Stratipath

A single deep learning model is the entire company. Stratipath Breast reads routine hematoxylin and eosin slides and returns a risk classification, with no viewer, laboratory, or assay of its own. The product exists only because the model extracts prognostic signal from images already produced in standard care, which is the purest expression of the computational pathology thesis in this index.

Produces a risk stratification rather than a diagnosis, which places it alongside the clinician rather than in the diagnostic path. The output is a low or high risk group assignment informing treatment intensity decisions, so the oversight question is less about diagnostic error and more about how much weight a clinician places on an algorithmic prognosis when deciding on chemotherapy. The company positions it as decision support, and the risk grouping format keeps interpretation with the clinical team.

Materially more transparent than most, because the validation is in the peer reviewed literature where methodology can be inspected rather than in a datasheet. The published work states the approach, deep learning whole slide image classification for prognostic stratification, the cohorts used, and the endpoints.

Training data composition and model architecture detail are not fully published, and there is no third party head to head benchmarking against molecular assays run by an independent group, but the primary validation is externally readable.

Strong on the provenance of the data the models were built from, silent on what happens to the data customers send, and those are different questions with this axis grading the second. The provenance position is better than most of this index and worth stating, because the company markets the scale and representativeness of its datasets and can actually account for them: development and validation ran on named national population cohorts and a university hospital series, drawing clinicopathological information from a national breast cancer quality registry, under stated ethical review approvals with the reference numbers published in the resulting papers.

A vendor that markets its training data and can name the ethics approval behind it, to the reference number, is doing something this index has repeatedly found absent, and it converts a marketing claim into something a reader can look up. What is missing is the operational half.

The product is delivered as software as a service, so customer slide images are transmitted to and processed by the vendor, and nothing retrieved states a retention period, a deletion position at the end of a contract, whether images are de identified before processing, or whether customer material may contribute to future model development. No processing agreement or privacy instrument was located.

The research data question is answered and the customer data question is not, and a buyer should not read the first as covering the second. Ask for the processing agreement, a retention and deletion schedule, and a written statement on model development.

Among the best evidenced single product vendors in this index, and the position has strengthened on the exact axis its earlier limitations named. The foundation is a retrospective multi site validation covering 2,719 primary breast cancer patients across two Swedish hospitals, assessing prognostic performance in two fully independent cohorts and published in the peer reviewed literature.

Two independent cohorts matter more than raw sample size, because prognostic models frequently fail to generalise beyond the population they were developed in. Two further pieces address what a retrospective Swedish record could not.

The model has been evaluated head to head against an established multigene assay in a real world series of clinically intermediate risk hormone receptor positive tumours, which is a comparison against the commercial standard of care rather than against no test, and is the comparison a laboratory weighing the cost of molecular testing actually needs.

And validation within TAILORx, one of the largest prospective randomised trials in early breast cancer, has been presented at a major oncology congress, which begins to answer the prospective question directly. That last item is recorded as conference presented rather than peer reviewed published, and should be confirmed in its published form before being relied on. What remains open is population breadth.

The published cohorts are Swedish, the head to head series is Swedish, and generalisation to different populations, care pathways and laboratory practices is still a reasonable question for a buyer outside the Nordics to raise, notwithstanding the prospective trial evidence now emerging.

Strong on the provenance of the data the models were built from, silent on what happens to the data customers send. Those are different questions and this axis grades the second. On provenance the position is better than most of this index and it is worth stating, because the company markets the scale and representativeness of its datasets and can actually account for them: development and validation ran on named Swedish population cohorts including a Karolinska University Hospital series and the SCAN-B study, drawing clinicopathological information from the Swedish national breast cancer quality registry, under stated Swedish Ethical Review Authority approvals with the reference numbers published in the resulting papers.

A vendor that markets its training data and can name the ethics approval behind it is doing something the index has repeatedly found absent elsewhere. What is missing is the operational half. Stratipath Breast is delivered as software as a service, so customer slide images are transmitted to and processed by the vendor, and nothing retrieved states a retention period, a deletion position at the end of a contract, whether images are de identified before processing, or whether customer material may contribute to future model development.

No data processing agreement or privacy instrument was located. The research data question is answered and the customer data question is not. Ask for the data processing agreement, a retention and deletion schedule, and a written statement on whether submitted cases can be used for model development.

No HIPAA position was located, and unusually for this lane the company has told a buyer why in advance. Stratipath states that Stratipath Breast is regulatory aligned for clinical use in the European Union and the United Kingdom, and its launch materials state plainly that the product is not marketed in the USA.

Where a vendor scopes itself out of a market explicitly, the absence of that market's compliance instruments is a disclosed consequence rather than a gap, and this note records it as such. That is the most candid of the four positions now visible among the European vendors in this lane, and the contrast is worth carrying: one European peer had its controls independently audited against the HIPAA Security Rule because United States laboratories required it, one operates a United States entity and names United States customers while publishing no HIPAA position at all, one simply has no United States presence, and this vendor says outright that it does not market there.

A United States laboratory should read this as the product not being offered to it rather than as a compliance weakness. If Stratipath enters the United States market, this axis needs regrading rather than reinterpreting, and the questions would then be the standard ones: business associate status, a business associate agreement, and the position under the Security and Privacy Rules for a service that receives patient slide images.

No security certification, attestation or trust centre was located. No ISO 27001, no SOC 2 of either type, no penetration testing statement and no vulnerability reporting route. No privacy policy, terms of use or security page was reachable from the company's homepage or found by direct search, which is itself unusual for a medical device vendor delivering a clinical product.

Scope note: the homepage may render legal navigation that was not captured on retrieval, so this records what could be located rather than proof that nothing exists. The absence matters more here than for several peers in this lane, and the reason is architectural rather than presentational. Stratipath delivers Stratipath Breast as software as a service, so patient whole slide images demonstrably leave the laboratory and are processed by the vendor.

Peers in this lane that publish little on security at least reduce the question: one states that all data stay inside the customer network, another runs inside the laboratory's existing image management system. Here the data goes to the vendor by design, which makes independent assurance over the vendor's own environment the central question rather than a secondary one, and nothing addresses it.

A buyer should ask for an ISO 27001 or SOC 2 position and timeline, the hosting arrangement and its location, the penetration testing history, and the incident notification terms, and should expect these in the contract rather than on the website.

Described in peer reviewed literature as the first CE-IVD marked AI based image analysis tool for primary breast cancer risk stratification available for routine clinical use, which is a category first rather than a me too certification. No FDA clearance was located.

Graded C on the established basis, real European authorization without US clinical clearance, with the note that a prognostic risk tool occupies different regulatory territory than a detection device and its US path would likely differ accordingly.

Validation across two independent hospital cohorts addresses site level generalization, which is the most common failure mode for prognostic models. No formal governance framework, monitoring commitment, or demographic subgroup analysis was located, and the validation population is drawn from two Swedish hospitals, so performance across more diverse populations is undocumented. This matters more than usual for a prognostic tool, since a risk score that behaves differently across populations translates directly into differences in treatment intensity.

The primary validation sits in the peer reviewed literature where the methodology can be inspected rather than in a datasheet, with the approach stated as deep learning whole slide image classification for prognostic stratification, the cohorts named, and the endpoints given.

For a prognostic product that combination is what makes external assessment possible at all, because a prognostic claim cannot be checked against the slide the way a detection claim can: a pathologist can look at an image and see whether a flagged feature is there, and nobody can look at an image and see whether a five year risk estimate was right.

The evidence has to come from cohorts followed over time, and publishing which cohorts and which endpoints lets a reader judge whether the follow up supports the claim. Held below the top grade on two points. Training data composition and architecture detail are not fully published, and there is no third party head to head comparison against the molecular assays this product competes with, run by an independent group.

That comparator is the relevant one, because a laboratory choosing this is usually choosing it instead of a molecular test with its own evidence base, and the question is not whether the image based score has prognostic value but whether it agrees with the established assay where they disagree matters. No warranty, indemnity or remediation commitment attaches. Ask for concordance with the molecular assays on shared cases, and performance in patients unlike the development cohorts.

Reach comes through platform distribution rather than direct integration engineering. Since 2025 the product is available through PathAI's AISight platform, which is separately indexed here and itself cleared for primary diagnosis in the US and CE-IVD marked in Europe. For laboratories on that platform, adoption is algorithm selection. A broader connector list for other image management systems was not located, so integration beyond partner platforms is undocumented.

The delivery model is stated without ambiguity, which is creditable, but almost nothing else on this axis is. Stratipath Breast is offered as software as a service on a subscription or pay as you use basis, either through integration with an existing digital pathology workflow or standalone through the company's own customer web portal. One integration route is named, with the product listed on the Sectra Amplifier Marketplace and purchased directly from Stratipath.

Geographic scope is stated plainly: clinical use in the European Union and the United Kingdom, and not marketed in the United States, with commercial activity reported as concentrated in the Nordic region. Being explicit that the product is delivered as a service is more useful to a buyer than it may sound, because it answers the question several peers in this lane leave open about whether patient images leave the laboratory. Here they do.

Held at C because the rest of the axis is unaddressed. There is no data residency commitment, no hosting provider or region named, no statement of where analysis is performed, and no named customer deployment or reference site was located. No implementation timeline, resourcing expectation or onboarding description is published. For a service based product the residency and hosting questions are the ones a laboratory's information governance team will ask first, and a buyer will have to obtain the answers directly.

No pricing is published. The commercial logic is nonetheless legible from the clinical positioning: the published rationale explicitly contrasts the tool against molecular multigene assays, which carry long lead times and high cost, so the value case is turnaround time and price relative to those tests. A buyer can therefore benchmark against a known comparator even without published rates, which is more than most vendors here allow.

A single indication, primary breast cancer risk stratification from surgical resection specimens. That focus is the source of its evidence depth, but coverage does not extend to other tumor types, to biopsy specimens, or to diagnostic as opposed to prognostic tasks. Buyers needing breadth will pair it with other tools rather than adopt it as a platform.