AI4Eyes

The model produces the output the product is sold for. AI4Eyes says its software delivers multimodal analysis of ocular surface images, condition detection and personalized treatment recommendations for dry eye disease, and it names the research lead who builds its diagnostic algorithms.

The capture device carries value of its own, though. The OSDx hardware mounts on a clinic's existing slit lamp and consolidates ten dry eye tests into one five minute scan, and the company prices it separately: an estimated US$15,000 for the device, with the AI software sold as a separate subscription charged per disease. The time, device count and footprint comparisons on the home page describe that hardware. The diagnosis and treatment suggestion do not exist without the model, which makes the AI the engine of a core module rather than the whole product.

The site does not say which of the ten test measurements the software computes, or whether the device returns anything usable without a subscription.

The workflow has three published steps: the device scans, the software returns an analysis, condition detection and a treatment recommendation, and clinicians approve results before care is personalized. Approval is a described step in the product's workflow rather than a slogan, and it is the control that stands between a model output and a patient.

What is missing is everything around that step: what the clinician sees when approving, whether a result can be edited or only accepted or rejected, what finding or confidence level sends a case to fuller testing, and what happens to a rejected result. Approval also feeds the model, since the company says the AI learns and adapts as clinicians approve results, so the quality of each approval is a training input as well as a safety check. The site is silent on how disagreement is recorded and whether rejected outputs train the model.

The approach is described in general terms. AI performs multimodal analysis of high resolution ocular surface images captured by patented hardware, and the company calls it fact grounded and built with data feedback loops, so that it improves with every scan and every patient. It names the research lead building its diagnostic algorithms, a strategic partnership with Mila, the Quebec artificial intelligence institute, and a technical advisory board of academic researchers. No model, architecture, training dataset or version is identified. The home page claims the largest ocular dataset without saying what it contains, how it was collected or how large it is.

The continuous learning claim raises the stakes of that gap. If the deployed model changes as clinicians approve results, a buyer needs a version identifier on each result and notice when the model changes, and neither is described.

The model layer is an in house build. The company names the research lead who builds its diagnostic algorithms, and no outside model provider is described, which makes the chain short at the model layer.

The rest of the party list is open. The software is sold as a subscription with continuous updates and a companion patient app, so images and results leave the clinic device for a service somewhere, and no hosting provider, subprocessor or storage location is named. The published privacy policy covers the website only (names, email addresses and browsing data) and says data may be processed wherever the parties involved in processing are located.

Two relationships carry no published terms. The company says its AI learns from every scan and every patient, so clinic images feed training, and it has a research partnership with Mila to develop further applications. Whether Mila or any other partner receives patient images, and on what deidentification and retention terms, is not published.

The performance claim is a vendor figure with nothing behind it. The home page says the platform elevates performance to 89%+ accuracy, against 85% for traditional methods, with no study, reference standard, population or denominator for either number. No peer reviewed validation of the diagnostic or treatment recommendation algorithms is cited.

Validation is funded and named rather than finished. In July 2026 the company announced acceptance into the MEDTEQ+ Avisé program, approved by Quebec's Ministère de l'Économie, de l'Innovation et de l'Énergie, with $700,000 to validate the dry eye platform in a real world clinical setting with CIUSSS du Nord-de-l'Île-de-Montréal. No protocol or reference standard for dry eye diagnosis is published for that study, and nothing commits to releasing results before the device ships.

The one data practice the company states is training use. In the home page's words, its AI evolves with every scan and every patient, learning and adapting as clinicians approve results, all while preserving data privacy. That settles whether patient images reach the model and leaves the terms open: no retention period, consent basis, de identification method or opt out is published. The privacy policy on the site is a generated website notice covering names, email addresses and browsing data, and it does not reach the images or results the device produces.

The data is high resolution imaging of the ocular surface, captured in the clinic. Nothing is published on safety engineering either: no guardrails, no handling of a result found to be wrong, and no process for a safety event.

No HIPAA or business associate position is stated, and no privacy document reaches the product. The only privacy policy published is a website notice built with a free policy generator, covering names, email addresses and browsing data collected through the site. It says nothing about ocular images, diagnoses or treatment recommendations.

The United States is in scope: the company takes refundable US$200 reservations for a device with an estimated retail price of US$15,000. No Canadian equivalent is documented either, since nothing addresses PIPEDA, British Columbia's Personal Information Protection Act or Quebec's Law 25 for the patient data the platform will handle. The company has not said whether it will sign a business associate agreement in the United States, what privacy terms will govern patient data in Canada, or when those documents will be published.

Nothing independent stands behind any security statement. No certification, attestation, penetration test or trust page is published. The privacy policy says the company strives to use commercially acceptable means to protect personal data and cannot guarantee its absolute security, which is the stock wording of a generated website notice rather than a description of controls.

The platform will hold high resolution ocular images and diagnoses in software sold by subscription with continuous updates, so the question is live before commercial launch. Nothing says where images will be stored, who at the company can access them, what certification the company is working toward, or what happens to data collected during validation if the device never ships.

No clearance or authorization covers the product, and the company does not claim one. openFDA lists no 510(k) or PMA record under the applicant name, and no Health Canada license is cited. The home page calls AI4Eyes an AI-powered Software as a Medical Device platform, which is a classification the company applies to itself rather than a regulatory status, and a news item describes the company as advancing toward regulated commercialization. No pathway is named: nothing says whether the company will seek a 510(k), a De Novo or a Health Canada device license, or for which indication.

The product is marketed now without a status label. The home page claims 89%+ accuracy and one trusted diagnosis for dry eye disease, and a reservation page takes US$200 reservation fees for devices the agreement says may never be produced, refunded if none ships by June 2027. No submission date is announced.

No bias evaluation, subgroup result or model governance framework is published. What is published is mission language: the company describes human and technology collaboration, expert validated insights, clinician approval of results, and a fact grounded AI built with data feedback loops. The board expansion in its news concerns corporate governance, not model behavior.

Two features of the product make the gap specific. The model is described as learning with every scan and every patient, and a medical device whose algorithm changes in use is what a regulator reviewed change control plan exists to govern; nothing describes how an update is validated before release. And dry eye presentation varies with age, sex and contact lens wear, yet no performance result is broken out across those groups. The MEDTEQ+ Avisé validation with CIUSSS du Nord-de-l'Île-de-Montréal is the natural place for that breakdown, and no plan for one is published.

A clinician approves each result before it is used. That is review before commit, a mechanism that lets a professional reject an output before it reaches care.

Nothing stands behind the output itself. No warranty, indemnity or remediation commitment attaches to a diagnosis or recommendation, the 89%+ accuracy figure has no method or denominator that would make it a testable commitment, and no route is published for a patient or clinic harmed by a wrong result. The reservation agreement covers only the reservation fee, refundable on request or if no device ships by June 2027, and says any purchase will be governed by standard terms of sale that are not yet published.

Until those terms appear, any limit on liability for the software's output is unknown, and no process is described for a clinic that reports an incorrect diagnosis.

No integration with a record system is described, planned or named, and no standard such as FHIR or DICOM is mentioned. The only integration on the site is physical: the device mounts on a clinic's existing slit lamp.

The output is a diagnosis and treatment recommendation that a clinician approves, which has to reach the chart to be useful, and nothing published says how it gets there, whether through automated or manual entry into a practice management or record system. A free patient app for engagement and education before and after the visit is a second data surface with its own connection questions. Image export in any standard format is not described either.

The shape is visible and the location is not. Capture runs on a compact device, 45 by 61 centimeters, that mounts on the clinic's slit lamp. Analysis is delivered as subscription software through an updatable clinician app store with continuous updates, alongside a patient app. That implies a vendor hosted service, and nothing says where it runs, whether analysis happens on the device or remotely, or where images rest. The privacy policy, which covers the website, says data may be processed and kept outside the user's province or country. The company's offices are in Vancouver and Saint-Laurent in Montreal, and validation is running in Montreal, which implies Canada without committing to it.

Residency matters here. Quebec's privacy law requires an assessment before personal information is communicated outside the province, and nothing says whether a United States clinic's images would leave the country. Nor is there any word on whether a clinic can keep images on premises.

A price and a basis are published. The reservation page gives an estimated retail price of US$15,000 for the OSDx device, with a refundable US$200 reservation fee for the first 100 devices, and the home page describes a one time hardware cost plus a software subscription charged per disease through a clinician app store. The home page also claims a total cost of $9,000 a year against $17,000 for traditional methods, without the assumptions behind either figure. No subscription rate is published, so the recurring cost cannot yet be sized.

The two reservation documents disagree on the discount. The reservation page offers 40% off the retail price; the reservation agreement, updated 25 July 2025, offers 50%, with a further 10% for buyers who reserve and purchase more than 25 devices.

Coverage is stated clearly and narrowly: dry eye disease, assessed in eye care clinics by optometrists and ophthalmologists, with the device mounted on an existing slit lamp. The company says it is expanding into myopia and corneal management and describes a pay per disease app store that would add conditions over time; both are roadmap.

Nothing validates the stated coverage yet. The real world validation with CIUSSS du Nord-de-l'Île-de-Montréal under the MEDTEQ+ Avisé program was funded in July 2026 and has no published results, and the device is not yet on sale anywhere.