The intelia.health methodology

122 criteria are computable out of 133 active ones — eleven target markets the engine does not evaluate today. Based exclusively on public and verifiable sources.

The intelia.health observatory is independent of the vendors it documents. No profile is bought: the scope is built on market criteria, and a vendor can neither enter, nor leave, nor rise by paying.

A vendor rules on nothing itself. It can submit supporting documents on its own profile through the vendor form: a declaration without a document is displayed and does not enter the count; a verifiable document reopens the criterion and can lead to a ruling on it, after verification by intelia.health, never by the vendor. A subscribed vendor obtains, on its own profile, strictly nothing more than a vendor without a subscription.

To consult the rest of the platform, it subscribes on the same terms as any other reader. No sponsoring, no advertising slot, no vendor pricing. What this independence guarantees to the reader:

• dated and revisable evaluations;
• a source for every element of the documented coverage, available to subscribers;
• publication independent of the result;
• no funding by an evaluated entity.

Each published verdict is backed by an official source (EUDAMED, PubMed, ANS, Bpifrance, etc.) or a publicly documented one, dated and cited. A negative verdict is published only if the consulted registry is exhaustive, if the product falls within its scope, and if the evidence carries a date of observation. Otherwise, the criterion remains « not evaluated ». Sources are available to subscribers.

Two-tier Architecture: EU Gates / FR Layer

The review distinguishes two complementary levels:

What coverage counts

For each axis, coverage compares the criteria on which information was found to those that apply to the solution. It measures the extent of the documentation work, never its outcome.

An axis fully verified whose criteria all conclude negatively shows high coverage and a result of zero. This is not an anomaly: it is the most solid measurement we can produce — everything was examined, nothing was found.

Not to be confused with “not evaluated”, which signals the opposite: no verdict could be established. Low coverage says “we found little”. “Not evaluated” says “we could not conclude”.

What “documented” and “consolidated” count — the two figures shown next to the total of referenced solutions. An axis is retained as soon as at least two of its applicable criteria carry a sourced verdict, true or false; vendor declarations do not count. A solution is documented when at least two axes are retained, consolidated when at least three are. Both counts concern the solutions examined criterion by criterion; they say nothing about the watch perimeter — organisations, facilities, countries — which counts everything referenced.

Sourced negative findings

An unmet criterion is not a criterion whose absence is established. A negative finding is published only when three conditions are met.

1. A single authoritative register exists for this criterion. EUDAMED for CE marking, ANS for Ségur referencing: appearing in it or not IS an answer.
2. The product falls within that register's scope.
3. The absence carries a date of observation.

PubMed is not a single register: the absence of a publication proves nothing there, it only fails to find one. In that case the criterion remains “not documented”, never “negative finding”.

A negative finding enters the calculation, on the same footing as a positive verdict. This is the distinction between absence of evidence and evidence of absence — and it is why a low result can be as documented as a high one.

Whether a source is valid for a criterion

The question is not “is the source public”, but “is it valid for this precise verdict”.

A scientific publication counts as clinical validation only if the product is the subject of the study — not if it is merely its data source. A paper using data produced by a solution does not validate that solution. This rule has led to reclassifying previously positive verdicts as negative.

It applies to every family of sources: a code repository is not a measure of model quality, user reviews are not a measure of clinical usability.

Contesting a verdict

Who — the vendor concerned, a professional user, or any reader observing that information is inaccurate.

How — by email to hello@intelia.health, stating the solution, the criterion contested, and the reference allowing verification.

Follow-up — an acknowledgement is sent; the contribution is examined by a person, and how long the review takes depends on the document provided; the revision is published with its date.

Assumed limitation. The framework carries only four statuses, and none credits evidence that exists but cannot be verified publicly. A document shared under a confidentiality agreement therefore falls into “not documented”, which is inaccurate without being arbitrary: we publish only what a third party can replay. A fifth status is under consideration.

• It does not rewrite history: earlier states remain consultable with the verdict that was theirs on their date.
• It grants no right to review before publication.
• It grants no right to the withdrawal of an accurate verdict.
• It gives rise to no financial consideration, in either direction.

The 8 axes · 122 criteria

A strict three-tier rule ensures the reliability of the displayed data.

Data is collected via scrapers from the 20 sources listed above. Scraped data is displayed as-is, without interpretation. It is timestamped and traceable to its source.

Listed vendors can complete their profile via a dedicated form — roadmap, pending certifications, usage context. Each contribution is examined by the intelia.health team before publication. A declaration is displayed and does not enter the count; a verifiable document — trial registration, notified-body document, entry in a public register — reopens the criterion and can lead to a ruling on it. The vendor rules on nothing itself: verification is carried out by intelia.health, and the revision is published with its date.

A profile is said to be validated when its data has been confirmed by a primary source (official publication, vendor response, regulatory database). This level of validation is independent of the completeness status: « partial » designates a number of evaluated axes, not a validation failure.