Robust, automated sleep scoring by a compact neural network with distributional shift correction
Accurately determining the sleep stage of experimental subjects is a key step in sleep research. Despite years of research into automated methods for scoring rodent sleep recordings, most scoring is still performed manually. Here, I present an automated, machine learning-based sleep scoring method that avoids the subjective and labor-intensive task of manual scoring. In the first chapter, I review recent advances in the field of sleep scoring. New algorithms have, over time, extracted more and more useful information from underlying physiological signals used as inputs. However, inter-laboratory and inter-subject differences have thus far prevented any single automated method from being widely applicable.
In the second chapter, I present a feature scaling algorithm, mixture z-scoring, that can eliminate many of these differences. Importantly, it also preserves changes in the amount of time a given subject spends in each sleep stage, which is not attainable using existing algorithms. I then present a neural network architecture which efficiently learns to score sleep from spectrograms of electroencephalogram recordings and evaluate it using a large, high-quality dataset. When mixture z-scoring is used as a preprocessing step, the network achieves state-of-the-art performance. I also introduce a free, open-source software package that allows even novice users to make use of the network and mixture z-scoring. This work is presented in the form of a published, first-author manuscript.
In the final chapter, I discuss the limitations of the scoring algorithm and its potential application for scoring data from other species. I also examine some remaining challenges in the field of sleep scoring as well as their possible solutions. As a whole, this work provides
computational tools that are designed to meet the data processing needs of the sleep research community.