Armghan Ans: The Stroke Decision Nobody Owns

Armghan Ans, Assistant Professor and Director of Stroke at UPMC Washington, Founder of MDAdopt, shared on LinkedIn:

”Somewhere right now, a stroke AI tool is running clinical logic that was written before January 26th.

The guidelines changed.

The logic did not.

Nobody flagged it.

Not because nobody cares.

Because nobody owns that job.

This is the governance gap in stroke AI. And it is the topic of the latest edition of TAIN.

Here is what the newsletter covers:

• What clinical logic actually is
• The three places ownership usually lives
• Why the ownership gap is getting worse, not better
• What good ownership actually looks like

The stroke AI industry is entering a new phase.

The first generation competed on detection accuracy.

The second will compete on clinical trustworthiness.

And clinical trustworthiness starts with knowing who owns your clinical logic.

Guidelines got more nuanced. AI got faster. Responsibility stayed undefined.

Somewhere right now, a stroke AI tool is running clinical logic that was written before January 26th.

The guidelines changed. The logic did not.

Nobody flagged it. Not because nobody cares. Because nobody owns that job.

That is not a hypothetical. It is the default state of most stroke AI products in the market today.

And it is the question this edition of TAIN is about.

When clinical evidence changes, who is responsible for making sure the AI tools reflect that change?

Not in theory. In practice. With a name attached to it. With accountability.

With a process that runs faster than a full product cycle.

At most stroke AI companies, if you ask that question directly, there is no clear answer.

Not because the companies are irresponsible. Because the role does not formally exist.

The AHA has signaled a move toward a living guideline model with more frequent iterative updates.

AI stroke tools are no longer peripheral alerts sitting at the edge of workflow.

They are embedded in triage decisions, transfer calls, treatment eligibility, and post-procedure monitoring.

The clinical logic inside them is load-bearing.

When load-bearing logic is wrong, the consequences are not a missed notification. They are a missed treatment or a wrong one.

The stroke AI industry has a governance gap. And until it is named, it cannot be fixed.

What Clinical Logic Actually Is

When most people in health AI hear the phrase ‘clinical logic,’ they think of something technical. Thresholds. Parameters. Code.

A configuration file somewhere in the product architecture that determines what fires and when.

That framing is incomplete. And the incompleteness is where the problem begins.

Clinical logic is not a technical artifact. It is a clinical judgment encoded into a product.

Every alert threshold, every contraindication classification, every treatment window calculation started as a physician’s understanding of the evidence.

At some point in time, that understanding was translated into a decision rule.

A rule that now executes in code, firing alerts and influencing decisions without a physician in the room.

That distinction matters enormously. Because it means that when the evidence changes, it is not a software problem.

It is a clinical judgment problem that then becomes a software problem. The chain runs in that order.

A human being with current clinical knowledge has to recognize that the judgment needs updating before any engineer can update the code.

The 2026 guidelines changed what the evidence says about medium vessel occlusions.

That change did not automatically update the alert logic in any product.

It required a physician to read the guideline, recognize the delta, understand its implications for the product, and initiate a review. Only then could the engineering work begin.

At most stroke AI companies, that chain has no formal owner.

The physician who originally informed the logic may not even still be advising the company. The clinical advisory board meets quarterly at best.

Nobody is monitoring guideline publications and triaging their implications for product logic in real time.

The code is maintained. The clinical judgment underneath it is not.

That is the governance gap. Not a technical gap. A human accountability gap sitting one layer beneath the technology.

The Three Places Ownership Usually Lives, And Why None of Them Work

Every stroke AI company has clinical advisors. Most have a clinical advisory board. Some have a medical director.

The ownership of clinical logic feels like it should be obvious.

It is not.

The product team

The product team owns the code. They do not own the clinical judgment. They can update a threshold when told to.

They cannot audit that threshold against a new guideline without clinical oversight.

And they are running product cycles and engineering sprints, not reading the Journal of Stroke at 6am.

The clinical advisory board

The clinical advisory board is consulted at launch and at major version releases.

They are not monitoring guideline updates on a rolling basis. The 2026 guidelines published on January 26th.

Most clinical advisory boards did not convene an emergency session in February. That is not a criticism.

It is a structural reality. Advisory boards are designed for product direction, not clinical surveillance.

The hospital

The hospital is the customer. They trust the vendor to keep the product current. They do not have access to the vendor’s codebase.

They cannot update alert logic themselves. They are the last line of defense and the least equipped to be it.

None of these ownership models produces a reliable update cycle. Each one assumes someone else is watching.

The product team assumes the clinical board flagged it. The clinical board assumes the product team is monitoring it. The hospital assumes the vendor handled it.

Nobody handled it.

That is not negligence. It is a gap that the industry has never had to formally close before. Until now.

Why This Gap Is Getting Worse, Not Better

This is not a static problem. Two forces are converging that make the governance gap more consequential every year.

Guideline acceleration

The AHA signaled a move toward a living guideline model at ISC 2026. More frequent iterative updates are coming.

The 2026 guidelines replaced a document that was seven years old. The next update will not wait seven years.

The interval between guideline publication and product update needs to shrink. Under current ownership models it will not.

Product depth

Early stroke AI tools were peripheral. An alert fired. A physician looked at it. The physician made the decision.

The tool was a notification layer, not a decision layer. Clinical logic errors were visible and interceptable.

Current and next-generation tools are different. They are embedded in triage workflows, transfer decisions, treatment eligibility assessments, and post-procedure monitoring.

The clinical logic is no longer at the edge of the decision. It is load-bearing infrastructure.

When peripheral logic is wrong, a physician catches it. When load-bearing logic is wrong, the physician may never see the error at all.

The alert does not fire. The patient does not get flagged. The pathway does not activate.

Two forces. Both are accelerating in the same direction. Guidelines are updating faster. Products are going deeper into the workflow.

The governance gap is growing wider every year it goes unaddressed.

What Good Ownership Actually Looks Like

This is not an unsolvable problem. But solving it requires treating clinical logic governance as a deliberate organizational capability, not an afterthought handled informally between product releases.

Three components make it work.

1. A named clinical owner

Not an advisory board that meets quarterly. A specific physician with a defined mandate and a supporting advisory team.

The physician owns the accountability.

They are the named clinical logic lead, the person who reads the guideline updates, recognizes the delta, and initiates the review.

The advisory team provides the clinical depth across territories the lead cannot cover alone.

Given that the AHA has signaled a living guideline model with more frequent updates, this is not a part-time role.

The velocity of the evidence base now requires a physician whose primary function is clinical logic currency.

2. A surveillance mechanism

A defined process for monitoring guideline publications and triaging their product implications.

Not a full product cycle is triggered every time something changes.

A rapid review protocol, lighter and faster, specifically designed to answer one question: did this guideline update change any clinical logic currently running in our product?

If yes, how urgently does it need to change? That triage decision should take days, not quarters.

3. A documentation standard

Every clinical logic decision is recorded with the evidence basis at the time it was made.

Which guideline. Which trial. Which recommendation class? What the threshold is and why.

When the guideline changes, the delta is immediately visible. The update is targeted, not comprehensive.

The engineer knows exactly what to change and why.

These three components are not expensive to build. They are expensive to ignore.

Companies that build this do not just reduce patient risk. They build a durable commercial advantage.

Every time the guidelines update, they move faster than competitors. Hospital trust compounds. Renewal conversations become easier.

The clinical logic governance system becomes a selling point, not a liability.

The Deeper Question

The stroke AI industry is maturing.

The first generation of products competed on detection accuracy.

The algorithm was the product. Getting FDA clearance was the finish line. Clinical adoption was assumed to follow.

It did not always follow. And now the industry is beginning to understand why.

The second generation will compete on clinical trustworthiness.

Not just whether the algorithm is accurate, but whether the clinical logic behind it stays current as evidence evolves.

Whether the hospital can trust that what fired last January still reflects what the guidelines say today.

That trustworthiness is not a feature. It cannot be shipped in a product update.

It is an organizational capability built over time, through deliberate governance, through named accountability, through a surveillance mechanism that runs faster than the evidence moves.

The companies that build that capability now, before the next guideline update, before the first adverse event that traces back to outdated clinical logic, will have an advantage that no algorithm improvement can replicate.

Clinical logic is not a product feature.

It is a responsibility.

And right now, in most stroke AI companies, no one owns it.”

Stay updated with Hemostasis Today.