The potential and limitations of AI algorithms in sleep care: Anuja Bandyopadhyay, MD
WATCHING TIME: 3 minutes
“An AI algorithm, for the most part, is only as good as the dataset it is trained on. For example, if I train it with data from a 4-year-old [population] who have a certain percentage of sleep and are normally healthy, after training my machine on this population, if I then take this algorithm and use it on, say, a 60 year old [population] with multiple morbidities, this may not work as well.
In sleep medicine, the American Academy of Sleep Medicine (AASM) has recently pushed to integrate artificial intelligence (AI) algorithms and tools into clinical care. Currently, their capacity is somewhat limited, but still useful, being used primarily for staging and sleep monitoring, which are currently labor intensive. The use of these AI tools could, therefore, reduce the time needed to collect rich data on sleep stages.
The AASM has set up a 2-year pilot program, called the AI/Autoscoring Certification Program, which will independently verify the performance of these auto-scoring systems.1 Currently, the focus is solely on polysomnograms for scoring sleep stages, with evaluated software needing to demonstrate equivalent or better accuracy than manual scoring to be certified, with program applications being accepted from late 2022 or early 2023.
NeurologyLive® spoke with Anuja Bandyopadhyay, MD, Assistant Professor of Clinical Pediatrics, Indiana University School of Medicine, and Chair of the AASM Committee on Artificial Intelligence in Sleep Medicine to learn more about what these AI algorithms could do for people in the field of sleep disorders. She spoke about the potential of these programs, as well as the importance of understanding their limitations and the importance of the data sets on which they are trained.
According to the AASM, companies using auto-scoring technology who would like to learn more about the AI/Autoscoring Pilot Certification Program can contact the AASM at [email protected]