Real-time REM sleep classification
LE3 .A278 2023
2023
Silver, Danny
Acadia University
Bachelor of Computer Science
Honours
Computer Science
We investigate and compare the ability of Convolutional Neural Networks (CNN) and Re-current Neural Networks (RNN) to classify Rapid Eye Movement (REM) sleep in real-time using data from a commercial smartwatch. iPhone and Apple Watch applications are developed that collect data from the Apple Watch sensors in real-time and store the data for model development and evaluation. Alternatively, this data could be provided to a model embedded in an iPhone app to classify the current sleep stage in real-time. Heart rate and acceleration data is collected and aggregated into 2-minute frames for input to the model. Ground truth data is also collected using interpreted polysomnography provided by a Cerebra Sleep Study system. Working with a psychology student at Acadia University we ran four sleep studies with four differentparticipants to gather data for our research. Deep learning models are trained and evaluated using cross-validation on data from asingle subject, and then from multiple subjects. We conclude that a one layer convolution model with batch normalization and max-pooling and a recurrent model with two gated recurrent unit (GRU) layers perform the best overall. We recommend the double GRU model for it’s higher true positive rate (TPR). The double GRU model achieved an AUC of 0.81, a correlation of 0.373, an accuracy of 77.3%, a TPR of 62.8%, and a TNR of 79% on the dataset from multiple subjects.
The author retains copyright in this thesis. Any substantial copying or any other actions that exceed fair dealing or other exceptions in the Copyright Act require the permission of the author.
https://scholar.acadiau.ca/islandora/object/theses:4030