Harrison.ai

Deep learning is the whole offering. Harrison.ai, which sold these products as Annalise.ai until a 2025 rebrand, offers detection software for chest X-ray (up to 124 findings in under 20 seconds), non contrast head CT (up to 130 findings) and chest CT (up to 167 radiological features), US triage and notification devices built on the same deep learning approach (the 510(k) summary for K222268 describes a convolutional neural network), and a radiology foundation model, Harrison.Rad 1.5. Take the models away and the imaging products have nothing left to sell.

Two items on the home page sit outside that description: HOP, linked only as a radiology platform with no description, and a prostate pathology line with no page behind it. Neither is described as a product that carries value without the models.

The home page publishes a completeness claim about the chest X-ray product. In a customer testimonial the company places on its own home page, a radiologist says Harrison.ai Chest X-ray "provides a critical safeguard, ensuring no detail is overlooked." The company chose to publish that sentence as marketing for its lead product, so it reads as the company's promise that the product misses nothing. The product's own cleared figures show that it does miss findings. The 510(k) summary for Annalise Enterprise CXR Triage Trauma (K222268, March 2023) reports sensitivity for vertebral compression fracture of 89.3 percent at one operating point and 85.3 percent at the other, so roughly one fracture in nine to one in seven went unflagged on the test set. Taken at face value, the sentence gives a site less reason to keep the review step that would catch a missed finding, and a reader cannot see a finding the software never raised. The company's own copy on the same page is comparative rather than absolute: "catch more. miss less" and "a second pair of eyes".

The oversight structure the US devices publish is otherwise clear. They flag and reorder; they do not read on their own. K222268 describes a workflow tool that runs alongside the standard reading workflow and raises a chest X-ray's priority in the PACS or RIS worklist when it detects features suggestive of vertebral compression fracture. It never lowers a study's existing priority, does not point to a location on the image, and is not to be used on its own for clinical decisions or to rule out a finding. The summary publishes two selectable operating points with their sensitivity and specificity, and names a condition outside what was tested: only erect films were evaluated, and supine films may perform differently.

The comprehensive products, which the company leads with, are described less fully. The home page markets them as clinical decision support for subtle and complex cases, but publishes no per finding thresholds for the 124, 130 or 167 findings, nothing on what the reader sees when the model is uncertain, and no list of uses the products are not intended for. Harrison.Rad 1.5, presented as the company's next generation foundation model, has no published description of its output or of how that output is reviewed.

The approach and the training data are described; version and update practice are not. The 510(k) summary for K222268 describes a convolutional neural network trained with deep learning on images sourced from datasets across three continents and a range of equipment makers and models. The company describes its chest X-ray model as trained on more than 750,000 images spanning departmental and portable imaging, inpatient and outpatient groups and varied demographics, with each case labeled independently by three radiologists under standardized procedures, which names the labeling method rather than leaving it vague.

The company also names a foundation model with a version, Harrison.Rad 1.5, with no model card, training data description or evaluation on its home page. The cleared triage devices carry dated clearances, and a Special 510(k) in April 2025 (K250831) shows a modification went back to FDA, but no release notes or statement of how customers are told a model changed are published. Architecture detail and per finding performance for the comprehensive products are not published either.

The chain is structurally short at the model layer. The models are the company's own: the 510(k) summary for K222268 describes a convolutional neural network the manufacturer trained with deep learning, and a change to a cleared device goes back to FDA, as the Special 510(k) for a modification in April 2025 (K250831) shows. No outside model provider sits in the path. The home page presents the foundation model, Harrison.Rad 1.5, as the company's own as well.

The rest of the chain is not enumerated. The privacy notice (last updated 20 November 2025) says service providers may hold personal information in countries including the United Kingdom, Germany and Australia, and lists infrastructure, channel and analytics partners by category without naming any. No hosting provider or subprocessor list is published in it. The company operates a trust center at trust.annalise.ai, which is the usual home for a subprocessor list.

The published evidence measures performance on test sets and in reader studies; none shows what changes in use. The 510(k) summary for K222268 (Annalise Enterprise CXR Triage Trauma, March 2023) reports a retrospective standalone test on 589 adult chest X-rays from four US hospital network sites, with ground truth from two board certified radiologists and a third on disagreement: vertebral compression fracture sensitivity 89.3 percent and specificity 89.0 percent at the first operating point. The comprehensive chest X-ray model is the subject of a retrospective multireader multicase study of radiologist accuracy with and without the model, published in The Lancet Digital Health in 2021. The company has reported a 32 percent improvement in head CT diagnostic accuracy averaged across findings, without the comparator or design alongside it.

Scale is large and is the company's own count: more than 35 percent of chest X-rays in NHS England, more than 130 UK hospitals across more than 45 NHS trusts, every public emergency department in Hong Kong, access for more than half of Australia's radiologists, and 3,500 clinicians at more than 1,000 sites on four continents. Scale of use is not evidence of benefit. No prospective or randomized study of the products in use, and no independent evaluation, is cited on the home page.

A categorical no training commitment is published; the retention schedule behind it is not. The company's privacy notice (last updated 20 November 2025), which covers its products and services as well as its websites, states that it does not use any content uploaded into its services, including demos, product trials and evaluations, for model training. Where the services are provided under contract, the notice says the customer organization controls the information and the company processes it only on the customer's instructions.

The rest is general. Uploaded content may be kept for as long as necessary to provide the services, with no period stated for images or derived findings and nothing on return or deletion when a contract ends. De-identification is asked of the user rather than described as something the service does: the notice asks users to supply non identifiable or synthetic data wherever possible, which fits the online demos and says nothing about identified studies in a clinical deployment. Safety engineering, meaning guardrails on erroneous output and how a safety event is handled, is not published.

The company describes its training data in detail on the model side: more than 750,000 chest X-ray images from three continents, labeled by three radiologists.

A service level privacy document is published; business associate status is not stated. The company's privacy notice (last updated 20 November 2025) covers Harrison-AI Pty Ltd and its affiliates, including Harrison-AI Medical Pty Ltd, the applicant on the most recent FDA clearance, and the US entities Harrison-AI Medical Inc and Harrison.AI Services Inc. It applies to the company's products and services as well as its websites, and says that where the services are provided under contract with an organization, that organization controls the information and the company processes it solely on its behalf and on its instructions.

That is a processing role stated in general terms. It names no regime and no agreement, so it does not stand in for a business associate position. HIPAA appears once, in the California section, among categories of health information that state law excludes. No business associate agreement, data processing agreement or subcontractor terms are published in the notice. The company operates a trust center at trust.annalise.ai.

No security attestation is published on the company's main pages. The harrison.ai footer carries privacy, terms and policy links but no security or trust page, and the privacy notice (last updated 20 November 2025) describes controls in general terms (industry standard safeguards, confidentiality obligations on staff and subcontractors, need to know access) without naming a certification or an independent assessment. The company does operate a trust center at trust.annalise.ai; any attestation it holds, such as SOC 2 or ISO/IEC 27001, with its scope, date and report, is what a buyer's security review needs.

The products are regulated devices, and FDA and European notified bodies review the manufacturer's cybersecurity documentation as part of clearance and CE marking. That examines the software as submitted, not how the company runs its own security operations.

FDA's 510(k) database lists nine clearances for the company, all radiological computer aided triage and notification software under 21 CFR 892.2080. Chest X-ray (product code QFM): K213941 Annalise Enterprise CXR Triage Pneumothorax, 24 February 2022; K222179 and K222268 Annalise Enterprise CXR Triage Trauma, both 28 March 2023; K250831 Annalise Enterprise, a Special 510(k) for a modification, 23 April 2025. Head CT (QAS): K223240 Annalise Enterprise CTB Triage Trauma, 3 April 2023; K231094 Annalise Enterprise CTB Triage-OH, 15 August 2023; K231384 and K231767 Annalise Enterprise CTB Triage Trauma, both 22 September 2023; K253818 Annalise Enterprise, 3 March 2026, the first filed under the Harrison-AI Medical name. The applicant on the earlier clearances, Annalise-AI Pty Ltd, gives the same Sydney address as the Harrison.ai site.

The scope is narrow. Each clearance is triage and notification for named findings, not detection or diagnosis. K222268, for example, covers adult chest X-rays with features suggestive of vertebral compression fracture, was validated on erect films only, and is not for ruling a finding out. The per finding indications of the other clearances are in their summaries, each public under its number. The comprehensive products sold as Harrison.ai Chest X-ray (up to 124 findings), CT Brain (up to 130) and CT Chest (up to 167 features) are not what these clearances cover. The company has announced EU MDR Class IIb CE marking for its chest X-ray and head CT products, states clearance for clinical use in more than 40 countries, and reported a Breakthrough Device designation for its obstructive hydrocephalus triage device.

Evaluation across patient groups is described but not reported. The 510(k) summary for K222268 says its test cohort included representation across sex, age, ethnicity and race, equipment make and model, positioning and projection, and coexisting findings. It publishes no counts or results for any of those groups, and no finding that performance held across them. The one limit it names is positioning: only erect films were tested, and supine films may perform differently.

No governance framework, AI management system certification or subgroup analysis for the comprehensive products appears on the home page or in the privacy notice. Chest X-ray classifiers trained on public datasets have been shown to underdiagnose some patient groups (Seyyed-Kalantari and colleagues, Nature Medicine, 2021), yet no results by sex, age, race and equipment are published for either the clearance studies or the comprehensive models.

The company's own FDA filing publishes an error rate for one of its chest X-ray triage findings, with method and denominator. In the 510(k) summary for Annalise Enterprise CXR Triage Trauma (K222268, cleared 28 March 2023), a retrospective standalone test on 589 adult chest X-rays (272 positive, 317 negative) from four US hospital network sites, with ground truth set by two board certified radiologists and a third resolving disagreement, flagged vertebral compression fracture with sensitivity 89.3 percent (95% CI 85.7 to 93.0) and specificity 89.0 percent (85.8 to 92.1) at the first operating point, and 85.3 and 90.9 percent at the second. Area under the curve was 0.954. Only erect films were tested, and the summary warns that supine films may perform differently.

These are test set results for one cleared finding, not performance in use. They do not cover the other triage findings, the comprehensive detection products (up to 124 chest X-ray findings, 130 head CT findings and 167 chest CT features), or the foundation model. For a triage product a miss in use leaves no artifact: the study keeps its place in the worklist and is read in the ordinary course. Neither the home page nor the privacy notice publishes a warranty, indemnity or remediation commitment attached to the output.

Integration is through imaging standards, with no system named. The 510(k) summary for K222268 describes the software receiving DICOM chest X-ray studies from image and order management systems such as a PACS or RIS, running in parallel with the standard workflow, and returning a case level result to the PACS or RIS for worklist prioritization.

No PACS, RIS, reporting or record system is named as a production integration on the home page or in the summary, and nothing describes results reaching the electronic health record or the radiology report.

Residency is partly stated and the processing path is open. The company's privacy notice (last updated 20 November 2025) says its service providers may hold personal information in countries including the United Kingdom, Germany and Australia, that information may be processed outside the country of residence, and that European transfers rely on Standard Contractual Clauses or similar mechanisms. The company has described two architectures: cloud processing for the enterprise products, and an edge option for point of care chest X-ray that returns findings in about ten seconds. The current home page lists no separate edge product.

Nothing published states which architecture a site runs, the hosting provider, the region for a US deployment, how customer environments are separated, or what the edge option transmits. The 510(k) summary for K222268 says the software takes DICOM studies from the PACS or RIS and returns its result there, without saying where processing happens. Because the products are sold in more than 40 countries with different localization rules, the region and provider matter.

No price is published. The products are sold to healthcare organizations (public hospitals, private hospitals and private radiology providers) by contacting the company, and the site states they are not available for purchase by the general public. The US triage devices and the comprehensive detection products for chest X-ray, head CT and chest CT are separate products, so a buyer can see what is being bought before calling, but not the license basis or the price.

The one public figure is on the payer side. Medicare approved a New Technology Add-on Payment for the obstructive hydrocephalus triage device (Annalise Enterprise CTB Triage-OH, K231094), which the company announced. That is what Medicare pays a hospital for eligible inpatient cases, not what the vendor charges, and add-on payments are temporary.

Coverage is named, and validation stands behind part of it. The company sells chest X-ray (up to 124 findings), non contrast head CT (up to 130 findings) and chest CT (up to 167 radiological features) to public hospitals, private hospitals and private radiology providers, with solution pages for emergency and acute care and for oncology management, and states clearance for clinical use in more than 40 countries.

The US triage clearances carry defined, validated populations. K222268, for example, covers adult chest X-rays, was tested on erect films at four US sites, and names supine films as possibly performing differently. The comprehensive chest X-ray model has a published retrospective multireader multicase study (The Lancet Digital Health, 2021). No validation is described for the chest CT product's 167 features, or by setting (emergency, inpatient, outpatient, screening).