AI4Eyes

Machine learning and generative AI do the diagnostic work the product exists to perform. The company describes multimodal analysis of high resolution ocular surface imaging producing condition detection and personalized treatment recommendations, with the stated aim of consolidating ten gold standard dry eye tests into one exam lasting a few minutes. The patented hardware captures the data, but the diagnosis and treatment suggestion are model outputs; without them the device is a camera.

The workflow is explicitly clinician gated: AI delivers analysis, detection, and personalized treatment recommendations, and the company states clinicians approve results, after which the system learns and adapts. Describing an approval step as an integral stage of the loop rather than an afterthought is the right structure for software producing a diagnosis and treatment suggestion before the physician has seen the patient. Buyers should note the corollary that clinician approvals feed model adaptation, which makes approval quality a training input as well as a safety control.

The approach is described at a functional level, covering machine learning and generative AI applied to multimodal ocular surface imaging with a continuously learning loop, and the company has a stated partnership with a major academic machine learning institute to develop next generation applications. No model architecture, training data provenance, or performance metrics were located, which is the expected position for a company still pre commercial.

No published statement on patient data handling was located: no retention period, deletion procedure, consent description, de identification method, sub processor list or residency commitment. What makes this more than routine early stage thinness is that the company has already answered the hardest question in the affirmative and left every follow up unaddressed.

Its own material states that with every patient it sees the system becomes smarter, and that the system learns and adapts as clinicians approve results. That is an explicit statement that patient data improves the model. Most vendors in this index leave the question ambiguous; this one resolves it in marketing copy and then says nothing about the terms, which is an unusual combination and a more useful starting position for a buyer than silence would be, because there is nothing left to establish except the conditions.

Those follow directly. Whether patients consent specifically to their images being used for model improvement, or whether clinician approval of a result is being treated as sufficient, which would substitute a professional's workflow action for a patient's decision about their own data. Whether images are de identified before entering the training path and by what method. What is retained, for how long, and what happens at the end of a clinic's relationship.

The material is high resolution imaging and video of the ocular surface captured in clinic, which is identifiable in ways a tabular record is not. Ask whether a clinic can use the product without contributing data.

Evidence is in progress rather than published, and the company is transparent about that stage. A subdivision of Quebec's regional health agency is assessing image and annotation quality alongside the accuracy of the diagnostic and treatment recommendation algorithms, which is independent evaluation by a public health body rather than a vendor study.

The company also cites the clinical problem specifically, that over 50 percent of people entering an eye clinic are affected by dry eye while only 8 percent are diagnosed. No accuracy figures, peer reviewed publications, or completed validation results were located.

No published statement on patient data handling was located: no retention period, deletion procedure, consent description, de identification method, subprocessor list or residency commitment.

What makes this more than routine early stage thinness is that the company has already answered the hardest question in the affirmative and left every follow up unaddressed. Its own material states that with every patient it sees the system becomes smarter, and that the artificial intelligence learns and adapts as clinicians approve results. That is an explicit statement that patient data improves the model. Most vendors in this index leave that question ambiguous; this one resolves it, and then says nothing about the terms.

The unanswered questions follow directly. Whether patients consent specifically to their images being used for model improvement, or whether clinician approval of a result is treated as sufficient. Whether the images are de identified before they enter the training path, and by what method. What is retained, for how long, and what happens at the end of a clinic's relationship with the company. The data in question is high resolution imaging and video of the ocular surface, captured in the clinic, which is identifiable in ways a tabular record is not.

Buyers and participating clinics should establish the consent basis for model improvement in writing, and should ask whether a clinic can opt out of contributing data while still using the product.

No HIPAA position, business associate agreement or related terms were located, and in this case that is the expected and appropriate answer rather than a shortfall.

This is a Montreal company, incorporated in Quebec, pre commercial, currently validating its platform in Quebec clinics in collaboration with a regional public health authority and funded in part by Canadian federal and provincial innovation programmes. The applicable privacy framework is Quebec and Canadian health privacy law, not HIPAA, and nothing retrieved indicates any United States commercial activity. A vendor with no United States operations has no occasion to publish a HIPAA position, and its absence carries no information about the quality of its data handling.

The grade records what a United States buyer would find, which is nothing usable, and that is the correct signal for this index to send. It should be read alongside the observation that the company is at an early stage in a single provincial market.

The practical guidance differs by buyer. A Canadian clinic should ask how the platform meets Quebec provincial privacy requirements, including the more recent modernised obligations, rather than asking about HIPAA at all. A United States buyer should treat the absence of any HIPAA framework as a genuine gap that would need to be built before deployment, and should ask what the timeline for that work is.

No trust centre, security page, certification or third party attestation was located, and no security controls are described in any retrieved material.

The stage accounts for this. The company was founded in 2020, remains pre commercial, and is currently in real world validation rather than production sale. Independent security assessment is not a normal expectation at that point, and this grade should be read as recording what exists today rather than as a criticism of the engineering.

One element is nonetheless worth putting on the record. Canadian federal innovation funding was awarded specifically to develop the cloud infrastructure and data structure underlying the platform, so a cloud component holding ocular imaging exists and has been deliberately built. A buyer can reasonably ask what security design decisions were made during that work, which is a more productive question at this stage than asking for a certificate that will not exist.

Clinics participating in validation should ask where images are stored, who at the company can access them, and what happens to data collected during the validation period if the product does not proceed to commercial release. That last question is specific to pre commercial deployments and is easy to overlook.

No FDA clearance and no Health Canada authorization were located, and the company itself describes advancing toward regulated commercialization, which is a candid acknowledgment that it is not there yet. This is the central constraint: software that produces a dry eye diagnosis and treatment recommendation is squarely within device regulation, so current deployment sits in evaluation and clinical assessment contexts rather than routine authorized care.

Board expansion was explicitly framed around strengthening governance as the company advances toward regulated commercialization. Grade reflects clear disclosure of pre clearance status rather than any regulatory achievement.

No bias assessment, subgroup performance analysis, responsible artificial intelligence statement or model governance framework was located.

The absence matters more here than the company's size would suggest, because of a design choice it advertises. The company describes itself as a software as a medical device developer, and separately states that with every patient it sees the system becomes smarter, learning and adapting as clinicians approve results. A medical device whose algorithm changes in the field as it accumulates data is precisely the object that predetermined change control planning exists to govern. A small number of vendors in this index have obtained regulator agreed change control plans for exactly this situation. Nothing published here describes how model updates are validated before deployment, whether a clinician sees which version produced their result, or how performance is monitored for drift once learning is live.

The subgroup question is also specific rather than generic. The platform captures high resolution imaging of the ocular surface and consolidates ten diagnostic tests, and ocular surface appearance varies with pigmentation, age and tear film characteristics. A validation programme is currently running with a regional public health authority, which is the natural opportunity to stratify results and publish them.

Buyers should ask how model updates are controlled and validated, and should ask the validation partners whether subgroup performance is part of the study design.

The approach is described at a functional level, covering machine learning and generative techniques applied to multimodal ocular surface imaging with a continuously learning loop, and the company has a stated partnership with a major academic machine learning institute to develop further applications.

No model architecture, training data provenance, evaluation methodology, performance figure or warranty, indemnity or remediation commitment was located, which is the expected position for a company still pre commercial and is recorded as such rather than as evasion. One property of the design deserves flagging now rather than at commercial launch, because it is easier to build in than to retrofit.

A continuously learning loop that adapts as clinicians approve results means the deployed model is not a fixed artefact, so a performance figure established at any point describes a system that no longer exists, and a clinic cannot validate once and rely on it.

It also means the training signal is clinician approval rather than patient outcome, and a model learning what clinicians accept will converge on their existing practice including its errors, which is the same structural point this index recorded where a coding model learned from payer decisions. Ask how versions are identified and communicated, what performance monitoring runs after each update, and what the model is optimised against.

No electronic health record integration was located, and no integration standard, named record system or clinical documentation pathway is described.

The product's shape explains part of this. It is a combined hardware and software system: patented imaging hardware captures ocular surface data in the clinic, and the software returns a condition detection and a personalised treatment recommendation for the clinician to approve. A device that produces a result at the point of examination can function without touching the record at all, in the same way a standalone imaging instrument does, and a pre commercial company will reasonably prioritise the examination itself over downstream integration.

That said, the output is a diagnosis and a treatment recommendation that a clinician approves, which is exactly the kind of content that has to reach the chart to be clinically useful and medicolegally sound. How it gets there is unaddressed. A separate patient facing application is mentioned for engagement and education, which is a second data surface with its own integration questions.

Buyers should ask how an approved result is documented in their record system, whether that is automated or manual re entry, what integration is planned and on what timeline, and how the patient application relates to the clinical one.

No deployment documentation, hosting arrangement, implementation timeline, availability commitment or data residency statement was located.

The architecture is nonetheless partly visible and is a hybrid rather than a pure software deployment. Patented imaging hardware sits in the clinic and captures the ocular surface data, while Canadian federal innovation funding was awarded specifically to build the cloud infrastructure and data structure behind the platform. So a clinic adopting this takes on a physical device, a cloud dependency, and whatever integration follows, and the balance between what runs locally and what runs remotely is not published.

Real world deployment exists but is validation rather than commercial. The platform is being tested in Quebec clinics under a provincially funded programme, with a regional public health authority assessing image and annotation quality alongside the accuracy of the diagnostic and treatment recommendation algorithms. That is a more rigorous arrangement than most early deployments and is worth crediting as evidence the product is being examined by someone other than its maker.

Residency is unaddressed and is a real question for a Canadian company, since provincial health information rules can constrain where data may be stored. Buyers should ask whether patient images remain in Canada, and specifically within the province, and what the answer would become if the company expands beyond it.

No prices are published, but the intended model is disclosed in outline and is unusual enough to note: an updatable application store for clinicians with pay per disease pricing, plus a free patient application for engagement and education. Pay per disease implies the platform expands beyond dry eye by adding condition modules a practice buys individually, which tells a buyer how cost will scale even without a rate.

Single condition and single specialty today: dry eye and ocular surface disease within ophthalmology and optometry clinics, addressing what the company frames as a large undiagnosed population. The pay per disease application store model signals intent to broaden into additional ocular conditions, which should be treated as roadmap. Geographic footprint is Montreal clinics with Quebec health system assessment underway, so this is not a US available product today.