The AI Health Index Brief

The week in one line

The FDA cleared a pathology platform along with its plan for future change, and URAC issued one of its first healthcare AI accreditations. The trust paperwork for clinical AI is multiplying into species. Telling them apart is now part of the job.

Theme 1: Clearance became a process

Proscia received a new FDA 510(k) clearance for its Concentriq AP-Dx digital pathology platform, and the interesting part is not the clearance. It is the Predetermined Change Control Plan that comes with it, alongside newly cleared cloud deployment and interoperability with the Leica Aperio GT 450 DX scanner.

Under the change control plan, laboratories can adopt future FDA cleared scanners and displays without waiting for Proscia to file additional 510(k)s. The agency did not just clear a product. It cleared a procedure for changing the product.

The classic kind of clearance had a good week too. Gleamer’s BoneView received 510(k) clearance for pediatric fracture detection, extending an adult tool to the population where missed fractures carry the longest consequences. And Overjet cleared IRIS, a real time image quality system that flags eight capture problems, cone cuts, overlapping contacts, missed coverage, the moment an X-ray is taken rather than after the patient has left. Quality control at the point of capture is the cheapest place it will ever happen. Healthcare has enough expensive ways to discover something went wrong later.

Then the paperwork arrived for the organizations rather than the devices. Guidehealth became one of the first organizations to earn the new Health Care AI Accreditation from URAC, an independent review of its AI governance, risk management, and trustworthiness practices. And Censinet built support for the new HHS Cybersecurity Performance Goals into its risk platform, so health systems can track themselves against the federal cyber floor without a spreadsheet.

Count the species: a device clearance, a cleared plan for changing a device, an organizational AI accreditation, and a federal performance goal framework. All of them will appear in vendor decks as the word “trusted.”

Our read

The assurance layer for healthcare AI is differentiating fast, and the Predetermined Change Control Plan is the piece buyers should study. It solves a real problem: models improve faster than 510(k) cycles, and a clearance frozen at submission describes a product you may no longer be running. But it also means the product you buy next year can lawfully differ from the one the agency reviewed, inside a plan you have probably never read. The clearance date on a spec sheet used to tell you what was evaluated. Increasingly it tells you when the evaluating started.

Buyer question

For every certificate, clearance, or accreditation in a vendor deck, ask three things: what exactly it attests, the product as shipped, the process for changing it, or the organization operating it; what version was actually reviewed; and how you will be notified when the thing it covers changes underneath it.

Theme 2: The foundation model got its journal day

Tempus published in Nature Medicine on PRISM2, the multimodal slide level pathology foundation model it developed with Microsoft. The model reads routine hematoxylin and eosin slides and predicts biomarker status and patient prognosis without specialized fine tuning for each task. That last clause is the claim that matters: one model, ordinary slides, many questions.

HOPPR expanded its vision language model portfolio with a 2D Mammography Narrative Model, a foundation layer for teams building breast imaging applications, extending the portfolio beyond chest X-ray and CT into women’s health. And Insilico Medicine published in npj Precision Oncology, using its PandaOmics platform to map therapeutic targets for a rare sinonasal carcinoma by integrating transcriptomic data with pathway biology.

Last issue closed on the quiet failure mode of pathology AI: models that perform beautifully at the institution that trained them, then degrade at the next site’s scanner. A slide level model claiming task generality is the same problem approached from the opposite end, with scale instead of robustification. The journal route is the right door to walk through with a claim like that. It is also only the first door.

Our read

Foundation models are entering pathology and imaging through the literature rather than the sales deck, which deserves credit, and the verification column reflects it. But a publication demonstrates performance; it does not define a deployment pathway. Between a Nature Medicine paper and a tool a pathologist can rely on sit external validation across sites and scanners, a regulatory route, and the unglamorous integration work this category is famous for underestimating.

Buyer question

For any foundation model claim, ask what external validation across sites, scanners, and stains exists beyond the flagship study, and what the deployment pathway actually is: a cleared device, a lab developed test, or research use only. The answer changes who is allowed to act on the output.

Theme 3: Findings started triggering workflows

Azra AI integrated with Blackford’s portfolio of more than 130 medical imaging AI applications, so a finding detected by any of them can automatically trigger patient navigation and follow up workflows in Azra’s platform. A second Azra integration, with RevealDx, routes a lung lesion characterization score directly into care coordination and survivorship workflows.

The radiology finding used to end in a report. It now opens a care pathway by itself. That is where “AI assisted” starts becoming operational rather than descriptive. Once a finding can cause another system to act, the interesting question is no longer only whether the model was right. It is what happens next when it is wrong.

Autonomize AI shipped the same idea for the payer side twice in one week: Genie AI, an agent that lets frontline teams assemble production ready utilization management and authorization review workflows from natural language prompts, drawing only on approved enterprise capabilities, governed data sources, and validated agents; and a payment integrity application for fraud, waste, and abuse review, built natively on ServiceNow’s healthcare data model and sold through its store. Hyro connected its patient facing voice and digital agents to ServiceNow workflow management the same week.

Note the design choice inside Genie: the agent composes workflows only from parts that were individually approved and governed. That is a vendor answering the composition question before being asked, and it is the right question. A collection of approved components does not automatically become an approved outcome simply because someone connected them with arrows.

Our read

Healthcare AI is wiring detection to action, and the governance surface is moving with it. When a model’s finding automatically opens a navigation workflow, a false positive stops being a statistic in a validation table and becomes a scheduled appointment, a worried patient, and a downstream cost. The composed workflow, not the individual model, is becoming the unit that needs review.

Buyer question

For every automated trigger between systems, ask what a false positive costs after the handoff, who reviewed the assembled workflow before it ran on real patients or claims, and whether the parts being composed were each approved, or only the platform composing them.

Market notes

The scribes kept expanding their job description. Sully.ai released a scribe built for nursing shifts, handling timestamps, flowsheets, and shift handoffs, which is a different documentation species from the physician visit note: continuous, structured, and handed off rather than signed off. Upheal added pre session preparation that surfaces client context before the appointment begins, so the scribe now starts working before anyone speaks. And Freed shipped dictation, live only on its Premier and Group tiers, a small reminder that last issue’s advice to price the tier you would actually deploy keeps earning its keep. Last issue the scribe volunteered for administrative privileges. This week it picked up the night shift and started arriving early.

The reference giants both went shopping for corpus. UpToDate integrated Lexidrug’s medication data into its Expert AI assistant, aimed at drug dosing questions the vendor says account for nearly 30 percent of clinician queries on the platform. OpenEvidence integrated Springer Nature’s journals and textbooks to ground its answers. The conversational layer of clinical reference is converging on similar interfaces; what diverges underneath is the licensed reading list. In this category the moat is starting to look less like the model and more like the library card.

Hinge Health acquired Cylinder Health, a virtual first digestive care company, for 105 million dollars, adding a gastrointestinal program beside its musculoskeletal and migraine offerings. Last issue Hinge published a falls trial; this issue it bought a condition. Evidence and acquisition are the two gears of the same platform strategy, and it is worth watching whether the evidence discipline travels to the new conditions. Buying adjacency is easy. Validating it remains irritatingly empirical.

Doctolib launched a clinical AI research program with public institutes including Inria and Inserm, developing AI tools using patient medical histories, with user health data included automatically unless the patient actively opts out via a form. The entry is recorded Partially Verified in the log. Whatever the final details, the structural point stands on its own: a consent default is a policy decision wearing a form. If your organization partners with platforms that touch your patients, the training default belongs in the contract, not the FAQ.

The imaging workbench kept automating. Circle Cardiovascular Imaging shipped cvi42 v6.5 with automated pre segmentation for 4D flow and editable plaque segmentation, and TeraRecon released a neurovascular package that turns CT and MRI perfusion into penumbra maps for stroke workups. Neither is a headline. Both are minutes, and stroke care is priced in minutes.

The action item: ask for the change control plan

If your organization runs FDA cleared imaging or pathology AI, add one question to the next vendor review: does your clearance include a Predetermined Change Control Plan, and what is inside it.

A vendor with a change control plan can update models and add cleared hardware without new filings, which is genuinely good for keeping pace. It also means the version you validated is not necessarily the version you will be running by renewal. Ask for the plan’s scope, ask how you are notified when something changes under it, and record the version you actually validated in your own files. The clearance is the agency’s record. The version history is yours to keep, because nobody else will keep it for you.