and AdaBoost. Metrics used for evaluation included sensitivity, accuracy, precision, F1 score, and Matthews Correlation Coefficient (MCC). UTSN-L performed better overall, especially for detecting REM sleep.
UTSN-L (UTSN with Long Short-Term Memory) The overall accuracy was 90% For REM sleep detection, UTSN-L achieved 91% sensitivity and 98% specificity.
Read this paper and discuss key takeaways for our project here.