Detecting Sleep With a Phone Instead of a Wearable

How does phone sleep tracking work without a wearable?

MaxDex runs three independent detectors over the night — screen-interaction gaps, operating-system sleep classifications, and accelerometer stillness windows — then selects the best-supported candidate segment and scores it for efficiency and disturbance.

Running three detectors in parallel rather than one is the core design choice. Each fails differently: interaction gaps break if you leave the phone untouched during a long film, classification events are unavailable on some devices, and stillness windows misread a phone left on a desk. Requiring agreement, or picking the best-supported candidate, is more robust than trusting any single signal.

Adjacent fragments separated by less than 90 minutes are merged into one episode, so a 3am bathroom trip does not split a night into two short, low-confidence records.

What does "best effort" mean in the algorithm?

Rather than discarding a night that looks imperfect, the engine keeps the episode and scores it — efficiency, movement index, disturbance score, and estimated awake minutes — so an interrupted night is recorded as an interrupted night instead of vanishing.

This is the opposite of the common approach, which is to throw away readings below a confidence threshold. Discarding produces a tidier chart and a worse model: the nights most worth knowing about are exactly the broken ones, and a gap in the record reads to the sleep-debt calculation as though the night never happened.

Each episode carries a classification derived from its confidence and disturbance scores, so the planner and the AI can treat a fragmented six hours differently from a clean six hours.

How does MaxDex handle an uncertain night?

A duration-confidence weight blends the reading toward your personal sleep need. Effective hours = confidence × measured hours + (1 − confidence) × your sleep need, so a barely-detected night regresses toward "probably normal for you" rather than swinging the score.

Without this, a single noisy detection would crater a sleep score and a week of debt figures. With it, the model says the honest thing: we are not sure what happened, so we will not pretend to be.

Manually entered sleep carries no confidence value and is trusted at face value. The same duration-confidence function is shared by the Android detector and the iOS Apple Health path, so a short or fragmented night degrades identically on either platform rather than one being quietly more optimistic.

Does it work for shift workers and day sleepers?

Yes. Time of day is explicitly not a gate on what counts as sleep. A main sleep episode starting at 9am is treated as a first-class night, not discarded as an anomalously timed nap.

Most consumer sleep tracking bakes in an assumption that sleep happens at night, then behaves strangely for the roughly one in five workers whose schedules do not. Typical sleep timing is used as a weak prior when disambiguating between candidate segments — never as a filter that rejects them.

Naps are detected separately, with a lower minimum duration, and contribute a bounded pressure relief rather than overwriting the main sleep record for the day.

How accurate is phone-based sleep tracking really?

Less accurate than a validated wrist device for duration, and not comparable at all for sleep staging. Phone detection infers sleep from behaviour and stillness; it cannot measure heart rate, heart-rate variability or blood oxygen, and MaxDex has not run a validation study against polysomnography.

The honest framing is that this is a good-enough estimate of when you slept and roughly how well, built for planning your day rather than diagnosing anything. If you need sleep-stage data or clinical-grade measurement, a validated wearable or a sleep study is the right tool, and we would rather say so than lose your trust later.

Reviews of consumer sleep technology have consistently found that even wrist-worn devices perform well on total sleep time while performing poorly on stage classification. A phone is a step further removed again. What a phone genuinely offers is that it is already on your nightstand, charged, every night — and consistency of measurement has real value that accuracy comparisons tend to miss.

How is the sleep score itself computed?

score = (effective hours ÷ your sleep need × 100) − debt penalty + consistency bonus − oversleep penalty, clamped to 0-100. The debt penalty caps at 30 points, the consistency term is (consistency − 0.6) × 20, and oversleep beyond 1.5 hours over need costs up to 12.

The oversleep penalty is the part people query, so it is worth stating the reasoning: the relationship between sleep duration and metabolic and mortality outcomes is an inverted U, not a straight line. Large meta-analyses have found elevated risk at both short and long durations. The penalty is kept deliberately gentle — capped at 12 points and only engaging beyond 1.5 hours over your need — so one genuinely restorative long night is not punished.

Frequently asked questions