Infermedica

Triage is the product, and the engine that produces it contains learned parts. The company describes its inference engine as probabilistic models built on modified Bayesian networks and machine learning methods, reasoning over a physician curated Medical Knowledge Base of more than 1,800 symptoms, 900 conditions and 340 risk factors.

The engine chooses each next question from the answers so far and ends with the most probable causes paired with a five level care recommendation: self care, consultation, consultation within 24 hours, emergency, or ambulance. Natural language processing turns a patient's free text into those concepts, and the newer Conversational Triage and pre-triage voice agent put a general purpose large language model, speech to text and text to speech in front of the same engine.

Every product on sale, from the Triage, Intake and Follow-up modules to the Nurse Triage Co-Pilot and the Engine API, runs on that engine, so the capability does not exist without it. Even so, the company's own account places much of the value in curation rather than in anything learned.

Its AI page calls the system an expert system built on explicit rules and logic, the knowledge base page calls the knowledge base the foundation of every product, and 36 physicians and more than 155,000 hours of physician input are credited for it.

How much of the engine's behavior comes from parameters estimated from data, and how much from physicians' judgment about the links between symptoms and conditions, is not stated, and that split governs how the engine's behavior shifts when it is updated.

The oversight structure differs by product and is described for each. In patient self triage and Conversational Triage the patient answers questions with no clinician present and receives a recommendation on a five level scale from self care to ambulance, and the deploying organization configures the rules and routing that follow.

The pre-triage voice agent gathers symptoms before a nurse picks up and sorts callers by clinical urgency, and its launch post says it can also guide a caller to self care, a live nurse or emergency care at any hour. In the Nurse Triage Co-Pilot the engine proposes a triage level and the nurse on the call advises the patient; the results screen shows the engine's level, the nurse's recommendation and the patient's decision side by side, which gives a supervisor an override trail.

The scope is stated plainly: the intended use is guiding users through symptom questions to an appropriate care setting. What is missing is the threshold. The probability or rule at which a case moves from one care level to the next is not published, and neither is the point at which the voice agent must hand a caller to a nurse rather than finish the assessment itself.

That matters here because the one independent safety test the company lists, Luckraj and colleagues (2024) on 78 standardized vignettes for Healthdirect Australia's nurse helpline, found 19 percent undertriage against 15 percent for the system it replaced, alongside much less overtriage.

The company's lines that its approach "guarantees maximum accuracy" and that the platform produces "the same results in all circumstances" speak to ambition and consistency, not to a promise that no answer is wrong.

A good deal of the method is published. The engine is described as probabilistic models built on modified Bayesian networks and machine learning methods over a knowledge base whose size is stated (more than 1,800 symptoms, 900 conditions and 340 risk factors, adult and pediatric) and whose content is drawn from more than 12,000 published sources and curated by a named physician team; the applied scientists who build the engine are named on the company's AI page, and the company says its content creation procedures and acceptance test cases are open, with the full list of symptoms and conditions on its developer portal.

The natural language layer has a published evaluation from 2021, with the method explained: precision 0.852, recall 0.816 and F-score 0.833 on about 2,500 patient messages. A regression testing practice for the knowledge base is described, and the engine is tested against thousands of documented clinical cases. The generative layer is less clear.

Conversational Triage combines "a general-purpose Large Language Model" with the engine, and the voice agent adds speech to text and text to speech, none of them named. The company publishes no version or update practice. Nothing says when the knowledge base or the engine changed, what changed, or how customers are told, and the accuracy figures on product pages (93 percent engine accuracy on cases from BMJ, NEJM and JAMA; 96 percent conditions accuracy) carry no version, date or denominator.

The core is the company's own build: the knowledge base and the inference engine are in house. The newer products add parties it does not name. Conversational Triage runs on "a general-purpose Large Language Model", and the pretriage voice agent combines large language models, speech to text and text to speech with the engine, with no provider named for any of them, so the outside model that receives a patient's own description of their symptoms is unknown, as are its retention and training terms.

Hosting is stated two ways: the platform page says the technology is hosted on Google Cloud, while the security portal names Amazon Web Services as the infrastructure provider. The portal links a subprocessor list, but the list itself is not public. Microsoft's Health Bot service, which embeds the engine, and partners such as Hyro, Kore.ai, Talkie.ai and Vonage sit downstream as integrators of Infermedica rather than as its suppliers. The architecture is described and a host is named, but no model provider is.

The company's headline is safe triage that moves patients to the right level of care at national scale, and it lists a long run of peer reviewed papers, most written by its own staff. Two independent evaluations stand out. Boğa (Western Journal of Emergency Medicine, 2026), a single emergency physician at a state hospital in Istanbul, randomized 18,000 adult patients to Infermedica supported triage, the Emergency Severity Index or the Manchester Triage System, and reported lower in-department mortality than Manchester (odds ratio 0.39), fewer complications (4.42 against 10.25 percent) and higher satisfaction (9.0 against 7.0); the author notes the single center design and short observation period, and the setting is emergency department triage rather than the patient and call center products the company sells.

Luckraj and colleagues (Studies in Health Technology and Informatics, 2024), from Healthdirect Australia and Macquarie University, tested the nurse helpline decision support on 78 standardized vignettes: 66 percent triaged correctly against 57 percent for the previous system, overtriage 15 against 28 percent, and undertriage 19 against 15 percent, concluding it was at least as safe. The company's research page reports the overtriage reduction and not the undertriage figure.

Company authored work covers the deployments: care seeking intent before and after triage at Healthdirect (Mayo Clinic Proceedings: Digital Health, 2026), Médis in Portugal, Reliant Medical Group in the United States and an Australian health plan; a comparison with Schmitt-Thompson protocols on live calls; and a JMIR AI 2026 study of 116,890 website encounters comparing the questionnaire with the language model conversational interface.

Most deployment outcomes are what patients said they intended to do rather than the care they went on to use. Named deployments with dated figures and methods, and independent evaluations that answer a narrower question than the headline, one in an emergency department and one on vignettes rather than live national use. Nothing is published on the voice agent's performance.

General assurances that stop short of the questions artificial intelligence raises. The company states compliance with GDPR and HIPAA, encryption of patient interactions, configurable access controls and audit logs, and says the platform is designed to meet strict data residency requirements. What is retained from a symptom interview, and for how long, is not stated.

Reuse is implied rather than bounded: the platform page says a feedback loop ensures continuous improvement of the engine, and the knowledge base page says physicians' experience of millions of real cases helps improve its accuracy, without saying what data drives that, whether it is de-identified, or whether a customer can opt out. The only privacy policy published covers the company website, not the product, and the security portal's documents sit behind an access request.

The generative products add an unnamed outside language model that receives the patient's own words, with nothing published on what that provider keeps. Safety engineering is described in design terms (the language model handles conversation while recommendations come from the engine, with deterministic and explainable decision paths), but no guardrail testing, hallucination measurement or safety event process is published.

Compliance is claimed without a document a buyer can read behind it. The regulatory page marks HIPAA as completed and describes encryption, access controls and audit logs, and GDPR is stated alongside it for the European business. No business associate status is stated, no agreement or data processing terms are published, and the only privacy policy covers the company website rather than the service that handles patient interviews.

The security portal lists HIPAA, GDPR and the NHS Data Security and Protection Toolkit among its frameworks, with the documents behind an access request. For a company whose product usually runs inside another organization's app or phone line, the question of who is whose business associate is a real one.

Recognized certifications are named with versions, and the reports sit behind a request. The company names ISO/IEC 27001:2022 for information security and ISO 13485:2016 for its medical device quality system, and its home page cites a SOC 2 Type 2 report.

The security portal at security.infermedica.com lists SOC 2 Type 1 and Type 2, ISO/IEC 27001 with its statement of applicability, ISO 13485:2016, Cyber Essentials and Cyber Essentials Plus, the NHS Data Security and Protection Toolkit, PCI DSS and NIS2, and offers an ISO 27001:2022 report, a SOC 2 report and a penetration test report through a "Get access" request, with access lapsing after one month.

The portal's public page gives no report dates, auditors or SOC 2 trust services categories, and one company page still cites ISO 27001:2017. Recognized certifications named in the company's own material, with the artifacts released on a request the company controls.

The regulatory position is stated market by market. In the European Union the Medical Guidance Platform, with its Triage, Intake and Follow-up modules, is certified as a Class IIb medical device under the Medical Device Regulation, which requires a notified body to assess the technical documentation; the legacy Triage module remains a Class I device under the older directive until the end of 2028.

The products are registered with the MHRA in the United Kingdom and with ANVISA in Brazil, registered in Switzerland, and in progress with Australia's TGA and under the EU AI Act, where the company classifies them as high risk. For the United States the company states that its software "operates under FDA enforcement discretion due to its clear Intended Use: guiding users via symptom questionnaires to recommend appropriate care settings", and the nurse co-pilot page says Triage does not require FDA clearance.

That position gives its reasoning, and the reasoning corresponds to FDA's published policy for software that lists possible conditions from common symptoms and advises when to see a provider. Enforcement discretion here is still the company's own reading, and there is no entry in FDA's databases to check it against.

The stated intended use also describes symptom questionnaires, while the newer Conversational Triage and pretriage voice agent put a general purpose language model and speech recognition in front of the engine, and nothing says whether the Class IIb certificate or the US position covers them. No certificate number or notified body is named.

Multilingual operation is marketed with no performance published by language or population. The knowledge base runs in 26 languages, the voice agent in English, Spanish, Portuguese and Polish, and Conversational Triage is said to interpret symptoms described in natural language accurately, including vague language, spelling mistakes and colloquialisms. The only published recognition result, from 2021, is one overall F-score of 0.833 on about 2,500 messages, with no breakdown by language.

The company says customers "have conducted comprehensive testing of our technology, demonstrating its safety and fairness across diverse geographical areas, populations, and ethnicities", without naming the tests or publishing results. For a probabilistic engine this matters twice over: prevalence is an input, so the same symptoms carry different likely causes in different populations, and a caller the system misunderstood leaves no complaint and no record.

What is published is process: a clinical validation team testing the engine on thousands of documented cases, regression testing of the knowledge base, and papers describing who uses the tool (pregnant users, people with obesity, displaced Ukrainians), which describe use rather than performance. The medical device certification and quality system are counted under regulatory status. D, because recognition quality is sold as a strength across many languages with no subgroup results behind it.

Error rates for the triage engine are published with their method and denominator, though the company's own summary leaves part of them out. The Healthdirect Australia evaluation the company lists (Luckraj and colleagues, Studies in Health Technology and Informatics, 2024) tested the nurse helpline decision support on 78 standardized vignettes over two cycles and reported 66 percent correct triage, 15 percent overtriage and 19 percent undertriage; the company's research page repeats the overtriage improvement and omits the undertriage figure.

A company authored comparison with Schmitt-Thompson protocols on live calls reports no critical mis-triage, meaning no emergency sent to self care. Mechanisms to challenge an output exist: each question and recommendation can be traced to the engine's reasoning, the nurse product records the engine's level beside the nurse's decision, and call histories are searchable. Nothing stands behind the output: no warranty, indemnity or remediation commitment is published, and the only published terms cover a sales chatbot on the company website.

Named systems with data flowing one way. The Triage module is listed in Epic's App Orchard to add symptom checking to MyChart, and the Intake module is integrated with ProAssist, with other record systems reached through the Infermedica API. Intake sends a structured pre-visit summary and a list of possible conditions to the physician, and Conversational Triage supports pre-fill from patient records and maps findings to ICD-10 and SNOMED.

The engine also ships inside Microsoft's Health Bot service and in partners' conversational platforms (Hyro, Kore.ai, Talkie.ai, Vonage), which is distribution rather than record integration. Named systems with reads and one directional writing.

No bidirectional integration or production use with a named record system is published, and triage usually happens before a patient becomes an encounter, so whether the voice agent's structured handoff reaches the chart depends on the customer's own build.

A single hosted service, with residency offered as a design goal rather than a commitment. Customers reach the engine through the API, an embeddable iFrame or white label apps, all hosted by the company. The home page says the platform is "designed to meet strict regulatory and data residency requirements", which ministries of health in more than 30 countries would need, but no hosting region, regional deployment option or tenant isolation is stated. The host itself is given two ways: the platform page names Google Cloud and the security portal names Amazon Web Services. One hosted option with its location implied.

No price or unit of charge is published, but how the product is sold can be established before contact. Organizations license modules (Triage, Intake, Follow-up, Conversational Triage, the voice agent and the nurse co-pilot) delivered through the API, an iFrame or white label apps; developers can request an API account; and the company publishes a savings calculator and a cost to savings ratio of up to 1 to 10 drawn from national deployments. That is a value claim rather than a price.

A buyer cannot tell whether the charge is per assessment, per call, per member or per license until a sales conversation, and the answer changes the bill a great deal for a national front door or a large health plan.

Coverage is named broadly and validated in part. Clinically the engine covers undifferentiated presentations in adults and children across specialties, with content extended over the years from common illnesses to skin conditions, mental health, surgical complications, injuries and pediatrics.

It is delivered as self triage on websites and apps, pre-visit intake, follow-up, a nurse call center co-pilot, a patient phone line and chat, to ministries of health, health plans, providers, telemedicine companies and pharmaceutical companies. The company reports use in more than 30 countries and 26 languages and more than 27 million checkups.

Validation covers part of that: vignette testing and deployment studies for Healthdirect Australia's nurse line and national digital front door, a Portuguese insurer's nurse line, a US medical group, Australian and Middle Eastern health plans, and one Turkish emergency department. The voice agent, available in four languages, and most of the 26 languages have no published validation.