Power Spectral Density Based Sleep Scoring Using Artificial Neural Networks

Abstract: Sleep scoring is the process that medical researchers use to analyze the sleep of a subject. By looking at signals in the brain and muscles, it is possible to determine the current sleep state of the subject. The procedure is traditionally done manually, requiring a lot of tedious processing of data. In this report, a machine learning system that automates the process of sleep scoring is studied and developed. The system works by estimating the power spectral density of the electroencephalography (EEG) and electromyography (EMG) signals, and training an artificial neural network to classify the correct sleep state. The signal processing was done in Python and the artificial neural network was implemented in Keras, using a TensorFlow back end. Finally, the implemented system proved to have an accuracy comparable to that of manual sleep scoring on five different rat datasets. Additionally, the system was able to generalize beyond the rat specimens it was trained on, meaning it could potentially be used on specimens that lack labeled sample data.