The short answer
Using thirty HRV features, linking one window to another from the same recording among 202 candidates achieved 30.7% at rank one and 55.4% within the top five. Random selection would be about 0.5% and 2.5%. This unreviewed analysis used the same day, device and posture in a closed candidate set. It does not establish identity verification across days or devices, or authentication that can reject an unknown person.
Key points
- Across 202 people in a public dataset, thirty HRV features linked a person to their own later window
- Rank-one 30.7% (chance 0.5%); top-five 55.4% (chance 2.5%)
- Permutation test against a fixed random-label null: p=0.00005
- But this is within a single session; stability across days, devices or sleep and wake was not shown
- It is not an Apple Watch evaluation, and not a measure of authentication performance
Why we tested it
HRV summarizes beat-interval variation in several ways. It is influenced by posture, breathing, age and measurement conditions, but it may also carry repeated person-level patterns or features shared by one recording.
For a health app, the degree to which two records can be linked matters when assessing the risks of storage, export and sharing.
How it was measured
Use public data
From PhysioNet's Autonomic Aging dataset, we used the 211 subjects already frozen as the held-out split. 202 of them met the conditions.
Take two windows per person
Two consecutive, non-overlapping 300-beat windows. The first became the enrolment template, the second the query. The query window was never part of its own enrolment.
Compare on thirty HRV features
SDNN, RMSSD, frequency-domain and nonlinear measures — thirty in all — standardised with training-frozen parameters and compared against all 202 people by cosine similarity.
Compare against chance
Picking at random from 202 gives 0.5% at rank one. That is the baseline the measured accuracy is set against.
Results
- Rank one: 62/202 = 30.7% (95% CI 24.7–37.4%; chance 0.495%)
- Top five: 112/202 = 55.4% (95% CI 48.6–62.1%; chance 2.48%)
- Permutation test (20,000 replicates): p < 0.0001
- Swapping which window enrols and which queries: rank one 34.2%, top five 55.9%
What this result does not mean
Read as numbers alone, this looks like "HRV can verify identity". It cannot. The test has limits worth stating plainly.
- Enrolment and query come from the same recording session — same day, same device, same posture
- The candidate set is closed at 202; the ability to reject an unknown person was not measured
- Stability across days, devices, or sleep and wake was not demonstrated at all
- Session, sensor or preprocessing artefacts specific to a recording may contribute to the linkage
- This is not an Apple Watch evaluation. The signal used was clinical ECG
How this shows up in Feelmo
Because of this result, Feelmo keeps HRV-derived data and scores on the device as a rule, with no automatic upload to Provider servers. Only when you start an export or a share does the content you chose go where you send it.
This analysis does not decide the legal status of the data, but it does show that a same-session record can retain clues that help link two windows, so we use a cautious design.
The data and how to reproduce it
The aggregate results, the uncertainty intervals, the sensitivity analysis, SHA-256 hashes of every input and code file, and the limits listed above are published as JSON. The script that produced it is included.
Per-subject features are not published — this research is itself the reason. A researcher with an independently generated, compatible feature artefact can reproduce the same calculation with the included script.