By Sree Roy

CPAP adherence rates haven’t budged much in decades. Sure, the statistics have shifted for specific sleep practices or durable medical equipment providers, as well as individual patients; however, there have been no industry-wide leaps. But now, a grand-scale solution to minimizing CPAP failures is poised for entry. Artificial intelligence (AI)—with its capability of distilling reams of data into usable recommendations—could revolutionize sleep apnea therapy in several ways: by coaching CPAP users, personalizing therapeutic pressures, and even predicting which patients would do better on an alternative therapy. 

Coaching CPAP Users

Reaching out to struggling CPAP users is good practice. When a sleep tech or respiratory therapist is responsible for more patients than they can reasonably call, AI can step in. 

When there are upwards of thousands of patients, a person “would never be able to process and synthesize that much data to predict accurately and then create an organized priority list and task list for coaches,” says Sam Rusk, co-founder and chief AI officer of EnsoData. But an AI-powered solution like EnsoTherapy excels at this. “It’s just doing a higher dimensionality prediction of the future, more so than any simple algorithm could do,” he says. The sleep coach then has the most impact by troubleshooting with patients who need immediate help.

In the future, AI chatbots will likely also coach patients directly, as well as escalate individuals to a live clinician as needed. Ankit A. Parekh, PhD, assistant professor, medicine – division of Pulmonary, Critical Care, and Sleep Medicine and an assistant professor in the department of AI and Human Health at the Icahn School Medicine at Mount Sinai, says, “Chatbots or ‘conversational agents’ are being utilized in several different fields of medicine, and I think sleep is perfect for them.”

Personalizing CPAP Pressure

In-lab titrations and APAP algorithms personalize device pressures today, and AI will take CPAP pressure personalization to the next level—ideally improving disease burden, therapy tolerability, and adherence.

For example, Harvard Medical School researchers, along with Nox Medical, are working on an AI-powered model that analyzes sleep apnea endophenotypes. The Food and Drug Administration (FDA)-cleared tool, expected to integrate into Nox Medical’s solutions shortly, will provide sleep physicians with insights into parameters including arousal threshold. If, for example, a patient has a high arousal threshold, the AI may note that they can tolerate high CPAP pressures. “The next thing on the to-do list is: How do we present this information to physicians to help them understand the patient’s pattern, gain confidence that they understand the outputs of the device, and that they can feel empowered in drawing conclusions from them?” says Jon S. Agustsson, PhD, vice president of AI and data science at Nox Medical.

Meanwhile, a team at NovaResp Technologies Inc has trained AI on millions of breathing samples from polysomnography recordings and CPAP data downloads, and it can now predict CPAP users’ upcoming normal breathing, obstructive apneas, and central apneas. NovaResp’s software (which has not been submitted to the FDA yet) then adjusts CPAP pressure in advance of forecasted events. “We very much respect CPAP and APAP; it has saved countless lives,” says Hamed Hanafi, PhD, founder and CEO of NovaResp. “This is just an inevitable evolution of bringing prevention, instead of reaction, of apneas to market.”

The company has nearly completed a clinical trial comparing adherence of its proactive positive airway pressure software versus traditional CPAP. So far, the AI-powered solution leads, with the gap between it and CPAP enlarging as time passes. “They’re experiencing more comfort, less leak, lower pressures, and better sleep staging,” Hanafi says. “That must be why they’re sleeping longer” on the AI-powered software.

NovaResp’s roadmap includes even more personalization. The company plans to earn approval for, in order: its fixed AI model, which generates population-level probabilities of upcoming respiratory events; AI models tailored for sleep apnea phenotypes; and, ultimately, individual AI models that continuously learn from each sleep apnea patient’s breathing patterns.

Faster Alternative Therapy Routes

Another tactic to minimize CPAP failures is to stop prescribing the devices to people who are not going to use them (much easier said than done). AI models will eventually assist here too, perhaps by sharing a probability of how likely an individual is to have their sleepiness resolve with specific therapies.

The biggest hurdle isn’t the AI; it’s the humans charged with maintaining electronic medical records. “One of the biggest requirements is good, clean, labeled data,” Parekh says.

That’s why Mount Sinai, as well as other institutions, is now prioritizing medical record comprehensiveness. The records must capture the clinical care timeline from initial screening to symptom resolution. Blanks sometimes include exact CPAP usage hours (not simply whether a payor’s adherence criteria are met), titration records from alternative therapies, and sleep care from other health systems. “With the advancement of the GLP-1s and other medications, I feel this is going to be the biggest utility: knowing whether firstline treatment of CPAP or something else should be preferred for the patient,” Parekh says. “These AI models, I think, can definitely help.”

Heidi Riney, MD, chief medical officer at Nox Health, echoes the sentiment. “Obstructive sleep apnea is a chronic condition that, when left untreated, can drive healthcare costs and utilization and worsen outcomes for patients with other chronic diseases, making it even more critical to guide each patient to the right treatment modality from the start. AI is helping us tailor therapy by analyzing data from sleep studies, comorbid conditions, and patient preferences to recommend treatments patients are more likely to succeed with—empowering physicians to make more informed, patient-specific decisions that ultimately improve patient adherence and outcomes.”

Monitoring Therapy for the Long Term

EnsoData’s Rusk adds, “There’s also less confidence in alternative therapies because there’s a lack of monitoring of the effectiveness, especially over the long term.” Though the person may initially be titrated with diligent monitoring, “after a year or two, your body might change or your response to the therapy might change.” Typically, by then, no one is interpreting the data from the device (or an add-on wearable sleep test) through the lens of potential adjustments. 

“At-home monitoring over a longer-term time scale results in a lot of data,” Rusk says. “The idea that [clinicians] can monitor effectiveness of treatment for somebody over years is very powerful and not really possible without an AI capability.”

Learn More About Sleep AI: