Waiv

Both products are algorithms operating on slides the laboratory already produces, with no assay, reagent, scanner, or sequencing step of their own. That is the entire commercial thesis: extract prognostic and biomarker signal from routine H and E rather than from an additional molecular test. Remove the model and there is no product and no test.

Both products are structurally positioned to inform a clinician rather than replace a test, but they do so differently and the distinction matters. RlapsRisk BC produces a risk score used alongside clinico-pathological factors, explicitly as an additional layer of information rather than a standalone determination.

MSIntuit CRC is a pre-screen that rules OUT microsatellite stable cases so pathologists concentrate confirmatory testing on the rest, which means a false negative removes a patient from confirmatory MSI testing entirely. A rule-out design placed upstream of confirmatory testing carries different risk than a flag placed alongside it, and buyers should understand which they are deploying.

Development and validation cohorts are disclosed with real specificity for RlapsRisk BC: more than 1,800 patients in training and validation on independent cohorts of more than 1,200 slides, conducted blind and single shot, which is a stronger methodological statement than most vendors make.

Performance is reported as a concordance index, 0.81 for the AI score combined with clinico-pathological factors versus 0.76 for those factors alone, which quantifies the incremental contribution rather than the model's standalone accuracy. That framing is the honest one for an adjunct test. Architecture detail and per-subgroup performance are not published, and equivalent methodological detail for MSIntuit CRC was not located.

The architecture puts patient material in the vendor's hands and the published record does not describe what happens to it. Tests run through a platform where slides are uploaded and analysed, so whole slide images of patient tissue are transmitted to and processed by the company, and nothing located states a retention period, a deletion position at the end of a relationship, whether images are de identified before processing, or whether customer material may contribute to development of future models.

That last question carries particular weight because the company's own positioning rests on it: it describes proprietary foundation models trained on multimodal data, a patient data network inherited from a decade of research, and participation in a large multi institutional consortium. A vendor whose competitive claim is the scale and quality of its training data owes a buyer a clear statement on whether their cases join it. Two things are credited.

Instructions for use are published openly rather than behind a login, which is uncommon for software as a medical device and lets a laboratory read the operating conditions before purchasing. And a patient facing information page is maintained, which is rare for a business to business diagnostics vendor and suggests the patient is treated as a stakeholder rather than only the laboratory. Scope note: a privacy policy is published and was not opened on this pass, so this records what was located rather than proof of absence and should be read first on refresh.

Genuine evidence with an important caveat the vendor itself discloses. RlapsRisk BC was developed with Gustave Roussy following a 2019 regional AI for Health challenge grant, with research presented at ESMO 2021 and USCAP 2022, and blind validation on independent cohorts exceeding 1,200 slides.

The caveat, stated in the vendor's own footnote, is that the validation cohort contained patients treated with endocrine therapy only and was ENRICHED with patients who experienced metastatic relapse, which inflates apparent discrimination relative to an unselected screening population. Independent validation is underway through a partnership with Cerba Path across routine digital workflows. Conference presentations and enriched cohorts sit below the standard of the peer reviewed multi-cohort validations achieved by Stratipath and DoMore in adjacent indications.

The architecture puts patient material in the vendor's hands and the published record does not yet describe what happens to it. Waiv's tests run through Destra, where slides are uploaded and analysed, so whole slide images of patient tissue are transmitted to and processed by the company.

Nothing located states a retention period, a deletion position at the end of a relationship, whether images are de identified before processing, or whether customer material may contribute to development of future models. That last question carries weight here because the company's own positioning rests on it: Waiv describes proprietary foundation models trained on multimodal data, an extensive patient data network inherited from a decade of research, and participation in a large multi institutional European data consortium.

A vendor whose competitive claim is the scale and quality of its training data owes a buyer a clear statement on whether their cases join it. Two things are credited. The company publishes instructions for use openly rather than behind a login, which is uncommon for software as a medical device and is the same disclosure habit credited elsewhere in this lane.

And it maintains a patient facing information page, which is rare for a business to business diagnostics vendor and suggests the patient is treated as a stakeholder rather than only the laboratory. Scope note: a privacy policy is published and was not opened on this pass, so this records what was located rather than proof of absence, and it should be the first document read on refresh. Ask for a data processing agreement covering slide images, a retention and deletion schedule, and a written statement on model training.

No HIPAA position was located. HIPAA is not named, no business associate status is stated, no business associate agreement is referenced, and the published instruments address the General Data Protection Regulation and French data protection law, with the cookie policy explicitly framed around the requirements of the French supervisory authority.

For a Paris based company whose two commercially available tests carry the EU CE mark, and whose third product remains research use only, that is a coherent position rather than a gap, and it should not be read as a compliance failure. It does mean a United States laboratory has nothing published to work from. Two things make this worth revisiting sooner than for the other European vendors in this lane.

The company has stated an intention to expand globally following its March 2026 financing, and a documented pharmaceutical partnership programme spanning multiple large drug developers gives it commercial reasons to reach US sites through clinical trial work even before it pursues US clinical clearance.

The pattern established elsewhere in this lane is that European vendors build a United States compliance posture when United States customers require it rather than in advance, so the absence here reflects current market scope rather than a settled position. A buyer in the United States should ask whether Waiv will sign a business associate agreement, and separately whether any test is available in the United States at all, since CE marking does not confer that.

No security certification, attestation or trust centre was located for Waiv in its own right. The site's legal section carries terms of use, a cookie policy, a privacy policy, a patient information page and published instructions for use, and nothing on information security. This grade requires a note about corporate history because the obvious inference would be wrong.

Waiv was spun out of Owkin as an independent company in March 2026 with $33 million in financing, having previously operated inside Owkin as its diagnostics division. Owkin holds ISO 27001 certification and maintains a trust centre. Those belong to Owkin. A newly separated company does not inherit its former parent's certifications by operation of the spin out, and this index grades the posture a buyer can actually read and contract under, so nothing has been imported here.

That is a statement about what is published rather than a judgement about the underlying engineering, which may well be unchanged from the day before separation. It does mean a buyer has no independent assurance to rely on at present.

The practical questions are specific and a buyer should put them directly: does Waiv hold information security certification in its own name, is it operating under a transitional services arrangement with Owkin and if so what does that cover and when does it end, and what is the timeline to certification as a standalone entity. Worth revisiting within a year, since certification in the new entity's name is the normal next step after a spin out of this kind.

Two CE-IVD approvals granted September 2022, both reported as first in class in Europe: the first CE-IVD digital pathology AI predicting breast cancer relapse risk, and the first CE-IVD AI identifying microsatellite stable colorectal patients from routine histology. No FDA clearance was located.

One detail a buyer should catch: independent survey work indexes RlapsRisk BC as CE-IVD General under the older IVDD directive rather than under IVDR, the same distinction flagged against Visiopharm's DeepBio-supplied prostate APP, and IVDD carries a materially lower evidence bar than IVDR. Graded C on the established basis of real European authorization without US clearance.

The vendor discloses its cohort enrichment openly, which is a form of honesty about validation limits that many competitors omit, and blind single shot validation on independent cohorts guards against overfitting. What is absent is demographic subgroup analysis, which carries specific weight here because the company markets RlapsRisk BC as levelling access to precision testing for patients who cannot reach gene expression testing. A tool justified on equity grounds should demonstrate it performs equitably, and that evidence was not located.

Two methodological choices here are better than the lane norm and both are stated rather than implied. The validation is described as conducted blind and single shot on independent cohorts of over a thousand slides, following training on a separate cohort of well over a thousand patients.

Single shot is the important word: it means the test set was used once rather than iterated against, which is the difference between a validation and a tuning exercise, and most vendors never say which they did. And performance is reported as a concordance index for the score combined with clinicopathological factors set against those factors alone, which quantifies the incremental contribution rather than the model's standalone accuracy.

That framing is the honest one for an adjunct test, because the question a clinician actually faces is not whether the model has predictive value in isolation but whether it adds anything to the information they already have, and a standalone figure would look far more impressive while answering a question nobody asks.

Held below the top grade because architecture detail and per subgroup performance are unpublished, equivalent methodological detail was not located for the company's other test, and no warranty, indemnity or remediation commitment attaches. Per subgroup matters for a prognostic score, since the incremental contribution may be concentrated in patients who were already clearly high or low risk. Ask for the incremental gain in the intermediate group specifically.

Delivery runs through pathology laboratory workflows rather than the EHR, with integration into routine digital pathology pipelines pursued via the Cerba Path partnership across that network's laboratories. No published connector list, scanner compatibility matrix, or platform integration detail equivalent to Primaa's multi-platform disclosure was located, so integration depth is asserted through partnership rather than enumerated.

Two routes are offered and the second is the more interesting one, but neither is evidenced in deployment. Destra is a common platform through which all of Waiv's tests are accessed: slides are uploaded singly or in batches, whole slide images are viewed and navigated in the browser, multiple AI products run against them, results are reviewed with interactive visualisations, and cases are organised into folders and monitored.

Results are exportable as PDF, CSV and JSON, which is a small but genuine interoperability property, since a laboratory can move outputs into its own systems without a bespoke integration. The alternative route is a Destra API allowing the tests to be integrated into a laboratory's existing image management system rather than requiring staff to work in a second application, which is the arrangement most laboratories will want.

The company also states that Destra hosts compatible software AI products including software as medical devices, so the platform is designed to carry more than its own portfolio. What is missing is the deployment evidence itself. No customer institution or reference site was named, no implementation timeline or onboarding process is described, no scanner compatibility or image format support is enumerated, and no laboratory information system is named for the API route.

On residency, nothing states where uploaded images are processed or stored, which matters because the platform receives patient tissue images by design. For a company four months into independent operation, thin deployment disclosure is unsurprising, and this is the axis most likely to move first as customers are announced.

No rates are published, but the commercial case is more legible than most because both products define themselves against a priced comparator. RlapsRisk BC is positioned against gene expression tests the company describes as expensive and not always accessible, and MSIntuit CRC against the volume of confirmatory MSI testing it displaces.

In both cases a buyer can model value against a known line item, the cost of the molecular test avoided or deferred, rather than against an abstract efficiency claim. That is the same reasoning applied to Stratipath.

Two indications, each narrowly specified: ER positive HER2 negative early invasive breast cancer post surgery for relapse risk, and colorectal cancer for MSI pre-screening. Both are high volume decisions where the alternative is expensive molecular testing, which is a coherent strategy rather than scattered coverage. But scope is confined to those two tumour types and, for the breast product, to one receptor subtype and one treatment context. The company states intent to expand biomarker pre-screening and outcome prediction across further disease areas.