Skip to main content
The SDK helps you show that your app changes the body, not only how people say they feel. This page collects what has worked in the NEUROFIT app and the validation study.

Pre/post around a session

The cleanest single-session design is a reading immediately before and immediately after your intervention (a breathing exercise, a meditation, a coaching session).
  • Same posture, same place. Seated, feet on the floor, phone resting on a table or the lap, both times.
  • Same finger. The guidance handles the pressure, but tell users to use the same fingertip.
  • Let the body settle first. A minute of quiet sitting before the pre reading avoids measuring the walk to the chair.
  • Expect some retries. A reading that fails the quality check returns no numbers and restarts by itself (how retries work). Every completed result counts as a data point.
  • Store VitalsResult whole (Result JSON), so you can audit and re-analyse later.
For a trend, the reading conditions matter more than the count.
  • Morning, before caffeine, seated. A consistent time of day removes most day-to-day noise.
  • Weekly averages beat daily values. HRV (RMSSD) varies day to day in healthy people, so a 7-day rolling mean is a steadier signal for a trend line.
  • Record context your app already knows (sleep, training, illness) alongside readings, so a dip has an explanation rather than becoming a worry.

Comparing a person with themselves

Absolute HRV differs widely between people (age, fitness, genetics), so a change from a person’s own baseline is meaningful where a population table usually is not.
  • Build a baseline from several readings under the same conditions each time: the same time of day and posture, seated and still, before caffeine or exercise. A handful over a week or two makes a usable baseline.
  • Compare averages, not single readings. One reading can move for ordinary reasons.
  • Compare HRV (RMSSD) on a log scale. HRV changes are proportional: a rise from 20 to 25 ms is a similar shift to a rise from 40 to 50 ms. Average the natural logs of the RMSSD values and compare the two averages. The difference converts back to a percentage change. Heart rate and breathing rate can be compared as they are.
  • Judge a change against that person’s own variation. Treat a change as meaningful only when it is larger than how much their baseline readings usually vary from day to day.
  • Use completed results only. Every result passed the quality check, so there is nothing to filter for heart rate or HRV. For breathing rate, leave out readings whose breathingRateConfidence is low or withheld.
For example, five morning readings in a baseline week and five in a later week averaging 42 ms and 48 ms on the log scale are a rise of about 14%. That is meaningful if the person’s readings usually differ by about 10% from day to day, and not yet if they differ by about 25%. For a group claim, pre-register the metric (HRV (RMSSD) is the usual primary), the reading protocol and the analysis. Report how often a finished reading failed the quality check along with the results: the number of cancelled(qualityCheckFailed) events divided by the number of measurementComplete events.

Measuring reliability in a pilot

To compare your pilot with NEUROFIT’s figure, measure it the same way:
  • Count sittings, not screen opens. Treat a user’s sessions less than 30 minutes apart as one sitting, keep the sittings in which a reading started, and report the share that ended with a completed result, retries included.
  • Count a try once the reading has run 10 seconds. A reading the SDK cancels in its first few seconds restarts by itself and is not a separate attempt.
  • Test with real users, on their own phones. Include first-time users, since people improve quickly after their first few readings, and include the placement guidance in your UI.
  • Use enough people. Results vary a lot between individuals, so test with at least 30 people and report by person as well as by reading.
To find causes, these slices pay off, in this order:
  1. Quality-check failures (cancelled(qualityCheckFailed) per measurementComplete) by device model. One model failing across many users points at the device.
  2. Cancel reasons by device model and by user. Frequent poorSignal for one user usually means a cold finger, hard pressure or a hand-held phone, and for one model a thick case or weak optics.
  3. How often positioning shows lowFrameRate, by device model. It usually tracks Low Power Mode or Battery Saver.
  4. Positioning time, from start() to started. A long wait points at the placement, the guidance copy, or the lighting on a device without a torch.

Wording for users

Calm and specific wording keeps people engaged with a metric that moves around.
  • Prefer “your HRV (RMSSD) was 42 ms this morning, in your usual range” to “your HRV is low”.
  • Show change over a week or a month before change over a day.
  • Explain a failed quality check as an issue with the camera or the signal, not a health signal.
  • Never present the metrics as diagnosis. The SDK is not a medical device.

Why a finger scan, not a face scan

RMSSD is the HRV number WHOOP, Oura, Polar and Fitbit report, and the hardest one to measure with a camera. It is built from tiny beat-to-beat changes, so any error in timing a beat adds straight to it and makes HRV read high. That is a physics problem for face scans. On a face, the pulse is a faint colour change in reflected room light, easily swamped by shifts in lighting and small head movements. A fingertip pressed over the flash is lit by a bright, steady light right against the skin, giving a pulse about 6 times larger, or over 30 times the signal power. The SDK’s average HRV (RMSSD) error is 3.4 ms, while the best face-video result in a 2023 benchmark was 10.5 ms, with seated, still participants.
Facial pulse amplitude: Wang et al., Biomedical Optics Express, 2017 (15 participants, medians read from their Figure 2a), compared with the 30 NEUROFIT study participants using the same method. Face-video RMSSD: Bioengineering, 2023 (UBFC-rPPG dataset). Different studies, cameras and references, so not a head-to-head comparison.
A finger scan also helps outcomes work: the reading does not depend on room light, and the torch-lit fingertip is what the validation study measured.