Despite historical successes of population-based sleep medicine, individual differences in sleep necessitate personalized approaches. Accordingly, in its recent Strategic Vision, The NHLBI included among its Compelling Questions: “What are the major determinants of individual and sex differences in breathing patterns in sleep, susceptibility to insomnia, and other sleep behaviors?”, and among its Critical Challenges: “Robust tools and algorithms are needed to evaluate objective biomarkers of sleep health and dysfunction.” Our application directly addresses these questions and challenges.
“Deep learning” (DL) refers to a class of powerful new computational methods that use large data sets to train flexible neural network algorithms to perform tasks previously only able to be performed by human experts. Sleep medicine is poised to benefit tremendously from DL: sleep disorders are common; carefully annotated multi-modal physiological recordings are routinely acquired; numerous treatment strategies are already available to patients; and disordered sleep produces a wide variety of measurable symptoms and outcomes.
Our central hypothesis is that DL methods applied to Big Data can enable efficient discovery of clinically actionable sleep phenotypes and to advance the field of Personalized Sleep Medicine. While the heterogeneity of clinical data has historically been viewed as a limitation compared to standardized research data (e.g. the National Sleep Research Resource; NSRR), in fact it represents an advantage for training DL models and increases the external validity of discovered insights.
We will develop a Big Data platform using DL for sleep phenotyping, with three Aims: (1) Develop DL algorithms for known sleep phenotypes using subsets of PSG sensors, to overcome limitations of home testing and mobile health sleep applications. These algorithms will be validated against an external large research dataset from the NSRR. (2) Leverage DL methods to discover new phenotypes that support diagnostic and outcome prediction superior to models based on usual clinical metrics, and use these to construct physiologically-based prediction models for sleepiness (in sleep apnea) and sleep misperception (in insomnia). (3) Develop DL methods combined with at-home sleep monitoring data to predict clinical response to positive airway pressure (PAP) treatment in a prospective cohort of newly diagnosed adults with sleep apnea.
This project will benefit the field of personalized sleep medicine and, by extension, the many health problems that are linked to poor sleep, in three main ways. 1) Development of robust tools and algorithms to increase efficiency of clinical practice, and enable identification of clinically relevant sleep phenotypes; 2) Tools and data to overcome current limitations of at-home methods, based on principled analysis of large clinical gold standard datasets; and 3) Demonstrate real-world application in a clinically common use case for new physiologic biomarkers that allow quantifying and monitoring clinical response to PAP in sleep apnea patients.
Collaboration with Massachusetts General Hospital