Sleep labs start certifying AI

A sleep study is the medical version of wiring a car to a dashboard full of sensors and then watching what happens overnight. A full polysomnogram records brain waves, breathing, oxygen levels, heart rhythm, eye movement, muscle activity, and leg movement while a patient sleeps in a lab. (aasm.org) A human scorer then turns that raw signal into a report by marking when a person is awake, when they enter each sleep stage, and when events like apneas or limb movements happen. The American Academy of Sleep Medicine says the new certification covers adult sleep stages and events from full polysomnography data, not just one narrow slice of the test. (aasm.org) Autoscoring software is built to do that first pass automatically, the way speech-to-text software drafts a transcript before a person fixes the mistakes. Sleep labs have used these tools for years, but the academy’s new Full Polysomnogram Autoscoring Certification Program is an outside check on whether the software performs well enough to trust in clinic workflows. (aasm.org) This is not the academy’s first try. The earlier pilot program focused on sleep stage scoring from in-lab polysomnograms, while the new program expands to respiratory events, arousals, and limb movements using real-world records from accredited sleep laboratories. (aasm.org 1) (aasm.org 2) The certification also sets a regulatory floor. The academy says eligible software must already have Food and Drug Administration 510(k) clearance or be in the submission process, which means the program is aimed at commercial clinical products rather than research demos. (sleepworldmagazine.com) (aasm.org) That matters because sleep medicine has a bottleneck problem. A full-night study can generate hours of multi-channel data, and every minute has to be reviewed, which is why vendors have spent the last few years selling software that promises to cut scoring time without replacing the technologist or physician. (aasm.org) (ensodata.com) The second shift is happening after the test, not during it. Mount Sinai researchers reported a machine-learning model that estimates cardiovascular risk for people with obstructive sleep apnea and predicts how continuous positive airway pressure treatment could change that risk for a specific patient. (mountsinai.org) (nature.com) Obstructive sleep apnea is the form where throat muscles relax and the airway repeatedly collapses during sleep, like a straw pinching shut over and over. Mount Sinai says about 25 million people in the United States have the condition, and it is linked to higher risks of stroke and heart disease. (mountsinai.org) Continuous positive airway pressure, usually shortened to CPAP after the first mention, is the bedside machine that pushes air through a mask to keep that airway open. The Mount Sinai team says its model is the first to estimate whether CPAP is likely to lower or raise an individual patient’s cardiovascular risk instead of assuming the same effect for everyone. (mountsinai.org) (nature.com) The researchers built the model from the Sleep Apnea Cardiovascular Endpoints trial, a dataset with more than 2,600 participants across 89 sites in seven countries. In the paper in Communications Medicine, they reported that patients in the tertile predicted to benefit from CPAP had roughly 100-fold better event-free survival when randomized to CPAP, while the tertile predicted to be harmed had more than a 100-fold increase in major cardiovascular outcomes. (digitaljournal.com) (nature.com) Put those two developments together and the direction of travel is pretty clear. One part of sleep medicine is trying to certify artificial intelligence to do the scoring work inside the lab, while another part is using machine learning to sort patients by downstream risk after the diagnosis is already made. (aasm.org) (mountsinai.org)