Label-Surroundings pleasant Sleep Staging Using Transformers Pre-trained with Place Prediction
Authors: Sayeri Lala, Hanlin Goh, Christopher Sandino
Abstract: Sleep staging is a clinically essential exercise for diagnosing quite a few sleep points, nevertheless stays tough to deploy at scale on account of it on account of it is every labor-intensive and time-consuming. Supervised deep learning-based approaches can automate sleep staging nevertheless on the expense of big labeled datasets, which can be unfeasible to acquire for quite a few settings, e.g., uncommon sleep points. Whereas self-supervised finding out (SSL) can mitigate this need, newest analysis on SSL for sleep staging have confirmed effectivity constructive facets saturate after teaching with labeled info from solely tens of matters, due to this fact are unable to match peak effectivity attained with greater datasets. We hypothesize that the speedy saturation stems from making use of a sub-optimal pretraining scheme that pretrains solely a portion of the construction, i.e., the perform encoder, nevertheless not the temporal encoder; subsequently, we advise adopting an construction that seamlessly {{couples}} the perform and temporal encoding and an appropriate pretraining scheme that pretrains your total model. On a sample sleep staging dataset, we uncover that the proposed scheme presents effectivity constructive facets that do not saturate with amount of labeled teaching info (e.g., 3–5% enchancment in balanced sleep staging accuracy all through low- to high-labeled info settings), reducing the amount of labeled teaching info needed for high effectivity (e.g., by 800 matters). Based totally on our findings, we advise adopting this SSL paradigm for subsequent work on SSL for sleep staging.
