The research team at Nox Medical is presenting a new abstract at this year’s World Sleep Congress in Rome. The abstract, titled ResTNet-Arousals: A end-to-end deep learning approach to arousal detection, presents an end-to-end deep learning approach to robustly identify arousals from standard polysomnogram recordings (PSG) and from Self-Applied Somnography (SAS) studies. The SAS setup allows patients to self-administer frontal EEG and EOG leads in a home sleep study, which reduces cost and is more convenient for the patients. Nox SAS is currently available for research purposes only and was recently covered in Nox Medical’s live webinar with Naresh Punjabi, MD, Ph.D. During the webinar, Dr. Punjabi explained how he and his team utilized the Nox SAS in a multicenter AIDS cohort study (MACS). You can find a recap of the session here.
At World Sleep, the abstract will be presented by Sigurður Ægir Jónsson during the Oral Abstract Session in room 33:
Oral Abstract Session at World Sleep
Wednesday, March 15, 10:45 – 12:30
Room 33
Below you can read a short summary of the abstract. Please find the session link here
ResTNet-Arousals: A end-to-end deep learning approach to arousal detection.
Sigurður Ægir Jónsson, Eysteinn Finnsson, Dagmar Lukka Loftsdóttir, Eydís Arnardóttir, Jón Skírnir Ágústsson
Nox Research, Nox Medical ehf, Reykjavík, Iceland
Arousals are defined as abrupt shifts of electroencephalography (EEG) frequency that last at least 3 seconds, preceded with at least 10 seconds of stable sleep. The identification of arousals is important for the evaluation of sleep continuity and diagnosis of sleep disorders. Arousals are difficult for human experts to score and the low inter scorer agreement makes this a particularly challenging task for artificial intelligence (AI) models to learn. However, a well-designed AI model might be helpful in improving scoring consistency, leading to more consistent clinical results. The ResTNet-Arousals model structure was inspired by ResNet convolutional neural network architecture, which has been highly successful in image recognition tasks. The model makes predictions from the raw EEG, EOG and EMG signals, in an end-to-end fashion to avoid manual feature extraction. The model output is a temporal sequence of arousal probabilities, which are then used to generate discrete arousal events in a post-processing step. Furthermore, the output probabilities are calibrated to correct for the poor calibration of modern neural networks.
Topic: Events
