Sleep + Digital Health

The landscape

More people now receive nightly quantitative feedback about their sleep than have ever seen a sleep specialist. Rings, watches, straps, mattress sensors, bedside radar, and hundreds of apps produce scores, stage breakdowns, and readiness metrics.

This is a genuine opportunity. Longitudinal, ecologically valid data on sleep timing and regularity — collected for months in a person's own bed — is something the clinical field never had at scale. It is also a genuine risk, because the interpretive layer sitting on top of that data is frequently unvalidated and delivered with more confidence than the underlying signal supports.

What wearables actually measure

No consumer device measures sleep directly. Polysomnography stages sleep from brain activity (EEG), eye movement, and muscle tone. Wearables infer sleep from proxies:

• Accelerometry — movement, the historical basis of actigraphy and still the primary sleep/wake signal.
• Photoplethysmography — optical heart rate, from which heart rate variability and respiratory rate are derived.
• Skin temperature — used for trend and cycle detection.
• Blood oxygen estimates — present on many devices, but not validated for apnea diagnosis.

Stage estimates are produced by proprietary algorithms combining these inputs. They are inferences, they differ between manufacturers, and they change silently with firmware updates.

Where the validation evidence stands

Head-to-head studies against polysomnography converge on a consistent picture. Modern multi-sensor devices detect sleep versus wake well, with high sensitivity for sleep. Total sleep time is typically accurate within tolerable error for tracking purposes. Wake after sleep onset is systematically underestimated — devices tend to score quiet wakefulness as sleep, which matters specifically for people with insomnia.

Stage classification is the weakest area. Agreement with polysomnography for individual stages is moderate at best, with deep and REM sleep frequently misclassified. Nightly stage percentages should not be treated as measurements.

Two further cautions. Accuracy is generally established in healthy sleepers and degrades in exactly the clinical populations where precision would matter most. And algorithms are proprietary and revised without notice, so a change in your numbers may reflect a software release rather than your sleep. Detailed accuracy review.

Using consumer sleep data well

• Read weeks, not nights. Single-night values carry too much measurement noise to act on.
• Prioritize timing and regularity. Bed and wake times and their consistency are the most accurate and most clinically actionable outputs.
• Ignore the composite score as a verdict. It is a proprietary weighting, not a measurement, and it is the main driver of tracker-related anxiety.
• Treat stages as texture, not fact.
• Trust how you feel. If you feel rested and the app disagrees, the app is the less reliable instrument.
• Bring the data to clinical care. Multi-month timing data is genuinely useful to a behavioral sleep clinician.

Orthosomnia

Orthosomnia describes a pattern first named in the clinical literature in 2017: people whose sleep worsens because of a preoccupation with achieving perfect tracker data. They extend time in bed to raise their numbers, become anxious about scores, and interpret normal variation as pathology.

The mechanism is well understood in behavioral sleep medicine. Insomnia is maintained by sleep-related anxiety and monitoring; a device that supplies a nightly score to worry about reinforces both. For a patient who checks their score before getting out of bed, a period of not tracking is often a legitimate and effective clinical intervention.

Digital CBT-I

Digital delivery of CBT-I is the strongest evidence base in consumer sleep technology. Meta-analyses show clinically meaningful improvements in insomnia severity, sleep onset latency, and wake after sleep onset, with effects maintained at follow-up. It somewhat underperforms in-person treatment on average, and it dramatically outperforms the realistic alternative, which for most people is no treatment at all.

The important distinction for consumers and buyers alike: many products marketed as "sleep apps" contain relaxation content and education but no actual CBT-I. Genuine dCBT-I includes sleep diary entry, individualized sleep-window prescription that adapts to the data, stimulus control instruction, and cognitive restructuring — and published trial evidence for the specific product. What CBT-I involves.

Building clinically credible sleep technology

Most digital sleep products fail on the same points, and all of them are addressable at design time:

1. Match claims to evidence. Say what the sensor supports. Overclaiming is both a regulatory exposure and the fastest way to lose clinical credibility.
2. Design the feedback for behavior, not engagement. A score that produces anxiety generates sessions and worsens outcomes.
3. Make the intervention adaptive. Fixed content sequences are education; responsive protocols are treatment.
4. Plan for non-adherence explicitly — it is the default state, not the exception.
5. Build clinical escalation paths. Screen for apnea risk and severe mood symptoms and route those users to care.
6. Measure real endpoints. Insomnia Severity Index, sleep diary metrics, retention to outcome — not app opens.
7. Involve behavioral sleep expertise before launch, not after the first outcomes review.

This is the substance of our work with digital health companies, device makers, and employers. See consulting services.

Evaluating a sleep product

• Is there peer-reviewed validation of this product, not the category?
• Were validation studies conducted in relevant populations, including poor sleepers?
• Does it deliver an evidence-based intervention, or content and relaxation audio?
• Does the intervention adapt to the individual's data?
• Are outcomes reported as clinical measures or as engagement?
• Is there qualified clinical involvement in its design?
• Does it screen for conditions requiring medical evaluation?
• How is the data handled, and who has access to it?