How AI is changing the way
we diagnose sleep problems

Why diagnosing sleep problems has always been hard

The gold standard for diagnosing sleep disorders has always been polysomnography — an overnight sleep study in a lab where you're wired up with sensors measuring brain waves, eye movements, muscle activity, heart rate, and breathing. It's accurate, but it's expensive, it requires specialist facilities, and it's deeply uncomfortable. Many people who need a diagnosis never get one because of the wait times and cost involved.

The result: obstructive sleep apnea — which affects an estimated one billion people worldwide — is dramatically underdiagnosed. Most people with it don't know they have it. This matters because, as we covered in our article on sleep and high blood pressure, untreated sleep apnea is one of the most common drivers of difficult-to-treat hypertension.

What AI is doing differently

Since 2020, the AI related sleep research has exploded. The shift has moved from relatively simple machine learning models to deep learning architectures that can process raw sensor signals the way a neurologist reads an EEG.

The headline result: AI models for obstructive sleep apnea detection now achieve between 85% and 99% accuracy in controlled settings. Some outlier models hit 99.9%. Deep learning approaches for sleep staging — working out how much time you spend in light sleep, deep sleep, and REM — now match human expert performance.

What this means for your Oura ring or Apple Watch

This isn't only happening in research labs. The same AI techniques are already embedded in consumer wearables. When your Oura Ring tells you how much deep sleep you got last night, it's using a machine learning model trained on thousands of sleep studies to classify your movement and heart rate data into sleep stages. When Apple Watch detects an irregular heart rhythm, it's a neural network doing the classification.

These devices aren't yet reliable enough for clinical diagnosis — but they're becoming genuinely useful for spotting patterns over time. Seeing consistently low deep sleep for weeks, or a spike in your resting heart rate at 2am every night, are the kinds of signals worth taking to a doctor — and AI is what makes that kind of continuous monitoring possible outside a lab.

The real limitations — and why they matter

The main problems are three: generalisability, interpretability, and bias.

Most AI sleep models are trained on data from specific populations — typically middle-aged Western adults attending specialist clinics. When those models are applied to people from different age groups, ethnicities, or health backgrounds, performance often drops. A model that detects sleep apnea with 97% accuracy in a US clinical trial may perform much worse on a population with different body shapes, different comorbidities, or simply less well-represented in the original training data.

Interpretability is the other issue. Deep learning models are notoriously difficult to explain — they can tell you that you probably have sleep apnea, but they often can't tell you exactly which signals led them to that conclusion in the way a human clinician can. That matters for both clinical trust and for catching errors.

And algorithmic bias — the risk that AI systems work better for some groups than others — is a real concern in healthcare AI more broadly. If the patients who benefit most from AI-assisted diagnosis are already the ones with the best access to care, the technology risks widening rather than closing health inequalities.

Where this is heading

The direction is clear. Sleep studies that currently require a hospital overnight — with a technician manually scoring hours of data the next day — will increasingly be replaced by home-based tests analysed automatically. AI won't replace sleep specialists, but it will make their work faster, more consistent, and accessible to many more people.

For anyone dealing with sleep problems now, the practical implication is this: if you snore, wake unrefreshed, or feel excessively tired during the day, it's worth asking your doctor about a sleep study — and it's increasingly likely that study won't require a hospital bed. Home sleep apnea tests using AI scoring are already available in several countries, and the technology is improving quickly.