AI Gets the First Look at Sleep Studies: SOMNUM Wins FDA Clearance to Identify Three Types of Sleep Apnea

An overnight polysomnography study generates vast amounts of data, including brain waves, respiratory airflow, blood oxygen levels, and chest and abdominal movements. The truly time-consuming work involves having trained professionals mark respiratory events segment by segment. The U.S. Food and Drug Administration (FDA) recently cleared HoneyNaps’ SOMNUM V3.0 for marketing, allowing artificial intelligence software to perform the initial detection of apnea and hypopnea events and further distinguish among different types of sleep apnea.

The FDA database shows that the product was submitted through the traditional 510(k) pathway under K253390, received on September 30, 2025, and determined to be “substantially equivalent” on June 27, 2026. It is classified as a Class II medical device and categorized as automated event-detection software for polysomnography used with electroencephalography. It was reviewed through the neurological medical device review system without third-party review.

SOMNUM V3.0 analyzes multichannel physiological signals and, after identifying apnea and hypopnea, can further classify apnea as obstructive, central, or mixed. This distinction is clinically meaningful: obstructive apnea typically involves upper-airway collapse during sleep, while central apnea is associated with a pause in respiratory drive, and mixed apnea has features of both. Accurate classification affects physicians’ assessment of the cause and subsequent management.

HoneyNaps said that, in the validation results submitted for regulatory review, the overall percentage agreement for every type of respiratory event exceeded 97%. Figures reported by the South Korean newspaper Seoul Shinmun ranged from 97% to more than 99%, while the sleep medicine publication Sleep Review also cited overall agreement exceeding 97%. However, currently available public information does not specify the validation dataset’s sample size, participant composition, disease severity, reader configuration, or coverage of equipment used across different hospitals. Nor does it fully present the sensitivity and specificity for each event type.

These gaps mean that the “agreement rate” must be interpreted cautiously. When a particular type of event accounts for a very high or very low proportion of the data, a single overall percentage may not reflect the risk of missing rare events. A high level of agreement between the software and the reference interpretation also does not mean that it has been proven to shorten reporting time, reduce manual corrections, or improve patients’ diagnostic and treatment outcomes.

This clearance extends the capabilities of the existing SOMNUM V1.1.2, advancing from assisting with sleep-study analysis to more detailed classification of respiratory events. The 510(k) pathway confirms that a product is substantially equivalent to a legally marketed predicate device; it does not independently endorse the algorithm’s clinical benefits across all populations, hospitals, and workflows. In actual use, professionals must still verify the automated annotations against the raw signals.

HoneyNaps also said it is developing quantitative metrics including hypoxic burden, arousal burden, and ventilatory burden, and plans to submit them for regulatory review in subsequent versions. For these metrics to move from research tools into clinical decision-making, their applicable populations and reproducibility across devices must be clearly defined, along with whether they can provide additional, actionable information compared with the conventional apnea-hypopnea index.