How Machine Learning Could Transform the Way We Diagnose Narcolepsy
Researchers at Virtual SLEEP 2020 explained machine learning’s potential in sleep medicine.
By Lisa Spear
Even under the care of a sleep medicine clinician, narcolepsy patients may have a long wait before receiving a correct diagnosis. Researchers think that machine learning and the use of neural network analysis could speed up the diagnostic process, and pave the way for more precise care.
“Within the next few years, sleep scoring by technician will be replaced by automatic deep learning networks that will automatically annotate the sleep study and the task of the technician or doctor will only be to verify the event,” sleep scientist Emmanuel Mignot, MD, director of the Stanford Center for Narcolepsy, said during a presentation.
During Virtual SLEEP 2020, Mignot and a panel of other experts spoke about the future of artificial intelligence and machine learning in sleep medicine.
The panel discussed how these tools could potentially improve treatment for countless patients, lead to the use of phenotyping for diagnosing obstructive sleep apnea, and speed up the diagnostic process for those who experience narcolepsy with cataplexy, also known as type 1 narcolepsy.
By using statistical methods to find features unique to narcolepsy type 1, including a short REM latency period, researchers can build machine learning systems to help diagnose the disorder, Mignot explained. In the future, he said, all narcolepsy type 1 cases will be able to be detected remotely, from the patient’s home, over the course of a few days.
“We believe that this will be applied soon, where you will be able to wear a simplified device that you will wear at home for an entire weekend. Then, maybe you will have a blood test, combined with a deep learning algorithm, and that will get a beautiful diagnosis for narcolepsy.”
During his presentation, Mignot spoke about how he and his Stanford team have been working on creating machine learning systems to pinpoint narcolepsy type 1 cases. According to work presented during SLEEP, his team has demonstrated that using deep learning with polysomnography (PSG) could help clinicians bypass the multiple sleep latency test (MSLT), while producing an accurate diagnosis.
The team created a score for a sample of PSG recordings to reflect how close the recordings are to narcolepsy type 1 patterns. “The model generalized remarkably and had a high predictability for diagnosing narcolepsy,” said Mignot, professor of psychiatry and behavioral science at Stanford University.
Machine learning, the researchers said, can analyze large swaths of data without human bias. By applying machine learning and analyzing the data further during nocturnal PSG in narcolepsy type 1, clinicians could raise the specificity of detecting different sleep stages and transitions, Mignot said.
Another advantage is the ability of machine learning to score very brief periods of time. Instead of the 30-second epoch, which is typically scored by a human technician, machine learning can score up to a 5-second window, said Mignot, who is leading a sleep analytics project called the Stanford Technology Analytics and Genomics of Sleep (STAGES).
“Similarly, also you can compare the performance of the machine learning network to each technician. What you can show, in fact, is that the machine learning routine is closer to the consensus of all technicians than any single technician. In summary, machine learning is doing better than any single technician at recognizing all the sleep stages,” Mignot explained. “It has a superior performance.”
While it is known that narcolepsy with cataplexy is caused by a hypocretin deficiency, measuring hypocretin requires a lumbar puncture, an invasive procedure that is unpleasant for patients and not commonly used as a sleep medicine diagnostic tool.
Typically, diagnosis instead involves an overnight stay for an in-lab p PSG, followed by a daytime MSLT. The MSLT measures excessive daytime sleepiness by asking patients to nap 4 to 5 times for 20 minutes every 2 hours during the day. During these naps, sleep latency and the presence of REM sleep are observed.
Unfortunately, the MSLT can produce both false positive and false negative results, says Mignot.
“The fact that the MSLT is not a perfect test for narcolepsy, and the fact that it takes quite awhile for the MSLT, PSG to be conducted, during the night and then during the day, led us to believe that there might be a better way to analyze the data of a patient with narcolepsy and maybe diagnose narcolepsy with only one night of sleep using machine learning,” Mignot said.
Another flaw in the current diagnostic process for detecting narcolepsy type 1 is using the observation of cataplexy as diagnostic criteria.
“The problem with cataplexy is that it is subjective, so it cannot be 100% certain as a predictor,” Mignot said.
Cataplexy is also not present in all cases of hypocretin deficiency, said Nathaniel Watson, MD, MSc, director of the Harborview Sleep Clinic and co-director of the University of Washington Medicine Sleep Center in Seattle, during the Virtual SLEEP 2020 session.
Watson explained that machine learning systems “could save time and increase the probability of diagnosing patients in sleep clinics.”
Lisa Spear is associate editor of Sleep Review.
from Sleep Review https://www.sleepreviewmag.com/sleep-disorders/hypersomnias/narcolepsy/machine-learning-diagnose-narcolepsy/