Infermedica

A probabilistic inference engine over a physician curated knowledge base, which is more than a rules library and less than a learned model. The knowledge base covering roughly a thousand symptoms and conditions is the asset that took years to build and is maintained by clinicians, which is the moat is the dataset precedent this index applies repeatedly.

The inference layer is real reasoning rather than lookup: it computes condition likelihoods and urgency from reported symptoms and risk factors and uses that to choose the next question, which is why the flows are short. Machine learning appears in controlled feedback loops that refine the model against real clinical outcomes, not as the primary reasoning mechanism. Two things earn this a C rather than lower.

The company describes the mechanism accurately rather than calling probabilistic inference artificial intelligence and leaving it there. And the description is independently corroborated: a major cloud provider's own integration documentation characterises it as a deterministic triage engine using machine learning in controlled feedback loops, which is outside characterisation of the kind that let MedAware reach a higher grade and that C the Signs lacked entirely.

A patient answering questions unsupervised and being told what level of care to seek is a consequential recommendation made without a clinician present, and the company is clear about the boundary: the product assesses urgency and directs to a level of care, and is explicitly not for definitive diagnosis or for replacing clinical judgement. The nurse triage assistant module inverts the arrangement, placing the same engine behind a human who is already on the call.

What distinguishes this from a self declared boundary is that the oversight is externally enforced: a Class IIb certification under the European medical device regulation exists precisely because software that triages carries risk, and it required a notified body to review the technical file rather than the vendor to assert its own safety.

The residual question is the same one every triage engine faces and none publishes: where the urgency threshold sits, and therefore how often the engine sends someone to an emergency department who did not need to go, and how often it does the reverse.

Three routes to inspection exist here where most vendors offer none. The company describes its own architecture concretely as a curated knowledge base, an inference engine and a continuous validation process rather than in adjectives. A third party integrating the engine documents the mechanism independently in its own developer material, which is a form of external characterisation this index values highly because the integrator has no incentive to flatter.

And a Class IIb certification means a notified body has examined the technical documentation, which no self certified product can claim. Held at B rather than A because the substance behind those descriptions is not public: the knowledge base contents, the inference model, the vignette set used for validation and the results of that validation are all unpublished, so an outside party can establish that the engine was assessed without being able to see how it performs or where it degrades.

Deployment evidence is exceptional and accuracy evidence is not published. The company states use by ministries of health and national health systems across more than 30 countries, and claims cost to savings ratios reaching one to ten drawn from national deployments, with a calculator offered against those figures. National procurement at that scale is meaningful third party validation of something, though of procurement confidence rather than of diagnostic performance.

What is missing is the number this product class turns on. Independent systematic reviews have repeatedly found that the diagnostic and triage accuracy of symptom checkers varies widely between products, which makes vendor specific published validation more important here than in almost any other category, and two retrieval passes located no peer reviewed accuracy study for this engine. The continuous vignette validation the company describes is exactly the right practice and its results are not published.

A well constructed published position, unusual in this index. Compliance is stated against GDPR and against HIPAA together, reflecting a product sold into both regimes, and it sits alongside audited certifications rather than standing alone as a claim. The company states explicitly that it is designed to meet strict data residency requirements, which almost nothing else assessed in this session addresses at all.

Held at B rather than A because the operational detail is absent: no retention schedule, no minimisation statement, and no position on whether symptom interactions are used to improve the model, which matters here because the company describes machine learning feedback loops driven by real clinical outcomes and does not say what data feeds them or on what basis.

Compliance with HIPAA is stated for the United States market and with GDPR for Europe, which correctly reflects a vendor operating under both rather than asserting one and ignoring the other. The claims sit next to independently audited certifications, which raises their credibility above a badge standing alone.

Held at B rather than A because no Business Associate Agreement terms, subprocessor list or data processing agreement was located, and because a company selling through modular interfaces embedded in other organisations' applications has a more complicated processor and subprocessor picture than a directly deployed product, which is precisely the situation where published terms save a buyer weeks.

The strongest published security posture assessed in this index in some time, and it is earned by external audit rather than assertion. Three named, dated, independently assessed certifications are published together: ISO 27001:2022 for information security management, ISO 13485:2016 for medical device quality management, and a SOC 2 Type 2 report, the last being the version that tests controls over a period rather than at a moment.

Naming the standard version and the report type is itself a disclosure most vendors skip. This is comparable to the Ibex benchmark this index has used for combined certification portfolios, and it stands in sharp relief against the recurring finding of the last several lanes, which is that most healthcare artificial intelligence vendors publish nothing on this axis at all.

Certified as a Class IIb medical device under the European medical device regulation and registered with the United Kingdom regulator. That is the strongest European position in this index and it should not be read as equivalent to the class one markings recorded elsewhere here: class one conformity is self declared by the manufacturer, whereas Class IIb requires a notified body to audit the quality system and examine the technical documentation.

Two records in this index carry self certified class one status and this one carries an assessed certification two classes higher, for a product doing comparable work. Held at B rather than A on a jurisdictional asymmetry that is the interesting part: no FDA clearance, authorisation or submission was located, while the company states HIPAA compliance and sells into the United States.

The same engine that a European regulator treats as a moderate to high risk device operates in the American market with no device review at all, which is the same shape this index recorded in the medication safety lane where insulin dosing software was cleared as a device and alerting software of similar consequence was not.

The operating footprint implies serious generalisation work and none of it is published as governance. Running in 24 languages across more than 30 countries requires the knowledge base and the question flows to hold up across very different populations, disease prevalences and health systems, and prevalence is not incidental to a probabilistic engine, it is an input: the same symptoms carry different likely causes in different places, and an engine tuned on one population and deployed in another will be miscalibrated in ways that are invisible to the user.

No subgroup performance, no per market calibration description, no model card and no bias testing methodology was located. The threshold question compounds it, because where the urgency cut sits determines who is told to seek emergency care, and a threshold appropriate for a population with good primary care access is not appropriate for one without it.

Built interface first, which is the correct architecture for a component meant to sit inside other people's front doors. The engine embeds into websites, mobile applications, patient portals and call centre systems through modular interfaces, and it is available as a named component inside a major cloud provider's health bot service, which is a real and verifiable distribution channel rather than a claimed integration.

National health system deployments imply working connections into public digital front doors. Held at B rather than A because no named record system integration was located and the product's natural position is upstream of the record rather than inside it: triage happens before a patient becomes an encounter, so what the engine learns about a patient does not obviously arrive in the chart, and nothing published describes whether it does.

One of the few vendors assessed in this session to raise residency at all rather than leave it to be asked about: the company states that the platform is designed to meet strict regulatory and data residency requirements, which is what selling to ministries of health across more than 30 countries forces, since several of those customers cannot lawfully let citizen health data leave the jurisdiction.

Held at B rather than A because no hosting regions, architecture or deployment options are named, so a prospect knows the requirement is understood without knowing how it is met, and the difference between a configurable regional deployment and a single hosting location with contractual assurances is exactly what a national customer needs to establish.

No pricing or pricing mechanism published, and the interface based model makes the basis of charge a real question, since per assessment, per interface call and per covered life produce very different bills from the same deployment. What is published is better than most: a stated cost to savings ratio reaching one to ten, attributed to real world national health system deployments, and an offered calculator letting an organisation model its own case.

That is a value claim rather than a price, and it comes from the vendor, so a buyer can size the opportunity but still cannot compute the cost side without a sales conversation. Artrya remains the benchmark on this axis for publishing the price alongside the mechanism.

The broadest coverage of anything assessed in this session and among the broadest in the index. Clinically the scope is undifferentiated presentation, which is by definition every specialty, since a patient reporting chest pain or fatigue has not yet been sorted into one. Functionally it spans patient self triage, pre visit intake and a nurse assistant, so the same engine serves the patient, the front door and the clinician. Channels cover web, mobile, portal, chat, voice and call centre.

Buyers span providers, insurers, telemedicine platforms, public health institutions and national ministries of health. Geographically it reports more than 30 countries including the United Kingdom and Australia, in 24 languages, which is a footprint almost nothing else in this index approaches, since the overwhelming majority of vendors here operate in North America alone. Language coverage in particular is not a translation exercise for a symptom engine, because the clinical content and the question flows have to be right in each one.