Melissa Jones, MD, shares insights into recent artificial intelligence (AI) innovations in the realm of sleep studies.
Melissa Jones, MD, professor of psychiatry, Baylor College of Medicine, joined The American Journal of Managed Care® (AJMC®) to discuss notable innovations with artificial intelligence (AI) and sleep research, as well as to give her perspective on the validity and accuracy of wearable devices that track sleep data.
Jones presented at SLEEP 2024, the American Academy of Sleep Medicine and Sleep Research Society, Annual Meeting with her session covering “Artificial Intelligence and Big Data in Sleep Medicine.
Transcript
What do you believe are the most significant advancements in AI-driven sleep research over the past few years?
I think one of the major advancements has been the development of technologies that enable sleep staging with ambulatory technology, or wearable technology. I think this has the opportunity to enhance access to sleep medicine care. I think it also may help facilitate physicians who are facing reimbursement cuts, staff shortages, and burnout. It may help them get more sleep studies performed and more patients treated.
Can you explain how AI algorithms analyze sleep data from wearable devices? How accurate are these analyses compared with traditional methods?
The wearable devices collect raw data using various sensors; for instance, heart rate monitors or accelerometers that detect motion. And then this raw data is preprocessed and specific features are typically extracted from it such as heart rate variability, for instance. And then the machine-learning algorithm is trained on expertly-labelled data. So, in the case of sleep study staging, the 30-second Epics are manually labeled by sleep medicine physicians and that is what the machine learning algorithm is trained on. And then the algorithm detects patterns in the data. And then this machine learning algorithm is fed a validation set of unseen data, and it is validated against this expertly labelled data. So, the accuracy varies.
There's still a ways to go. One study, in the case of REM sleep behavior disorder, did report an accuracy of 96% to 100% in classifying patients with Parkinson disease and REM sleep behavior disorder vs no REM sleep behavior disorder; that was using actigraphy watch monitoring for 14 days and a support vector machine. But again, there's still a ways to go. And there needs to be more testing with large, diverse datasets.