How ONDA Science Checks and Grades Research

Why does method matter for heart rate variability?

Heart rate variability (HRV) is easy to record and hard to read. A research guideline notes that the methods, analyses and interpretations used in HRV studies vary widely, and that debates about its reliability remain [S1]. A methods review makes the same point from psychophysiology: the ease of measuring HRV can hide how easily its findings are misread [S2]. ONDA Science is built around that problem. This page sets out the rules every page follows, so that a reader can see why a sentence is worded the way it is.

How is each source checked?

• Identity. Before a journal source is used on a science page, its DOI is looked up in Crossref or its PMID in PubMed, and its title, authors and year are compared with the reference. A source that does not match is removed.
• Site-wide citation check. Across the whole site — science pages, articles, reviews and the glossary — every PMID, DOI and PMC identifier must be in a citation register filled from PubMed and Crossref. The site build fails if a cited identifier is missing from the register, does not exist, or is known to point to a different paper. Where an author and year are written next to an identifier, the build also fails if that author is not among the paper's authors or the year is more than one year off. The build does not compare titles; that is part of the check before use.
• Text. Abstracts and, where open, full texts are read through Europe PMC or the publisher. A claim that rests on a part of the paper that could not be read is not made.
• Quote. Each claim in a page's evidence table carries a quote copied word for word from the source, and the quote is compared with the abstract or full text before the page is published. If the quote states a neighbouring idea rather than the claim, the claim is rewritten or removed.
• Numbers. Every number on a science page comes from one list of approved facts; only years, dates and official regulatory designations are written directly. Each fact records its source and what it applies to — population, method, time of day — and a new number waits for the editor's approval before it can be published.
• Duplicates and replacements. In the register of candidate sources, the same study found twice is merged into one entry, and when a newer review replaces an earlier one, the earlier entry is kept and marked as replaced rather than silently dropped.

How is the strength of a claim graded?

Grades are attached to a specific claim, not to a product or a topic. Each claim in a page's evidence table carries exactly one evidence class:

• Established — consistent meta-analytic evidence, or long-settled textbook physiology (anatomy and basic mechanisms, not the effects of interventions). A single small trial is never labelled established, and the effects of breathing, cold, practices or devices are never established on textbook grounds alone.
• Guideline / expert consensus — a recommendation or methodological standard from a guideline. This is the strength of agreement among experts, not of outcome data.
• Context-dependent — supported, but only in a stated population, condition or measurement setting.
• Emerging — early or small studies point one way; the evidence is not yet consistent.
• Debated — credible sources disagree.
• Unknown — not studied enough to say.

Some question pages also sum up the evidence in plain words: shown, mixed, not shown and no data. Not shown means the studies looked for an effect and did not find it; it is not proof that the effect does not exist. No data means no study measured it, which is a gap, not a refutation. An association is never written as a cause.

Which studies count, and for what?

• Animal studies. A finding in animals is labelled as an animal finding and is not presented as evidence about people.
• Narrative reviews. A review without a systematic search is labelled as a narrative review in the evidence table, and its summary is never presented as a pooled result. On its own it can support an established claim only for long-settled textbook physiology, and only when it is a classic reference review.
• Single studies. A result from one study is described as one study, with its population and design, and is never labelled established.
• Maker-linked studies. A study run or funded by the maker of a device is flagged as such, and it does not move a claim into a stronger evidence class.
• Abstract and full text. When the abstract and the full text of a paper disagree, the full text is used.

Why is a group result not a personal verdict?

Almost every study reports averages across groups of people. A pooled effect says what tends to happen in a population like the one studied; it does not say what will happen to one person, and it does not travel to a different age group, sport or health condition. Pages name the population every time and keep results from, say, older adults apart from claims about young people.

How are device and regulatory claims handled?

• Official sources only. A device's regulatory status comes from the regulator's own database, not from the maker's website. Registering a device with a regulator is not the same as clearance or approval.
• Device documentation. Manufacturer and regulator documents are cited only for what a device records or what a clearance covers, never for a health or efficacy claim.
• Validation studies. When a study compares named watches with a reference, the result is reported as a finding about those models in that setting, not as a recommendation.

How do reviews use this evidence?

Device reviews on ONDA apply the same evidence rules, with a scoring layer of their own. Studies run or funded by the maker do not count as independent validation. A maker's regulatory claims are checked against the product's status in official databases: registration or clearance adds nothing to a score, and misrepresenting it costs points. When a maker's own materials claim clinical proof or studies that do not exist for that product, the score is lowered and the review quotes the maker's words. The full scoring rules are on the review methodology page.

How are funding and conflicts of interest reported?

Funding and declared conflicts of interest are read from the paper and noted with the source. Company funding or a maker's involvement is stated in the text when it bears on the claim. When the statement could not be read — for example, behind a paywall — the source note says it could not be checked. A missing statement is never presented as the absence of a conflict.

What are the limits of HRV measures?

HRV indices are highly sensitive to context: recording length, signal quality, breathing, posture, time of day, medication and the analytical method all change the numbers [S3]. The input signal — an electrocardiogram or a wrist or finger pulse sensor — and the place of recording matter too [S1]. ONDA Science therefore compares results only across similar recording conditions and says which conditions each study used.

Two cautions run through every page. HRV is not a specific marker of sympathetic activity or of the balance between the two branches of the autonomic nervous system [S1]. Single frequency-domain indices, the LF/HF ratio above all, are easily overinterpreted and are reported only as what the study measured, with that caveat attached [S3].

What it does not tell you

• A checked source is not a proven claim. Verification confirms that a paper exists and says what is quoted; the strength of the claim is set by its evidence class.
• Not shown is not proof of absence.
• These rules do not make ONDA a medical service. They govern how the science section is written.

In ONDA

The same rules apply to what pages say about ONDA itself: ONDA builds a personal baseline from HRV, resting heart rate and breathing rate values in Apple Health and describes how your latest values compare with your own range [S4]. It is not a medical device and does not diagnose [S4], and claims about the app are checked against how the app actually works. How ONDA's own measurements work is set out on how ONDA measures and interprets your signals.