Deep Models for Detecting Sleep-Related Problems
Open Access
- Author:
- Huang, Guanjie
- Graduate Program:
- Information Sciences and Technology
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- August 20, 2021
- Committee Members:
- Sharon Huang, Major Field Member
Fenglong Ma, Chair & Dissertation Advisor
Sencun Zhu, Outside Unit & Field Member
Ting Wang, Major Field Member
Mary Beth Rosson, Program Head/Chair - Keywords:
- deep learning
sleep stage classsification
sleep apnea detection
cross attention
evidential learning - Abstract:
- A large number of people suffer from different types of sleep-related problems, such as insomnia, narcolepsy, and apnea. One standard approach to detect the problems is to use polysomnography (PSG). However, it is difficult for medical practitioners to manually inspect the PSG data as the process is complicated and time-consuming. Traditional machine learning methods have been successfully applied to automate the diagnosis process by extracting hand-crafted features, which mainly depend on domain expert knowledge. For the physiological signal, its morphologies are very miscellaneous, which makes it even harder to design the features. Thus, the challenge is how to thoroughly learn the representative features of physiological data with miscellaneous morphologies. Deep learning models have shown superiority for automated diagnosis tasks compared with traditional machine learning ones. However, due to the complex and limited number of clinical data, it is hard to totally unleash those deep models' power. The hand-crafted expert knowledge-based features are still insightful, which increases the model's generalization and reminds the model of data characteristics. Therefore, incorporating hand-crafted features into deep learning models may boost the detection performance, but the challenge is how to combine them together effectively. Most studies only use one signal modality. However, multimodal PSU signals are more widely used to track the condition of the human body in a clinical setup. Thus, how to appropriately leverage the information from different modalities is challenging. Also, existing deep learning methods for sleep-related diagnosis only predict the probability of the predicted classes without telling how confident/reliable the prediction is. Thus, estimating the reliability/trustability of the predictions by the deep learning models should be an important and essential challenge in clinically relevant tasks. Besides the PSG, the accelerometer is also a widely used sensor to track human body activities. It is less intrusive and more scalable. Since it is naturally multivariate, there are some potential relationships between each variate. Hence, the challenge is how to learn the features from different multivariate collaboratively. This dissertation proposes several deep learning models for tackling the aforementioned challenges of sleep-related problems. Specifically, the SleepStageNet model is proposed to extract thorough features from a single EEG signal by using different feature maps parallelly in the designed CNN. The ConCAD framework is proposed aiming at fusing the deep features and the expert-curated features by contrastive learning-based cross attention mechanism to boost the performance further. We develop a TrustSleepNet model with a cross-modality attention to predict the problem with an uncertainty score using multimodal PSG data, which makes the prediction more accurate and trustworthy. Lastly, instead of using standard PSG signal, we propose an AccSleepNet system to fuse the multivariate data synergistically from three axes of a wrist-worn accelerometer by a novel axis-aware hybrid fusion module. These models are validated on two main sleep-related applications: sleep stage classification, which is a preliminary and essential step to diagnose other sleep disorders; and sleep apnea detection, which is one of the most common sleep problems. With the proposed models, the diagnosis procedure is simplified and automated. It can promote clinical experts to diagnose the problems more accurately and faster. Also, it helps the patients monitor and understand their health condition easier and better.