Here are summaries of Nox Research abstracts presented at World Sleep 2019.
BodySleep: Estimating sleep stages from respiratory and body movements
We propose a method for automatically estimating sleep states (wake, sleep, rem-sleep) from an HSAT sleep study. This is achieved by extracting features from the sleep study data and using a Recurrent Neural Network (RNN) for sleep state prediction. Our method was validated against a clinical PSG dataset and shows performance that is compatible with human scored PSG. This will enable sleep clinicians to provide a more detailed diagnosis when doing at-home PG studies, including estimating patient’s sleep structure, identifying rem-related sleep apnea and improving sleep time estimation.
End-to-end machine learning on raw EEG signals for sleep stage classification (co-authored KTH University)
We propose a method for automatic sleep stage classification by feeding raw EEG data into a Convolutional Neural Network (CNN), bypassing all feature extraction. Most previous machine learning based methods have consisted of an EEG feature analysis step before machine learning based classification. Having this EEG feature analysis step before significantly slows down the procedure. Various different neural network architectures were developed and validated on more than 400 anonymized nocturnal sleep studies using the novel Self Applied Somnography (SAS) setup. Results show that models using raw EEG data have similar predictive power as methods relying on pre-calculated hand-crafted features while consuming only a fraction of the classification time. Making them a promising alternative for classifying sleep stages using the SAS setup.
Respiratory inductance plethysomnography for the reliable assessment of ventilation and sleep apnea phenotypes in the presence of oral breathing (co-authored Harvard University)
Many patients with OSA exhibit a substantial amount of oral (mouth) breathing during sleep and this oral breathing is not captured with nasal cannula airflow measurements. In the context of OSA phenotyping, an accurate measurement of breathing is paramount and thus the cannula cannot be reliably used for PSG phenotyping. On the other hand, RIP measures lung volume changes and is therefore resistant to the route of breathing. We propose new methods for careful assessment of RIP ventilation that may provide reliable sleep apnea phenotyping for the clinical setting.