Artificial intelligence is already aiding in the diagnosis and treatment of sleep disorders. In the future, AI could drive precision medicine, ushering in a new era of patient care.

By Ann H. Carlson

When it comes to an individual patient’s treatment plan for sleep disorders, the more data, the better.

A single polysomnogram (PSG) provides invaluable information to help diagnose and treat one specific patient, but scientists have long understood that comparing hundreds—or even thousands—of tests can help identify otherwise undiscovered patterns across the general population. Using artificial intelligence (AI) solutions to analyze sleep testing data for new patterns is a logical step for researchers looking for connections to help predict sleep markers and even related health risks to promote better patient outcomes.

For example, in 2021, a team of French researchers led by Margaux Blanchard, PhD, published a study investigating the use of AI in helping clinicians evaluate the risk of cardiovascular disease morbidity and all-cause mortality in patients suspected of sleep apnea. The researchers used a tree-based machine-learning tool to analyze data collected from 9,876 patients, and the study recommends the AI tool be integrated into routine sleep testing moving forward.¹

This study is just one glimpse into how AI is helping to move sleep medicine toward more individualized interventions based on a more detailed understanding of a patient’s specific characteristics.

“In the future, AI can have wide applications in the field of sleep medicine, potentially driving precision medicine by helping clinicians identify the best possible, personalized treatment plans for each patient’s endophenotype,” says Jon Agustsson, PhD, vice president of AI and data science for Nox Medical, which incorporates a range of AI algorithms into its sleep medicine software, including those for sleep staging, respiratory flow analysis, apnea-hypopnea index (AHI), and periodic limb movement. “We are not there yet, but we have high hopes for the development of AI that can help us better understand each individual patient’s needs and how to best serve them,” he says.

Using AI to Detect Sleep Apnea Markers

The use of AI in sleep medicine has grown significantly in recent years. For example, deep neural networks—AI solutions based on a network of connected nodes inspired by the human brain—have been incorporated into some sleep medicine software solutions. This layered network approach helps AI detect patterns to more accurately predict sleep apnea markers and score sleep studies.

Traditionally scored manually, clinically established sleep apnea markers include the AHI, the oxygen desaturation index, and the arousal index.

“AI currently helps to automate annotation of sleep apnea markers,” says Sam Rusk, cofounder and chief AI officer of sleep software company EnsoData, who is also an AI committee member for the American Academy of Sleep Medicine. “AI is particularly good at processing large volumes of sleep studies and predicting sleep apnea markers consistently.”

AI can detect apnea by analyzing various data sources and patterns, explains Philippe Kahn, MS, founder and CEO of Fullpower-AI, a company that offers a contactless under-the-mattress sleep sensing system for at-home use.

“AI models are trained using labeled data that includes examples of apnea and non-apnea events,” he says. “The models learn to recognize patterns and features indicative of apnea, allowing them to detect and classify apnea events in real-time.”

One application is for AI to then analyze data from in-lab polysomnography. “For example, AI models can identify characteristic patterns in the respiratory airflow signal, such as the absence or reduction of airflow, and correlate it with other signals like SpO₂ or EEG abnormalities,” Kahn says. Other AI applications include under-the-mattress sensors and wearable devices, he adds.

How AI Learns

AI solutions use an iterative process based on repetition to learn and perform assigned tasks. “AI models contain sometimes millions of parameterized variables that are together trained to execute a task,” Rusk says. “The models learn patterns from the input datasets and are iteratively trained to increase the accuracy of a task.”

The models are then continuously improved and refined. “Once deployed, the model can make predictions on new data based on the patterns it learned during training,” Kahn says. “The iterative development process involves refinement, feedback, more data, and advanced techniques to improve the AI system over time.”

According to Agustsson, the most important ingredient for a successful AI application is the data itself. “The quality of the AI heavily relies on the quality of the data used for training and validation,” Agustsson says. “It is important to collect a large and representative dataset that aligns with the intended application of the AI.”

As a result, AI solutions learn best from manual data that provides consensus as well as clear definitions—why AI performs well when scoring sleep stages from EEG, EOG, and EMG data in sleep studies, Agustsson says.

“Generally, AI predicts sleep markers well where there is a good agreement among human scorers,” he says. “Additionally, AI has shown impressive performance in scoring sleep for other sleep disorders like narcolepsy, where consensus among sleep technologists is not as high. The classification of each 30-second period into wake, rapid eye movement (REM) sleep, or one of the three non-REM sleep stages (N1, N2, or N3) makes sleep stage scoring more manageable.”

However, Agustsson adds, AI progress has been focused on language and image processing rather than the unique physiological signals recorded in sleep studies. “These markers pose challenges in training AI models because the scoring of such markers can vary between individual scorers, hospitals, and geographical areas,” he says. “Even a single scorer may exhibit inconsistency in scoring over time, further complicating the learning process for AI.”

For this reason, AI can make errors in detecting respiratory events, desaturation events, arousals, limb movements, and hypopneas.

“Refinement of AI algorithms, larger and more diverse training datasets, and continued collaboration between experts and AI developers will contribute to the enhanced capability of AI in accurately detecting these challenging markers,” Agustsson says.

Despite current limitations, AI models positively impact sleep medicine overall by helping to standardize results, according to Agustsson. “AI is already making a huge impact on the efficiency and workflows of sleep labs across the globe,” he says. “I think it will continue to shape the future of sleep diagnostics as we develop new algorithms that provide new insights and inspire the evolution of our understanding of sleep patterns and physiology.”

Looking Ahead

As AI solutions continue to evolve, human expertise will continue to be integral to their implementation in sleep medicine.

“Human scientists and technicians play crucial roles in the machine-learning process,” Kahn says. “They are responsible for various tasks and decisions that contribute to developing and deploying AI systems, including problem formulation, data collection and preparation, feature engineering, model selection and configuration, training and validation, and ethical considerations.”

For the best outcomes for patients, it is also important for all stakeholders to have open and transparent communication to ensure the safe and effective use of AI, Agustsson notes.

“Overall, the collaboration and synergy between scientists, sleep technologists, medical doctors, and engineers are vital for the development of medical AI algorithms that are clinically relevant, reliable, and impactful in improving patient care in sleep medicine,” he says.

In the next decade, Agustsson predicts that AI will integrate more with electronic medical records. “Predicting patient outcomes and assessing the risk of certain outcomes may become more accurate and accessible through AI algorithms,” Agustsson says. “Additionally, AI can uncover subtle patterns and abnormalities in sleep data that might otherwise be challenging to identify, leading to more precise diagnoses.”

He also anticipates that precision medicine, aided by AI, will play a crucial role in the future of sleep medicine. “By analyzing extensive datasets, AI algorithms can identify patient profiles and characteristics that correlate with specific treatment responses or clinical outcomes,” Agustsson says. “This personalized approach can guide sleep professionals in tailoring treatments to individual patients, ultimately enhancing their efficacy.”

Of course, to make this possible, scientists require a huge quantity of reliable data to feed to AI solutions. “Among the greatest challenges for AI in sleep medicine is the availability of large, representative datasets for research and development,” Rusk says. “The greatest benefit is the massive potential to unlock clinically meaningful information from polysomnography recordings about sleep disorders and many other diseases.”

As the use of AI expands in sleep medicine, patients stand to benefit from new discoveries. “AI will continue supporting the sleep medicine field as it evolves to manage larger patient populations with better understanding of underlying sleep disorders,” Rusk says. “AI will help in multiple capacities, from automating diagnosis for sleep disorders to managing large populations undergoing treatment for sleep disorders.”

Augstsson also emphasizes the efficiency advantages that AI will bring to sleep medicine. “AI can optimize treatment approaches by identifying the most effective interventions for specific patients, leading to higher success rates,” he says. “Additionally, the integration of AI into clinical workflows can enhance efficiency, enabling sleep professionals to serve more patients and allocate their resources more effectively.”

While the evolving use of AI will continue to play a crucial role in patient outcomes, ultimately, of course, the patient’s treatment plan will always be in human hands.

“AI doesn’t solve the problem,” Kahn says. “AI gives practitioners the tools to deliver better and more customized patient care. And it will continually improve.” 

Ann H. Carlson is a freelance writer based in the Los Angeles metropolitan area.

Reference

1. Blanchard M, Feuilloy M, Gervès-Pinquié C, et al. Cardiovascular risk and mortality prediction in patients suspected of sleep apnea: a model based on an artificial intelligence system. Physiol Meas. 2021 Oct 29;42(10).

Photo 176515577 © Andranik Hakobyan | Dreamstime.com