Source: PR Newswire via Morningstar — 2026-07-16
Summary
HoneyNaps received U.S. FDA 510(k) clearance (K253390) for SOMNUM V3.0, AI software that analyzes overnight polysomnography (PSG) recordings and automatically detects and classifies obstructive, central, and mixed sleep apnea. The company reports a 97% agreement rate against human sleep technologist scoring.
Key Takeaways
- The clinically hard part isn't detecting "an apnea event happened" — it's classifying which subtype it was, because obstructive and central sleep apnea are treated completely differently, and getting the subtype wrong means prescribing the wrong therapy.
- 97% agreement with human-scored studies is the actual bar this clears — a direct comparison against the technologists who currently do this scoring by hand, not against some looser benchmark.
- 510(k) clearance is the FDA's substantial-equivalence pathway, faster than full premarket approval, signaling this builds on an already-recognized device category rather than introducing a brand-new one.
- A full-night PSG study is a dense multichannel biosignal recording (airflow, breathing effort, oxygen saturation, and more); this is a concrete example of a classifier operating on that kind of raw time-series data at clinical scale, not a chatbot wrapped around a symptom checklist.
Reel Script
Hook (~18s): Sleep apnea actually comes in three different flavors, and treating the wrong one doesn't just fail to help — for one of them, the standard treatment can make things worse. The FDA just cleared an AI that tells them apart automatically, agreeing with human sleep specialists 97% of the time.
Core Concept (~80s): A sleep study — a polysomnogram, or PSG — is an overnight recording of a bunch of signals at once: your breathing effort, your airflow, your blood oxygen, your brain activity. Someone has to comb through an entire night of that data and mark every apnea event. But detecting an apnea event is only half the job — you also have to classify it. Obstructive sleep apnea is a plumbing problem: your airway physically collapses, and a CPAP machine fixes it by splinting the airway open with air pressure. Central sleep apnea is a completely different problem — it's not the airway, it's the brain intermittently forgetting to send the "breathe" signal at all, and blowing air into an already-open airway doesn't fix that. Mixed apnea is both happening in the same patient. Get the classification wrong, and you can prescribe someone a CPAP machine for a problem CPAP was never going to solve.
Hands-On (~90s): The number here is 97% — that's SOMNUM V3.0's agreement rate against human-scored PSG studies, and the clearance itself is a 510(k), the FDA's faster "substantially equivalent to something already approved" pathway rather than a from-scratch approval. Conceptually, what's happening is a classifier running epoch by epoch across the whole night's multichannel recording — breathing effort, airflow, oxygen saturation — sorting each detected event into obstructive, central, or mixed, and handing a sleep physician an already-typed report to review and sign off on, instead of a technologist scoring hours of raw waveform by hand from scratch.
Takeaway (~25s): This is the unglamorous, actually-valuable end of medical AI — not a diagnosis-replacing chatbot, but a narrow classifier tackling a real scoring bottleneck, validated against a real clinical accuracy bar before it touched a patient. That's the template worth judging every other "AI in healthcare" headline against.