Publicação
Using deep learning for unobtrusive sleep stage classification
| Resumo: | Sleep represents a fundamental role to our well-being and today, as sleep disorders become more and more common, there is a growing necessity to monitor our sleep quality daily. Unobtrusive automatic sleep stage classification has made a tremendous breakthrough in this subject allowing regular users to monitor their sleep with day-to-day wearables, such as Fitbit Charge 2 tracker, contrary to the traditional manual sleep scoring based on polysomnography (PSG). Using cardiorespiratory signals to sleep stage has attracted increased attention as these signals can be obtained through unobtrusive techniques and have potential for continuous daily application. Therefore, in this thesis, deep learning frameworks based on Long-short-memory networks (LSTMs) and Convolutional Neural Networks (CNNs) are used to sleep stage classify, either just using respiratory effort signals, for example obtained from respiratory inductance plethysmography (RIP), or using the combination of respiratory and cardiac features, often based on heart rate variability (HRV) calculated from electrocardiogram (ECG). The dataset used was the SIESTA dataset that contains a total of 294 subjects (588 PSG recordings) of which 197 are healthy subjects, 51 suffer from obstructive sleep apnea syndrome (OSA), and the remaining from a variety of sleep or sleep related disorders. The classification problem was divided in a three-class and four-class sleep stage classification problem. As for the results, it was obtained with respiratory data for three stages classification (Wake, rapid eye-movement (REM) and non-REM stages) a Cohen’s kappa () of 0.46 for the overall pool of subjects (All), 0.50 for healthy subjects and 0.34 for OSA subjects. For four stages classification (Wake, REM, light sleep (N1/N2) and deep sleep (N3/N4) stages) it was obtained a Cohen’s Kappa () of 0.40 for the subject pool containing all subjects (All), 0.44 for healthy subjects and 0.31 for OSA. With cardiorespiratory data, for four stages classification, it was obtained a of 0.40 for the overall subject pool (All), 0.44 for healthy subjects and 0.30 for OSA subjects. With three stages, a of 0.46 for All subjects, 0.51 for healthy and 0.32 for OSA subjects. These results demonstrate that, with the developed frameworks, it is possible to achieve fairly good results as they are similar, in some cases moderately higher, to the current state-of-the-art but fail to generalize well, as significant differences can be found between subject types (All, Healthy and OSA). |
|---|---|
| Autores principais: | Prata, Marco André Ramos Dias |
| Assunto: | Engenharia e Tecnologia::Engenharia Médica |
| Ano: | 2018 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Sleep represents a fundamental role to our well-being and today, as sleep disorders become more and more common, there is a growing necessity to monitor our sleep quality daily. Unobtrusive automatic sleep stage classification has made a tremendous breakthrough in this subject allowing regular users to monitor their sleep with day-to-day wearables, such as Fitbit Charge 2 tracker, contrary to the traditional manual sleep scoring based on polysomnography (PSG). Using cardiorespiratory signals to sleep stage has attracted increased attention as these signals can be obtained through unobtrusive techniques and have potential for continuous daily application. Therefore, in this thesis, deep learning frameworks based on Long-short-memory networks (LSTMs) and Convolutional Neural Networks (CNNs) are used to sleep stage classify, either just using respiratory effort signals, for example obtained from respiratory inductance plethysmography (RIP), or using the combination of respiratory and cardiac features, often based on heart rate variability (HRV) calculated from electrocardiogram (ECG). The dataset used was the SIESTA dataset that contains a total of 294 subjects (588 PSG recordings) of which 197 are healthy subjects, 51 suffer from obstructive sleep apnea syndrome (OSA), and the remaining from a variety of sleep or sleep related disorders. The classification problem was divided in a three-class and four-class sleep stage classification problem. As for the results, it was obtained with respiratory data for three stages classification (Wake, rapid eye-movement (REM) and non-REM stages) a Cohen’s kappa () of 0.46 for the overall pool of subjects (All), 0.50 for healthy subjects and 0.34 for OSA subjects. For four stages classification (Wake, REM, light sleep (N1/N2) and deep sleep (N3/N4) stages) it was obtained a Cohen’s Kappa () of 0.40 for the subject pool containing all subjects (All), 0.44 for healthy subjects and 0.31 for OSA. With cardiorespiratory data, for four stages classification, it was obtained a of 0.40 for the overall subject pool (All), 0.44 for healthy subjects and 0.30 for OSA subjects. With three stages, a of 0.46 for All subjects, 0.51 for healthy and 0.32 for OSA subjects. These results demonstrate that, with the developed frameworks, it is possible to achieve fairly good results as they are similar, in some cases moderately higher, to the current state-of-the-art but fail to generalize well, as significant differences can be found between subject types (All, Healthy and OSA). |
|---|