By Robyn Woidtke, MSN, RN, CCSH, RPSGT (#340), FAAST
This article will address several key concepts: 1) the changing landscape of the provision of sleep medicine, 2) how artificial intelligence (AI) and machine learning (ML) can facilitate patient care; and 3) the changes necessary to implement and embrace AI and ML.
We are experiencing an avalanche of developments within the field of sleep medicine. The worldwide prevalence of obstructive sleep apnea (OSA) is estimated to exceed 1 billion!1 Concurrently, there is an emerging decline of a trained and knowledgeable sleep medicine workforce, including physicians, sleep technologists, behavioral sleep practitioners and knowledgeable clinical coordinators from a variety of backgrounds.2-4 Consequently, the field could find itself without sufficient practitioners to address patient demand: appointment backlogs may occur, and many patients may not have access to appropriate sleep care. The last several years have posed yet a new set of challenges; the pandemic, supply chain issues and recalls have been a wake-up call for all of healthcare. Changing and adapting how we deliver sleep health care has never been more important.
Sleep medicine’s adoption of technological advancements
Over time, the field of sleep medicine has had dramatic technological changes impacting the practice of sleep medicine. For decades, we relied on electroencephalograms modified to receive inputs from various ancillary equipment and record physiologic parameters on paper. Using ink and galvanometers to create the recordings (a beautiful sound!), we meticulously scored by hand, calculated the indices with a calculator, and manually created written reports. We stored boxes and boxes of paper recordings, and space was a challenge.
The early 1990s heralded the use of computerized sleep data acquisition systems. These systems enabled various permutations of recording montages with the flick of a switch; we could replay the night and examine the resulting data in ways that were previously impractical. Following scoring, calculations were done automatically, and a report was generated, saving hours and hours of sleep technologist time and effort.
The advent of commercialized continuous positive airway pressure (CPAP) devices in the mid-1980s changed the face of sleep medicine and quickly became the cornerstone treatment for obstructive sleep apnea (OSA). Technologists adapted and learned overnight titrations of CPAP, discovered how to craft to the best pressure, and worked with patients to adapt to home use using a large and noisy machine—while simultaneously we were also trying to figure it all out ourselves! Over the years, there has been a constant barrage of new masks to learn about and choose from to use humidification or not and what setting to choose, and we have been obliged to adapt to evolving advances for data storage and adherence reports.
Then, along came the AutoPap! This new technology initially created some angst among many sleep technologists; it seemed as if it might supersede in-lab titrations, and it was unclear how that would change the technical landscape and impact the role of sleep technologists. However, yet again, we adapted and modified our practice to accommodate the AutoPap device into clinical care.
Similarly, the variety of home sleep tests (HSTs) has exponentially increased over the past five years, including smaller sensors, apps, and an explosion of wearables and nearables. An important consideration when choosing the appropriate device is that some devices are over-the-counter/wellness devices, whereas others are cleared and regulated by the United States Food and Drug Administration (FDA). Monitoring sleep in normal environments for days or weeks engenders a new appreciation of what can be learned and eventually applied to patients with sleep health issues. Most of these technologies relate to sleep in general or OSA/sleep-disordered breathing and have limitations; thus, in many cases, in-laboratory recordings are still needed. However, recently, more HST devices have included electroencephalogram and eye movements to better assess actual sleep time.
In 2008, the Centers for Medicare and Medicaid made the decision to permit HSTs to fulfill the requirement for an OSA diagnosis for the purpose of prescribing CPAP.5-6 While this change posed some difficulties for sleep clinics in terms of revising their workflow and creating a new paradigm of care, the availability of HSTs also resulted in improved access to diagnosis and treatment for many patients. Again, the field rose to the challenge and integrated the use of HSTs into a robust practice of clinical care. In 2018, the AASM updated its policy statement for HSTs and reiterated the requirement to apply medical judgment: only skilled medical professionals with sleep expertise should be utilizing this technology.7
So, what is AI/ML and how does the field of sleep health benefit?
We are now entering the latest phase of our craft: artificial intelligence (AI) and machine learning (ML). The following are very simple descriptions of AI and ML: according to the English Oxford dictionary8, AI is defined as “the theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages.” Viewed as a component of artificial intelligence, machine and deep learning (ML/DL) are concepts to be familiar with and include the “ability of the system to automatically learn from trial and error, through experience and by the use of data.” AI is also regulated as software as a medical device (SaMD) with regulatory and quality systems standards to which manufacturers must adhere.9
AI/ML functions on a variety of levels that are capable of different tasks. Figure 1 provides an overview of the various types of machine learning.
Figure 1. Machine Learning (From TechTarget)
Currently, the use of AI/ML has been focused on scoring polysomnograms and home sleep tests. Using AI/ML programs has resulted in faster report turnover, improved patient throughput, and a growing knowledge of the application. Using these new technological advances, we can also begin to improve diagnostic capabilities and treatment, thereby enhancing outcomes. It is also reasonable to assume that in the near future, we will be able to identify phenotypes leading to personalized treatment or predict who will be a responder to an oral appliance.
Is AI/ML the same as autoscoring? In a sense, yes; however, AI goes far beyond what autoscoring can accomplish. Autoscoring works by a set of rules—“if then, else,” and it may not have the discriminatory capabilities to assess sleep changes that are not in the rule set, such as atrial fibrillation or subtle changes in rapid eye movement sleep. In contrast, AI/ML can view PSG/HST waveforms from a different “vantage” point, enabling it to discern electrophysiological distinctions that may not even be identified by the human eye.10-11 An AI-scored PSG/HST is typically completed in around 10 minutes. A technician then reviews the scored study and makes any corrections—usually within a few minutes. The really neat thing about AI/ML is that as clinicians or technologists make changes to the study and more data are collected, the training model can be applied, and then those adjustments can be validated to continue to evolve and provide more consistency in scoring.
The past few years have seen improvements in agreement between scoring with AI and well-trained, experienced sleep technologists.10 AI requires training datasets to learn from; these should come from many different sources, providing diversity in race, gender, and phenotypes to prevent bias in the algorithms. Verification and validation testing are needed to ensure a robust platform. In view of this, technologists and physicians gain more confidence in the AI-scored data, leading to improved diagnostics and time management. It is an iterative process. In addition to the technical requirements, education and training of sleep center staff is imperative for successful adoption. Implementation of workflow and logistics should also be considered.
As indicated above, using AI to score PSG and integrate report generation could provide clinical sleep staff with more time for patient education and training, coaching, and monitoring. This is an exciting notion! The promise of these technological advances puts sleep medicine in a position to explore sleep as never before. It also places our field in alignment with other medical specialties such as dermatology, radiology, surgery, and oncology. In fact, AI has demonstrated comparable or better outcomes compared with clinicians in these areas in determining cancer and other conditions.10 In 2020, the American Academy of Sleep Medicine (AASM) published a position statement supporting the use of AI, which they describe as being “well suited” for the large amounts of data captured by PSG;12 however, caution must be taken to ensure safety, ethical use, and legal considerations.
AI/ML must be carefully integrated into the sleep health care paradigm. AI is not only used for patient classification and diagnosis with PSG, but is also increasingly used in other applications. Oral appliance therapy (OAT) can be produced using additive manufacturing (AM/3D printing) to craft patient-matched devices. AI is used to enhance the computer-aided design (CAD) and computer-aided manufacturing (CAM) processes, thereby resulting in improved quality, durability, reduction in waiting times and better overall fit. Another application is computational fluid dynamics (CFD) modeling. In a retrospective study by Yeom et al., 13 the authors found that using CFD and ML for CT scans of the upper airway enabled researchers to discern OSA classifications in male patients with 80% accuracy in training and test cases.
AI adoption is a necessary leap into the future, but change is difficult
Innovation is an important tool for organizational revitalization, development, and competitiveness. However, attempts at innovation adoption do not necessarily lead to the expected benefits due to implementation failure rather than innovation failure. Often, implementation failure is the result of companies overlooking the importance of employees’ willingness to accept the change and continue using the innovation. According to Harvard Business Online, 14 changes within organizations fall into two areas: 1) adaptive and 2) transformational. Adaptive changes are typically small and occur over time, whereas transformational changes have a larger impact and are often sudden. What we are experiencing in AI is more of an adaptive change. However, although these changes may be small and incremental, ensuring a successful change management process increases employees’ acceptance and confidence in handling them.
Figure 2 Kotter’s 8-Step Model
Management can take steps to ensure a successful organizational change. Kotter’s 8-step model shown in Figure 2 above helps to illustrate how to create a climate for change, make the change happen, and implement and sustain the change.
Creating a climate for change involves clearly defining and aligning it to the business goals of the organization. These steps help to align the change’s value and quantify the effort that will be devoted to it. Making the change happen requires the removal of obstacles by clearly determining whom it impacts and developing a communication strategy, particularly for those not fully involved with the decision-making process. In times of transformation, leaders need to develop a change story that helps all stakeholders understand where the company is headed, why it is changing, and why implementing AI is important.
Finally, making changes stick involves implementing a support structure that helps employees emotionally and practically adjust and build the behaviors and skills they will need to achieve the desired business results. This process entails measuring the business impact of the changes and ensuring that continued reinforcement opportunities exist to anchor them. Change does not happen overnight for most of us; however, if done correctly, the desired results should be achieved.
The above brief historical summary speaks to the adaptability, creativity, and sustainability of the field of sleep medicine. AI/ML is yet another step in the journey and evolution of sleep medicine. Imagine that we can indicate which patients will accept and use CPAP or who will be a responder to an oral appliance merely by looking at a summary of data based on waveform analysis. We may be able to identify individuals who are at higher cardiovascular risk from sleep apnea or those who continue to have excessive sleepiness. It would be amazing to be able to identify individuals who might develop insulin resistance based on looking at physiological data from PSG, computational fluid dynamics, or other techniques. As we move toward the future, AI may enable all of this and more! The ability to integrate personalized sleep medicine will empower sleep health professionals to better inform our patients regarding outcomes and provide a more robust patient experience. We will be able to provide care that fits the individual rather than fitting the individual to the care.
The patient experience (PX) is different than patient satisfaction. PX encompasses the entire patient journey, including who they interact with, communication, timely appointments and getting results of tests. AI in this context can provide easier scheduling, faster sleep study results and more time to discuss therapeutic options. AI is therefore not only a tool for scoring; it also can be implemented in other ways to enhance the patient experience. Improved patient experiences can lead to better outcomes.
In conclusion, sleep health professionals must have the courage to adapt and change with the times. We must recognize that AI holds promise not only in scoring but also in the way that we care for patients. We are a world of people! Each one of us is unique with differing needs. Personalization of care using AI will improve access to “care that fits” with the optimism that it will and ultimately lead to better healthcare outcomes.
To quote Jimmy Buffett, “It’s these changes in latitudes, changes in attitudes. Nothing remains quite the same […] Oh, yesterday’s over my shoulder—So I can’t look back for too long—There’s just too much to see waiting in front of me.”
A good resource for sleep health professionals: American Academy of Sleep Medicine Website
Robyn Woidtke, MSN, RN, RPSGT, CCSH, FAAST is the Editor-in-Chief of Sleep Lab Magazine, a consultant to sleep and medical device companies, and the principal at Sleep for Nurses, a sleep-focused educational endeavor specifically developed for practicing nurses.
Conflict of Interest Statement
RW works as a consultant to the sleep and medical device industries. Of relevance to this article, she has received consulting fees from Ensodata.
- Grote, Ludger. “The global burden of sleep apnoea.” The Lancet Respiratory Medicine 7.8 (2019): 645-647.
- Watson, Nathaniel F et al. “The past is prologue: The future of sleep medicine.” Journal of clinical sleep medicine : JCSM : official publication of the American Academy of Sleep Medicine 13.1 (2017): 127-135. doi:10.5664/jcsm.6406
- American Association of Sleep Technologists. State of Sleep Technology and Workforce. Retrieved July 1, 2022 from https://www.aastweb.org/newsandevents/executive-summary-state-of-sleep-technology-and-workforce
- Collen, Jacob F., et al. “Losing sleep! Are we missing the future of sleep medicine?.” Journal of Clinical Sleep Medicine 16.4 (2020): 473-474.
- Centers for Medicare and Medicaid Services. Sleep Testing for Obstructive Sleep Apnea. National Coverage Analysis Decision Memo Retrieved from https://www.cms.gov/medicare-coverage-database/view/ncacal-decision-memo.aspx?proposed=N&NCAId=227&ver July 7, 2022
- Chediak, Alejandro D. “Why CMS approved home sleep testing for CPAP coverage.” Journal of clinical sleep medicine 4.1 (2008): 16-18.
- Malhotra, Raman K., et al. “Polysomnography for obstructive sleep apnea should include arousal-based scoring: an American Academy of Sleep Medicine position statement.” Journal of clinical sleep medicine 14.7 (2018): 1245-1247.
- Oxford Dictionary via Lexico Artificial Intelligence retrieved June 30 2022 from https://www.lexico.com/definition/artificial_intelligence
- United States Food and Drug Administration. Artificial Intelligence and Machine Learning in Software as a Medical Device Action Plan. Retrieved from https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-software-medical-device July 7, 2022
- Lovejoy, Christopher A., Abdul-Rahman Abbas, and Deeban Ratneswaran. “An introduction to artificial intelligence in sleep medicine.” Journal of Thoracic Disease 13.10 (2021): 6095.
- Nikkonen, Sami, et al. “Automatic respiratory event scoring in obstructive sleep apnea using a long short-term memory neural network.” IEEE Journal of Biomedical and Health Informatics 25.8 (2021): 2917-2927.
- Goldstein, Cathy A., et al. “Artificial intelligence in sleep medicine: an American Academy of Sleep Medicine position statement.” Journal of Clinical Sleep Medicine 16.4 (2020): 605-607.
- Yeom, Seung Ho, et al. “Computational analysis of airflow dynamics for predicting collapsible sites in the upper airways: machine learning approach.” Journal of Applied Physiology 127.4 (2019): 959-973.
- Harvard Business Online. 5 Critical Steps in the Change Management Process. Retrieved from https://online.hbs.edu/blog/post/change-management-process July 5, 2022