Big Data Approaches for Novel Mechanistic Insights on Disorders of Sleep and Circadian Rhythms

The National Heart, Lung, and Blood Institute (NHLBI) of the National Institutes of Health (NIH) hosted a two-day virtual workshop titled Big Data Approaches for Novel Mechanistic Insights on Disorders of Sleep and Circadian Rhythms on Thursday, May 2 and Friday, May 3, 2024. The goals of this workshop were to establish a comprehensive understanding of the current state of sleep and circadian disorders research and identify opportunities to advance the field by leveraging artificial intelligence (AI), machine learning (ML) and other related technologies. Speakers from academia, including clinical practice, described innovative approaches for detection and characterization of sleep and circadian disorders as well as potential directions for future research efforts.

Agenda

View the workshop agenda.

NIH Videocast

Day 1
Day 2

Background

Sleep is an essential element of health and promotes clearance of waste from the brain, enhance immunity, reduce inflammation and aid in moderating blood pressure and cardiovascular function. Sleep disorders, sleep deficiencies and insufficient sleep occur frequently in the general population and can cause reductions in health through disruptions of these processes. Monitoring of sleep and its disorders has traditionally been dominated by the polysomnogram, a multimodal physiological dataset that comprises the electroencephalogram, electrocardiogram, electrooculogram, blood pressure and oxygen saturation, respiratory signals, and limb movements. However, the recent development of wearables, remote monitoring devices, patient-reported outcomes (PROs), and the rich array of data available from biobanks have multiplied the computational requirements for analyzing sleep data, necessitating new and more powerful approaches, including artificial intelligence (AI) and machine learning (ML). These approaches are critical for analyzing the vast amounts of sleep data currently available in a variety of data repositories to increase understanding of sleep patterns and disorders and their health consequences. AI/ML could also aid future prospective sleep studies targeting specific populations and/or sleep disorders. This workshop highlighted currently available sleep data sources, new AI/ML approaches for their analysis and innovative approaches for remote monitoring of sleep and related physiological signals.

Summary

Counting Sheep 2.0: Role of AI in Sleep Research

The workshop began with a brief history of AI and provided some modern applications for its use. AI has advanced rapidly over the last decade. The evolution of computing chips with increasing power enabled miniaturization of smart devices and facilitated collection of massive amounts of digital data, particularly in health care and research. Sleep studies can produce an abundance of raw multimodal data amenable to AI analyses, which can facilitate identification of latent patterns, linkages, and potential biomarkers within these data. AI approaches can also improve harmonization and aggregation of data prior to analyses. However, it is critically important to ensure that AI-based approaches are trained using quality data to avoid bias and inaccurate interpretations.

Key Themes:

• Data privacy and intellectual property rights must be considered when developing and training AI technologies.
• AI can improve analysis of health-related data to improve outcomes for physicians, patients, and health care systems. 
• Quality and utility of results of AI analyses depend on reliable and reproducible input data.

Resources for Sleep Data Analyses and their Applications

This session included presentations on the wide variety of currently available sleep data resources, including NHLBI-supported BioData Catalyst (BDC) and the National Sleep research Resource (NSRR), which contain both sleep-related data and tools and methodology for analyses of these data. Application of AI approaches to existing sleep data can improve data analyses, harmonization, collaboration, cohort-building, and data extraction from EHRs. Speakers encouraged the responsible use of data, interoperability of resources and metrics, and integration of PROs and other subjective measures with sleep signals and data. They also cautioned against bias, utilizing AI without considering the consequences, insufficient sample sizes, and underrepresentation of key populations within the data.

Additional presentations focused on integration of sleep-related data generated by several types of remote monitoring devices. These instruments can provide rich and diverse signals in real-world settings that can be leveraged to improve health outcomes, train AI models, and increase availability of data sets for analysis. However, there are challenges with inter-device reliability, lack of transparency among consumer device and platform algorithms, differences between consumer grade and research grade devices, a lack of standardization among signals and methods for signal gathering, and the potential to exacerbate inequities due to varying device access and validation parameters.

Key Opportunities:

• Standardization of measurements, methods, terminology, study design, and analytical approaches to facilitate interoperability, integration, and harmonization of data sources and their analyses.
• Validation of new monitoring technologies for use with diverse populations to ensure equity in and accuracy of data collection and implementation. 
• Continuous collection of sleep data and signals over several nights could provide valuable insights on night-to-night variability of sleep, resulting in a more complete picture of sleep health.
• Harmonization of sleep data resources could enable international collaborations to leverage application of AI approaches to currently available data sets.
• AI/ML, large language models (LLM), natural language processing (NLP), and related technologies can help identify hidden patterns, biomarkers, and linkages within sleep health data and signals.
• Benchmarking data collected using consumer wearables should be considered since companies will likely evolve and update the algorithms used to collect and analyze data over time.
• Clinical trials and small pilot studies provide opportunities for evaluation of the capabilities, accuracy, and overall performance of remote sleep monitoring devices.

Panel Discussion: Pathways and Challenges to Utilization of Resources for More Precise Analysis and Diagnosis of Sleep Disorders (Circadian, OSA, Insomnia)

Panelists addressed key questions related to new discoveries resulting from the application of AI/ML approaches, the capabilities and limitations of these approaches, how to apply insights from big data and sleep studies to clinical practice, and what major clinical questions could be addressed through AI/ML.

Panelists noted that sleep data are multimodal, integrating measurements across multiple organ systems. Using AI technologies to analyze these data can help uncover hidden patterns and potentially identify novel biomarkers of sleep disorders and other conditions. Increased collaboration between clinical investigators and data scientists, as well as stronger standardization practices, can facilitate use of AI technologies to improve health outcomes. Additionally, panelists emphasized the importance of integrating contextual and subjective information with sleep signals and data. This additional information should include environmental factors, sex-specific differences, social and structure determinants of health, and PROs – all of which can influence sleep health and quality.

Key Opportunities:

• Integration of the various sleep-related signals available may permit a more comprehensive understanding of sleep and its interactions with various organ systems.
• Implementation of the Findability, Accessibility, Interoperability, and Reusability (FAIR) principles will ensure collection of representative high quality data.
• A large dedicated task force may be necessary to determine how to best approach the challenges surrounding standardization, benchmarking, interoperability, and harmonization of sleep-related data.

Examples of Applications of Big Data Approaches to Characterize Adverse Health Effects and Manifestations of Sleep Disorders

This session highlighted several applications of big data approaches in sleep research to provide insight into diverse sleep-related health conditions. For example, research has suggested mechanistic links between obstructive sleep apnea (OSA) and adverse health conditions, including cardiovascular disease, mental health issues, declining cognition, osteoarthritis), and early mortality. Definition of subtypes of sleep disorders may strengthen connections between poor sleep and adverse health outcomes.

One presentation focused on novel technology using radio signals as a remote and non-invasive approach to detect movement, nocturnal breathing patterns, neurological patterns, and other physiological signals during sleep. This approach enabled early detection of Parkinson’s disease, including assessment of the severity and progression of symptoms. Future investigations may use similar approaches for early detection of disease through analyses of aberrations in sleep-related signals.

Key Opportunities:

• Assessment of night-to-night variability of sleep and sleep patterns is a critical metric for characterizing overall sleep health..
• Further testing and validation of non-invasive technologies that measure sleep metrics could facilitate detection of sleep-related health issues with minimal patient burden.
• These remote, innovative approaches could improve understanding of the sleep-related causes of disease as well as examine the impact of medications and other factors on disease intensity and progression.
• Identification and better understanding of sleep disorder sub-phenotypes could lead to improved treatments.
• Integration of PROs and the physiological data collected could facilitate understanding of symptom subtype heterogeneity.

Big Data and AI Analyses of Sleep and Circadian Disorders Across the Lifespan and Their Differential Effects on Health

Session 4 focused on the effects of sleep disorders throughout the lifespan and the resulting adverse health impacts, including obesity, cognitive decline and neurodegenerative diseases. Insufficient sleep in adolescents is associated with greater risk of developing obesity. The links between sleep/wake disorders and neurogenerative diseases can be bidirectional. Additionally, sex-specific differences were noted, as menstruation, menopause, and societal factors impacted sleep health among women which in turn contributed to adverse cognitive and physiological health outcomes.

Speakers emphasized the importance of consistent data collection across sexes, ages, social determinants of health, and other factors so that data are truly representative. Development of new methodologies can ensure equitable data collection while leveraging AI, ML, LLM, and NLP approaches to improve data collection, harmonization, integration, and analysis.

Key Opportunities:

• Improving sleep habits for adolescents to decrease the risk of developing obesity.
• Balanced representation of men and women in sleep studies is critical for improved understanding of sex-specific differences in sleep-related health issues and for enabling more consistent and accurate diagnoses in men and women.
• Combination of sex-specific and life stage differences may lead to improved understanding of effects of childbirth and menopausal status on sleep and OSA.
• Moving beyond the use of averages and leveraging finer-grained demographic factors will improve risk predictions.

Focus on Population, Environmental Influences, and Ethical Issues

Population health, environmental influences, and ethical issues are all key considerations when collecting and analyzing sleep health data. Session 5 emphasized ensuring appropriate population infrastructure, consideration of environmental effects on the health of individuals and communities, and the ethical collection and leveraging of data using AI approaches. Consideration of societal context is important to ensure successful study design, community outreach, and implementation efforts. An effective data collection infrastructure can facilitate community engagement resulting in collection of information from individuals, families and communities.

The environments in which individuals and communities are situated can significantly affect sleep health. Environmental factors such as temperature, air pollution, light pollution, and noise can all contribute to poor sleep quality. The Ethical, Legal, and Social Implications (ELSI) framework was discussed in regards to AI, ML, and the use of big data in general. ELSI concerns such as transparency, consent, exacerbating biases, data privacy, liability, cybersecurity, and intellectual property are critical considerations in this space. When developing, training and validating AI technologies it is important to do so with an ELSI lens and ensure all relevant communities are well represented.

Key Opportunities:

• Incorporating geographic information at an appropriate scale can link environmental factors with sleep and circadian rhythm research and outcomes.
• Oversight mechanisms are necessary to ensure fairness and equity in application of AI technologies.
• Development of long-time meaningful partnerships with communities is critical to allow community oversight of development and implementation of big data approaches.
• Development and application of AI approaches must consider patient burden, community engagement, environmental factors, data privacy and security, bias mitigation, and ethical usage among many other factors.

Conclusions

Several new mechanistic insights have been gained through the application of AI/ML approaches to analyze sleep data in currently available databases. However, there are many opportunities to use these approaches to improve sleep-related health outcomes more equitably. Standardization of terminology, approaches, and data elements considered will enable future progress by promoting collaboration, harmonization, and data sharing. Additionally, a wide variety of individual and environmental variables should be considered in these analyses as they can exert significant influences on sleep and its relation to overall health. These variables in combination with traditional sleep metrics and PROs will permit development of a comprehensive picture of the determinants of sleep health.

Key Gaps and Opportunities

• Develop, train and validate new AI models, devices, and approaches in multiple diverse populations in a way that encourages community engagement and minimizes bias.
• Leverage new AI and machine learning algorithms to analyze existing databases that include multimodal data and signals that require analysis, harmonization, and integration.
• Through these analyses, gain improved understanding of the physiological manifestations of sleep, which factors have the greatest impact, and how to best mitigate these influences.
• Maintain progress towards standardization of sleep assessments to facilitate greater collaboration among investigators to ensure continued progress within sleep health and research.
• Delve further into identifying and understanding mechanistic linkages between sleep health, sleep disorders and adverse health outcomes such as cardiovascular complications, cognitive and neurological conditions, and obesity.
• Emphasize inclusion of ELSI principles to ensure equitable collection, analysis, and implementation of sleep data and interventions across various ages, sexes, races, communities, environments, and other populations.
• Design studies using AI/ML to achieve greater understanding of sleep health for women, especially at different stages throughout the lifespan, including those not traditionally included in research.
• Increase validation and implementation of novel wearable, nearable and airable technologies for remote collection and analysis of real-world sleep data.

Workshop Organizer and contact information:

Lawrence Baizer, PhD - Program Director, NCSDR, DLD, NHLBI

Workshop Co-Chairs and Moderators

• Shaun Purcell, PhD – Harvard University, Broad Institute, Co-chair
• Lauren Hale, PhD – Stony Brook Medicine, Co-chair
• Lawrence Baizer, PhD – NHLBI, NIH
• Inna Belfer, MD, PhD – NCCIH, NIH
• Todd Horowitz, PhD – NCI, NIH
• Dana Schloesser, PhD – OBSSR, NIH
• Sidd Shenoy, PhD – NHLBI, NIH

Workshop Speakers

• John FP Bridges, PhD – Ohio State University
• Regina Bures, PhD – NICHD, NIH
• Brian Cade, PhD – Brigham and Women’s Hospital, Harvard University
• Manisha Desai, PhD – Stanford University
• Massimiliano de Zambotti, PhD – SRI International
• Jeff Durmer, MD, PhD – Georgia State University
• Julio Fernandez-Mendoza, PhD – Penn State University
• Dina Katabi, PhD – Massachusetts Institute of Technology
• Marianthi-Anna Kioumourtzoglou, PhD – Columbia University
• Orsolya Kiss, PhD – Columbia University
• Sweta Ladwa, MPH – NHLBI, NIH
• Soomi Lee, PhD – Penn State University
• Diego Mazzotti, PhD – University of Kansas
• Kelton Minor, PhD – Columbia University
• Jonna Morris, PhD – University of Pittsburgh
• Girish Nadkami, MD, MPH – Icahn, Mount Sinai School of Medicine
• Ankit Parekh, PhD – Icahn, Mount Sinai School of Medicine
• Shaun Purcell, PhD – Harvard University, Broad Institute
• Carolyn Reyes-Guzman, PhD – NCI, NIH
• Azizi Seixas, PhD – University of Miami
• Bing Si, PhD – State University of New York at Binghamton
• Adam Spira, PhD – Johns Hopkins University
• Brandon Westover, MD, PhD – BIDMC, Harvard University

Disclaimer

The findings, knowledge gaps, and opportunities described here represent a summary of individual opinions and ideas expressed during the workshop. The summary does not represent a consensus opinion or directive made to or by NHLBI or NIH.