From identifying new sleep disorder phenotypes to rapidly scoring sleep studies, artificial intelligence may be a boon for precision sleep medicine.    

By Lisa Spear

Sleep studies record billions of biological waveform data points—the rise and fall of the chest as patients breathe, the electrical pulses between neurons in dream states, and the rapid eye movements that can be detected during the sleep cycle. These are a few of the clues to understanding patients’ health, but decoding the meaning of them and detecting all the underlying patterns is not always possible with just the human eye.  

Artificial intelligence (AI) and machine learning could identify patterns from large population datasets that may provide profound insights into health. Data gathered during sleep and analyzed by AI could help predict who is going to get sick in the future, long before any symptoms appear, says Cathy Goldstein, MD, MS, associate professor of neurology at the University of Michigan Sleep Disorders Center in Ann Arbor. Theoretically, a restless night’s sleep could predict who is going to later get dementia or changes in heart rate variability could signal the onset of pneumonia, information that could be used to help track infectious disease outbreaks.

There’s already evidence that certain sleep patterns can lead to other diseases. One study found that the identification of previously undefined subtypes in obstructive sleep apnea (OSA) were predictive of cardiovascular events. The researchers anticipate that routine identification of these and other subtypes from polysomnography (PSG) signals may help to better define how sleep-disordered breathing contributes to other disease states in the future.1 Another study found that sleep characteristics may predict all-cause mortality in some men.2 AI and machine learning could one day be used to drill into large population datasets to create more precision care.

“There are elements of sleep where AI could be useful and could be a big disruptor,” says Goldstein, head of the American Academy of Sleep Medicine (AASM) Sleep Medicine Artificial Intelligence Committee, a group that monitors advancements in AI in sleep medicine and provides information on how this technology could impact the field. 

Fifty to 70 million Americans are affected by sleep disorders and intermittent sleep problems, according to the National Sleep Foundation. Hidden in these statistics, medical researchers think there are undetected sleep disorder phenotypes, which if identified could lead to a better understanding of the pathophysiology of sleep disorders, earlier diagnoses, and faster treatment.

According to a paper from researchers at Yale University School of Medicine, machine learning is a promising phenotyping strategy that can integrate multiple types of data, including genomic, molecular, cellular, and clinical, to identify meaningful OSA phenotypes.3

“AI can look at signatures that you wouldn’t be able to analyze with regular statistics with massive data sets,” says Goldstein.

Machine learning and AI have already taken hold in other medical specialties, including radiology4 and pathology at Stanford University, where scientists are exploring ways to detect COVID-19 infections by monitoring Fitbit and other wearable device data. “My lab wants to harness that data and see if we can identify who’s becoming ill as early as possible,” Michael Snyder, PhD, professor and chair of genetics at the Stanford School of Medicine, says in a statement.

In sleep medicine, AI could also change the way clinicians use the in-lab PSG, the cornerstone of sleep diagnostic testing. These studies collect rich data sets, but currently, the PSG is typically scored by sleep techs in a way that is very structured, and overly simplistic, says Goldstein.

“Overall, we are seeing all this really interesting psychological information about our patients and we are just kind of oversimplifying it into metrics and into reports that say, ‘Do you have sleep apnea, yes or no?’”

PSG analysis today is often focused on the apnea-hypopnea index (AHI), which is used to diagnose sleep-disordered breathing. This AHI-focused approach likely contributes to the difficulties of better understanding the genetic and biological underpinnings of the disorder as well as to the modest treatment effects found in large randomized trials using continuous positive airway pressure (CPAP),3 the Yale researchers wrote.

“One way to address these challenges is to classify the disorder into smaller, more homogeneous categories. Such classifications, sometimes referred to as ‘phenotypes,’ can be based on clinical, pathophysiologic, cellular, or molecular characteristics,” according to the Yale paper.

The PSG study is typically scored by a sleep tech, who goes through the data, looking at each 30-second window of the patient’s sleep. This takes time and introduces human bias, explains Alyssa Cairns, PhD, head of sleep research at SleepMed BioSerenity, which develops data analytic tools through artificial intelligence to detect biomarkers.

Innovators in the field are working on ways to change the way sleep studies are interpreted and some are even hoping that sleep medicine can refine some of the current metrics that define sleep disorders.

Madison, Wisc-based startup EnsoData has made its way to hundreds of sleep clinics across the country with the promise that the company’s AI-powered software EnsoSleep, provides AI scoring that allows clinicians to review PSG studies and home sleep tests in minutes.

This frees up sleep techs to spend more time with patients, says EnsoData CEO and cofounder Chris Fernandez, MS.

Fernandez says this technology could lead to more affordable care and reach a greater volume of patients including underserved populations. “In order to drive good outcomes, we need to have access, in order to have access, we need to make healthcare more affordable,” says Fernandez.

The company is backed by $9 million in investor funding for EnsoSleep, which received U.S. Food and Drug Administration clearance to score sleep studies in 2017.

AI tools could also potentially help cut down on the long diagnostic delays for people with disorders that may otherwise take years to identify, explains Cairns.

Another diagnostic test that could see improvements from new AI programs is the multiple sleep latency test (MSLT), which is currently used to confirm a narcolepsy diagnosis but is not as effective at spotting idiopathic and physiologic hypersomnias.5

“Machines can help us determine the predictive probability of a central nervous system hypersomnia that would not otherwise get picked up on an MSLT,” says Cairns. Cairns and her team are working with about 600,000 raw patient data records on multiple projects to develop new AI-powered products to help the medical world glean novel insights about sleep population health.

“AI could allow us to derive more meaningful information from sleep studies, given that our current summary metrics, for example, the apnea-hypopnea index, aren’t predictive of the health and quality of life outcomes that are important to patients,” says Goldstein. “Additionally, AI might help us understand mechanisms underlying obstructive sleep apnea, so we can select the right treatment for the right patient at the right time, as opposed to one-size-fits-all or trial and error approaches.”

Lisa Spear is associate editor of Sleep Review.


1. Zinchuk AV, Jeon S, Koo BB, et al. Polysomnographic phenotypes and their cardiovascular implications in obstructive sleep apnoea. Thorax. 2018 May;73(5):472-80.

2. Wallace M, Stone K, Smagula S, et al. Osteoporotic fractures in men (MrOS) study research group, which sleep health characteristics predict all-cause mortality in older men? An application of flexible multivariable approaches. Sleep. 2018 Jan 1;41(1):zsx189.

3. Zinchuk AV, Gentry M, Concato J, et al. Phenotypes in obstructive sleep apnea: a definition, examples and evolution of approaches. Sleep Med Rev. 2017 Oct;35:113-23.

4. Reyes M, Meier R, Pereira S, et al. On the interpretability of artificial intelligence in radiology: challenges and opportunities. Radiol Artif Intell. 2020 May 27;2(3):e190043.

5. Lopez R, Doukkali A, Barateau L, et al. Test-retest reliability of the multiple sleep latency test in central disorders of hypersomnolence. Sleep. 2017 Dec 1;40(12).

Editor’s Note: The article has been updated to reflect that PSG studies are typically scored by looking at each 30-second epoch of a patient’s sleep, not 30 minutes.