What insights into sleep will we find when sleep studies are scored in a fraction of the time?

By Elisa Shoenberger, MBA

Polysomnograms (PSG) capture an incredible amount of data including brain activity, oxygen saturation, limb movements, heart rate, and so much more. Making sense of the data takes time and effort. And on top of it all, only a small part of the data is used regularly by sleep clinicians to help their patients.

But advances in artificial intelligence (AI) in sleep medicine may be able to change all that. Sleep researchers and medical device companies have been using supervised machine learning, a subset of artificial intelligence, to improve data scoring and working to predict patient outcomes.

But what exactly is supervised machine learning? It’s basically when you train a model with data inputs that then can make predictions or categorize data.

Working Smarter

Emmanuel Mignot, MD, PhD, says the sleep study results given to patients today represent only “one millionth of the richness of the data.” Artificial intelligence may provide a much more complete analysis of sleep. Photography by Steve Fisch for Stanford Health Care

For Emmanuel Mignot, MD, PhD, director of the Center for Sleep Sciences and Medicine and professor of sleep medicine at Stanford University, there’s a lot of potential with AI and sleep medicine. He says, “AI is most powerful when you have a lot of data sets in a very specific format. And that’s what sleep generates.”

With that data, researchers are working to have machine learning score sleep studies, like sleep staging, arousals, periodic limb movements, and more, alongside human scorers. It’s about “automating repetitive things” at the moment, explains Jón Ágústsson, PhD, director of analysis, data, and research at Nox Research.

Right now researchers are finding that machine learning has been effective in scoring sleep studies at least as good as human scorers, says Nathaniel Watson, MD, MSc, professor of neurology at University of Washington (UW) and co-director of UW Medicine Sleep Disorders Center. Better yet, they are faster and thus save time.

For Watson, it’s important that clinical providers understand the technology and not treat it like a black box. He says, “We have to be confident in these results to ensure that we are going to get the best possible treatment outcomes for our patients.”

Being able to more quickly and efficiently classify data will ultimately make clinicians available to do other work. The demand for sleep studies is growing, Ágústsson says, and the available sleep techs may not be able to meet it. AI can help techs be more efficient and free up their time to establish better relationships with their patients. 

But it’s important that people understand that machine learning is not a replacement for people, explains Amy Bender, MS, PhD, director of clinical sleep science at Cerebra, a digital health company focusing on sleep medicine. “Having input from a physician and a scientist is all part of making it work well together,” Bender says.

Hidden in Plain Sight

But there’s so much more potential for machine learning and sleep studies than scoring sleep stages and arousals. Mignot says the PSG results currently given to patients represent only “one millionth of the richness of the data.” There is so much more data that can be analyzed using machine learning like K-complexes, sleep depth, and heart rate changes.

In the real-world, of more interest than a patient’s sleep apnea itself is a prediction of daytime symptoms and comorbidities, for example, determining the patient’s tiredness and their risk for cardiac arrest or stroke. Mignot asks, “So if you kind of sit back and say, ‘why not predict that directly?’ That’s what I’m trying to do now.”

Watson agrees that AI is opening up new possibilities with sleep studies to see “things in it in ways that we’ve never been able to visualize before.” Or even thought to look for.

In fact, Watson notes, electroencephalography (EEG), not cardiopulmonary data, might be the most important part of the PSG. Watson says, “There’s machine learning and computational phenotyping that can be done just looking at the EEG that can come up with an assessment as to whether or not the individual has moderate to severe sleep apnea, with a fairly high sensitivity and specificity.”1 He thinks that it may be possible in the next 5 to 10 years that diagnoses of sleep apnea could be made based on EEG alone.

Sleep researchers at Flinders University in Australia are also using AI to analyze EEG data. These researchers are specifically looking at K-complexes, which occur roughly every two minutes during sleep and so are too labor-intensive for manual sleep scoring. 

Bastien Lechat, a PhD candidate at the Adelaide Institute for Sleep Health at Flinders University, says in a release, “A lack of K-complexes has been linked to various clinical problems, such as Alzheimer’s disease and insomnia, suggesting that K-complexes are an important part of normal sleep and health. While the meaning and role of K-complexes is rather unclear, one of the leading theories is that they reflect low-level decision processing to either wake up or stay asleep in response to sensory input during sleep.”

Lechat is first author on a paper detailing a deep-learning algorithm that automatically scores K-complexes.2 “We hope this algorithm will help to fast forward new discoveries regarding the mysterious K-complex waveform and its associated health outcomes,” he says.

AI might also facilitate a move away from one-size-fits-all sleep apnea treatments. AI is being used to determine in advance the patients who may not adhere to CPAP treatment, Bender says. If a clinician knows this information early on, they might be able to figure out a better way to get the patient to adhere.

Ultimately, AI could lead to “a path to personalized or precision sleep medicine,” says Watson.

Also exciting is the potential of machine learning analyzing PSGs along with other biometrics. By including people’s medical history outside of sleep, such as their weight, blood pressure, or even the medications they take, researchers might be able to better diagnose sleep disorders.

[RELATED: Pulling Back the Curtain on AI in Sleep Medicine]

Trust Builds Slowly

While AI shows a lot of promise for improving sleep scoring and patient outcomes, there’s been slow adoption in sleep labs. Education will be key. Watson says developers need to emphasize validation studies and that “sleep disorders diagnosed based on AI-scored sleep studies result in improved outcomes.”

Ágústsson thinks the approval of the US Food & Drug Administration (FDA) and the American Academy of Sleep Medicine (AASM) will be pivotal in gaining trust in the sleep science community.

In a release about the AASM’s 2020 position statement on artificial intelligence,3 lead author Cathy Goldstein, MD, says, “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. 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.”

The position statement recommends manufacturers disclose the intended population and goal of any program used in the evaluation of patients; test programs intended for clinical use on independent data; and aid sleep centers in evaluation of AI-based software performance.3

Mignot thinks it may be more difficult to see widespread adoption of AI into sleep labs/clinics given the current US medical healthcare system and ecosystem. Change is hard. He thinks AI use will occur through hardware manufacturers integrating the software into their devices.

However, some clinics have begun integrating AI. For instance, EnsoData’s sleep study autoscoring software EnsoSleep is currently used by 400 clinics across the United States. Chris Fernandez, co-founder and CEO of EnsoData, says its clients cite the time savings as opening up opportunities for increased patient interaction, such as giving them the manpower to launch sleep navigator programs. In a case report shared by EnsoData, John Childers, RRT, RPSGT, sleep center supervisor at St. Joseph Health, says, “EnsoSleep AI scoring frees up time for us to enhance patient care, develop new programs, build relationships with our referring network of physicians, and improve quality reporting initiatives.” 

Nox Medical’s Nox BodySleep, which estimates sleep states by processing respiratory data through advanced algorithms utilizing Nox calibrated RIP technology (without traditional EEG, electrooculography, and electromyography signals) has been integrated into the company’s products across Europe but not in the United States as the algorithm is not yet FDA approved.

At Home & In the Lab

How will AI impact home sleep studies? Bender explains the importance of at-home sleep studies since people do not tend to sleep as well in new environments.

Unfortunately, there’s limitations to at-home sleep studies since measuring devices can be invasive and difficult for the patient to use. Plus, it’s a challenge to get EEG information, since most home sleep apnea tests do not include the relevant sensors.

But there are companies that have developed more portable technology that can be used at home. For instance, Cerebra developed the Prodigy Sleep System, currently pending FDA approval, which can measure EEG, eye movements, muscle movements, respiratory channels, and more; it’s like “bringing the sleep lab into the home,” says Bender. While the company is working to miniaturize the device and hopes to add machine learning utility in the future, there’s a lot of possibility with collecting similar data as an in-lab PSG and using machine learning to make a diagnosis. It may be faster and cost less to get to that diagnosis, Bender says.

What’s more, Cerebra has already developed an algorithm—”odds product ratio” or ORP—that automatically calculates sleep depth based on the relationships of the powers of different EEG frequencies.4 Integrated into its Michele Sleep Scoring platform, ORP provides a single number between 0 (deeply asleep) to 2.5 (fully awake) to quantify sleep depth, adding an auto-scored element that goes beyond AHI.

Ultimately, as Watson says, “as long as our true north is always improving sleep health, not only for those with sleep disorders, but also the general population, as a sleep community, we can never go wrong. And you know, if we understand the ways that artificial intelligence can advance that goal and embrace it, then I think we have nothing but blue sky ahead of us.”

Elisa Shoenberger, MBA, is founder and owner of Bowler Hat Fox LLC and a freelance writer based in Chicago. This is her first article for Sleep Review.

Illustration 202885797 / Brain © Destina156 | Dreamstime.com

References

  1. Fernandez C, Rusk S, Glattard N, et al. 0932 Using novel EEG phenotypes and artificial intelligence to estimate OSA severity. Sleep (suppl_1). April 2019;42:A375.
  2. Lechat B, Hansen K, Catcheside P, Zajamsek B. Beyond K-complex binary scoring during sleep: probabilistic classification using deep learning. Sleep. 2020 Oct;43(10):zsaa077.
  3. Goldstein CA, Berry RB, Kent DT, et al. Artificial intelligence in sleep medicine: an American Academy of Sleep Medicine position statement. J Clin Sleep Med. 2020 Apr 15;16(4):605-7. 
  4. Younes M, Ostrowski M, Soiferman M, et al. Odds ratio product of sleep EEG as a continuous measure of sleep state. Sleep. 2015 Apr 1;38(4):641-54. 

(Sneak peek sponsor Cerebra is quoted in the article, along with other sleep medicine stakeholders who are working on AI solutions. It did not get prepublication approval rights for this content.)