Automated Rodent Sleep Analysis with Modern Machine Learning Methods

Sammanfattning: Sleep staging is the use of electrophysiological signals to determine the quality and architecture of sleep in an animal. Currently, to achieve this, researchers manually classify contiguous sections of electroencephalographic and electromyographic signals into defined sleep modes or stages. This is a slow and laborious process. Many attempts at semiautomated solutions can be found in the literature. In these attempts, a researcher manually classifies a portion of the data from a specific rodent. This data is used to train a model which can then be used to classify the rest of the data from that rodent. While such solutions can be found in commercial products, they still require hours of manual classification to be done by the researcher. In this thesis, I explore two machine learning methods in an attempt to fully automate the process of sleep staging. The automation consists of building a classifier that can classify data from a new rodent, using only manually classified data from previous rodents. This classifier should classify this new rodent’s data at a sufficiently high accuracy. While there have also been attempts at such a system in the past, none of them have reached a level of accuracy that is acceptable for use. The two methods implemented in this thesis are support vector machines (SVM) and convolutional neural networks (CNN). The results obtained are promising, with the results from SVM being on the cusp of real world usability for automated sleep staging.