What the EU AI Act Means for the Sleep Apnea Algorithms Inside Your Watch and Home Sleep Test

EU AI Act sleep apnea detection in plain language

EU AI Act sleep apnea detection rules govern the software that turns your overnight signals into a number. They do not govern the strap on your wrist or the sensor on your finger. If your home test was read by an automated scoring algorithm, that code is the regulated object.

An estimated 936 million adults aged 30-69 worldwide have mild-to-severe obstructive sleep apnea, including 425 million with moderate-to-severe disease, according to Benjafield and colleagues in The Lancet Respiratory Medicine (2019). For many people today, the first hint of that possibility comes from an algorithm rather than a clinician.

Your device's status under the Medical Device Regulation (EU) 2017/745 decides its status under the AI Act. MDR Class IIa, IIb and III software normally becomes a high-risk AI system. Self-certified Class I software with no notified body falls outside that category, unless it also appears in Annex III.

Why EU AI Act sleep apnea detection carries no right to an explanation

The Act's headline patient right does not reach your sleep algorithm. Article 86 grants a right to "clear and meaningful explanations of the role of the AI system in the decision-making procedure". It applies only to high-risk systems listed in Annex III.

Software built into a CE-marked product becomes high-risk through a different door: Article 6(1) and Annex I, the product-safety route. Annex III instead covers standalone high-risk AI, meaning systems that are not embedded in a separately regulated product. Because a CE-marked apnea algorithm sits in Annex I, Article 86 never touches it.

Most healthcare commentary implies the opposite, but you cannot demand a statutory explanation of the model that flagged you. Three routes remain: the instructions for use MDR already requires, Article 50, which requires that users be adequately informed they are interacting with an AI-based product, and GDPR.

Your sleep and oximetry readings are special category health data. GDPR Article 22 covers solely automated decision-making and can supply an explanation where the AI Act does not. Article 86(3) itself steps aside where another Union law already provides the right.

The deadline moved, so what governs your watch today

No part of the high-risk regime binds your sleep algorithm right now. The Digital Omnibus on AI reached provisional political agreement on 6-7 May 2026 and was confirmed by Council on 13 May 2026. It replaced the 2 August 2026 and 2 August 2027 dates that most published guidance still quotes.

Under the revised schedule, Annex III standalone high-risk AI applies six months after the Commission confirms that support measures are available, with a long-stop of 2 December 2027. Annex I product-embedded AI, including MDR and IVDR medical devices, applies twelve months after that confirmation, with a long-stop of 2 August 2028.

Meanwhile, AI-enabled medical devices in Europe are certified under MDR and IVDR alone. There is no separate AI Act conformity assessment today. When the rules arrive, sectoral conformity assessment takes precedence, so transparency and oversight duties fold into the existing notified body review.

Which regulatory box your device actually lands in

Sleep products can share almost identical sensors and still land in different regulatory boxes. Intended purpose decides, not the hardware.

MDR Recital 19 states that software intended for lifestyle and well-being purposes is not considered a medical device. Guidance document MDCG 2019-11 applies the same logic to the qualification and classification of software. A nightly sleep-score app with a wellness purpose never reaches Annex VIII Rule 11, the software classification rule, so it never becomes high-risk AI.

The Act will give you less protection on the sleep tracker you open every morning than on the one feature its maker chose to certify. Wellness trackers infer breathing from photoplethysmography (PPG), an optical blood-flow sensor, and from accelerometer-based wrist actigraphy, with no regulator reviewing the maths.

A real worked example

A consumer smartwatch sleep apnea feature obtained a CE mark for the European Economic Area, according to the manufacturer's official announcement of 5 June 2025, and expanded to 34 European markets. It is intended to detect signs of moderate-to-severe OSA in adults aged 22 and older across a two-night monitoring period.

The same announcement is explicit about limits. The feature is not intended for people already diagnosed with sleep apnea, and the manufacturer states it "should not be used to replace traditional methods of diagnosis and treatment by a qualified clinician".

What EU AI Act sleep apnea detection rules say about error rates

Article 15 will require high-risk AI to declare appropriate levels of accuracy, robustness and cybersecurity. Article 13 covers transparency and the provision of information to deployers, through instructions for use that state those metrics and known limitations. Neither is enforceable against your device yet, so for now you read the published evidence yourself.

Those figures come from a 2024 systematic review and meta-analysis of wearable AI indexed in PubMed Central (PMC). Detection accuracy of 0.869 (95% CI 0.81-0.92) pooled 13 studies and 2,015 participants. Sensitivity reached 0.938 (95% CI 0.89-0.97) while specificity stayed at 0.752 (95% CI 0.63-0.86).

High sensitivity means a false negative is relatively uncommon, so the technology rarely misses a real case. Mediocre specificity means it often raises an alarm in someone who is fine. That is the signature of screening, not diagnosis.

Severity is the weaker skill. Across 9 studies and 1,661 participants in the same 2024 review, pooled accuracy for assessing severity was only 0.651 (95% CI 0.54-0.75). Severity decides whether you are told mild, moderate or severe, and therefore what treatment you are offered.

Why the estimated AHI reads low

A 2025 real-world validation study indexed in PubMed Central compared a consumer smartwatch sleep apnea algorithm with polysomnography in 53 people. Watch-estimated AHI correlated with AI-scored PSG at rho 0.88 (ICC 0.87) and with physician-scored PSG at rho 0.85 (ICC 0.82). For moderate-to-severe disease at an apnea-hypopnea index (AHI) of 15 or above, sensitivity was 92.3%, specificity 92.6%, accuracy 92.5% and AUC 0.962.

The same 2025 study names three structural reasons the estimated AHI, or eAHI, runs low. Events are analysed in one-minute windows, so several brief apneas inside one minute count once. The eAHI has a theoretical ceiling of 60 events per hour. And the algorithm divides by total recording time rather than true sleep time.

Two further limits matter clinically. The algorithm cannot differentiate central from obstructive apnea, and the authors note it may underestimate in people with increased skin pigmentation. Our closer look at how wrist-based apnea detection performs against a lab study unpacks what that means night to night.

Who is accountable when nobody in a clinic sees the result

Accountability splits between the provider, meaning the manufacturer, and the deployer, meaning the organisation using the system professionally. When you buy the watch yourself, there is no deployer, so the chain stops at you.

Article 26 deployer obligations all assume an institution: human oversight with the power to override an AI prediction and safeguards against automation bias (26(2), 26(5)), input-data quality verification (26(4)), and retention of automatic logs for at least six months (26(6)). Article 12 record-keeping and Article 14 human oversight place matching design duties on the manufacturer.

Article 4 AI literacy is different, because it already applies. Since 2 February 2025, EU sleep clinics and hospitals must ensure staff understand a system's capabilities, its risks, how to interpret its output and how to spot a malfunction, across the whole lifecycle including feature updates. Your clinic is bound today; your wrist is not.

Post-market surveillance under MDR fills part of the gap for CE-marked products. Manufacturers must monitor real-world performance and report serious incidents through vigilance channels, a duty we cover in our guide to EU MDR post-market surveillance rules for anti-snoring devices. For a wellness app, that safety net is absent.

The training data you are not allowed to inspect

Article 10 on data and data governance will require that training, validation and testing datasets be relevant, representative and examined for bias. Sleep AI starts from a difficult position. In the 2024 wearable AI review, 36 of 38 studies (95%) were built on closed-source data, against only 2 (5%) using open-source data.

The same review reports that 24 of 38 studies (63%) tested commercial devices and 11 of 38 (29%) used wrist-worn hardware. Deep learning outperformed classical machine learning, with accuracy of 0.899 against 0.831 (P=.02). Better models trained on unpublished data still leave you unable to check who the model learned from.

A second problem sits underneath, in the labels themselves. In the AASM Inter-scorer Reliability Program reported in the Journal of Clinical Sleep Medicine (2013), human sleep-stage agreement averaged 82.6% across more than 2,500 scorers and over 3,200,000 scoring decisions on 1,800 epochs. Agreement fell to 67.4% for stage N3 and 63.0% for stage N1.

Manual scoring is the ground truth these algorithms learn from, and its inter-scorer reliability is imperfect. An algorithm trained on human labels inherits that variability rather than removing it. Treat any single night's number as an estimate.

National rules the AI Act does not replace

A CE mark lets a device be sold across the EU. It does not get the device paid for, and it does not centralise complaints. Both stay national.

In Germany, the DiGA fast-track run by BfArM accepts only Class I and Class IIa applications, with 68 listed DiGAs by late 2024. France adopted its analogous early-reimbursement route, PECAN or prise en charge anticipée numérique, in 2023. Neither route is automatic, and neither is granted by the AI Act.

Market surveillance under the AI Act also runs through national competent authorities, not one EU regulator. Your complaint route follows your country of residence, not where the software was written.

What to do when an algorithm flags you

An algorithmic estimate is a screening signal, not a diagnosis. The CE-marked consumer feature described above states plainly that it should not replace diagnosis and treatment by a qualified clinician.

The transparency the AI Act asks of algorithms should apply to device makers too. Back2Sleep is a CE-certified Class I soft silicone intranasal stent that keeps the nasal airway open during sleep, with no electricity, no noise and no tubing.

It is intended for snoring and for clinician-confirmed mild-to-moderate obstructive sleep apnea, where the obstruction is nasal or nasopharyngeal. It is not a treatment for severe OSA or central apnea, and it is never a reason to skip the diagnostic step this article argues for.

Frequently Asked Questions