AI Diagnostics: Proof Over Promise

The AI Diagnostic Marketplace, As it Is
[Not as the Pitch Decks Wish It Were]

Patient and doctor interaction is critical to the development of successful diagnostic platforms; its absence is delaying their adoption.

By Michael Schmanske with Levan Gogichaishvili, MD, PhD

Overall, the U.S. FDA has now authorized hundreds of AI-enabled medical devices, with radiology still the gravitational center. The agency’s public list shows an accelerating cadence of 510(k)s and De Novo applications across imaging, cardiology, ophthalmology and more. Radiology-heavy updates through the summer of 2025 pushed authorizations near the 1,000 mark, underscoring that AI is no longer “adjacent”—it’s embedded in mainstream device pipelines.

However authorizations are not adoption. And the delay resides primarily in real-world application and benefits for a number of reasons, but economics usually tops the list. To spur adoption, clinical practices must experience improved outcomes and input costs, but they also need security around compensation. For example, insurance practices and coverage payment drags the real-world curve. 

In the United States, reimbursement for diagnostic AI still runs through three uneven channels: 

1. Status-quo fee-for-service via Category I CPT codes (rare for AI); 
2. Category III tracking codes (common, but coverage is discretionary); and 
3. Exceptional programs like CMS’s New Technology Add-on Payment (NTAP) for inpatient use. 

Yet many of these channels themselves are not particularly stable. The last item, NTAP, successfully created a temporary bridge for stroke triage software (e.g., Viz LVO) with per-case add-ons—making it a useful precedent, but not a general solution. And by design NTAP sunsets, forcing a second transition to durable coverage. 

Translation: many “FDA-cleared” AIs remain unfunded experiments unless hospitals absorb costs or negotiate one-off contracts. (Details: CMS.gov)

Fewer Grey Zones

Meanwhile, a sizable slice of clinical “AI solutions” in hospitals are never even touched by the FDA because they fit (or try to fit) within clinical decision support carve-outs or “wellness” claims. That regulatory ambiguity is shrinking. The FDA has started to police overreach in wearables and “AI health” apps with warning letters—and more will follow. The bottom line for the market: Regulatory authorizations are up; durable payment is patchy; enforcement is getting sharper; and buyers have learned to ask harder questions.

Policy Tailwinds That Actually Matter Right Now

Three shifts are worth our attention because they directly affect adoption speed, lifecycle updates and go-to-market strategies.

1) Predetermined Change Control Plans (PCCPs) are real, and they’re here.
This is the essential unlock for iterative models—plan the permitted model updates (data scope, retraining triggers, performance bounds) in advance, then update without redoing a full submission each time, within that plan. (PCCP details: FDA)

2) eSTAR and e-submissions reduce avoidable friction.
De Novo requests now route through eSTAR (electronic templates) hopefully eliminating  format hiccups which typically delay review. It won’t rescue a weak clinical package, but it shortens the dumb parts of time-to-decision. (De Novo details: FDA)

3) Real-world performance (RWP) is moving from buzzword to task list.
FDA requested input this fall on how to measure and monitor AI devices in the wild—detecting drift, reporting performance changes, managing updates. Expect RWP plans and dashboards to become a de facto requirement for serious buyers. (Request for Public Content details: FDA)

Recent Successes

Considering the glacial pace of institutional and bureaucratic evolution, it is remarkable that as many companies have managed to navigate the regulatory unknown as they have.  These brave pioneers have established their individual value proposition to users and investors and present a useful model for other entrepreneurs to follow.

That adds up. It also explains why the “AI replaces radiologists” narrative died while the “AI tools everywhere in radiology” reality arrived.

Where the Wheels Came Off

While some startups successfully threaded the gauntlet, others got caught up on evolving regulations and ambiguous efficacy. That is to be expected in a new and developing field of innovation. However savvy founders should pay heed to the ways these companies failed and educate their own strategies accordingly.

When “wellness” crosses into diagnosis, FDA bites.
WHOOP’s wrist-based Blood Pressure Insights promised daily BP estimates without clearance. FDA called it what it is: a medical device marketed without authorization—adulterated and misbranded. 

• The broader point isn’t about WHOOP alone. It’s about the collapsing daylight between consumer wearables and clinical diagnostics; blood pressure is inherently medical, and once you imply accuracy and health impact, you’re in device land. Expect more of this. (WHOOP details: FDA)

SeniorLife.AI and Exer Labs: diagnostic claims without a regulatory base.
SeniorLife’s mobile app marketed fall-risk and cognitive screening (including Alzheimer’s) without authorization. Exer Labs touted screening/diagnostic functionality for “Exer Scan.” Many of these products live entirely as software, and their developers are not always aware of the difference between healthcare and consumer products.  

• Be prepared to provide efficacy data in line with a clinical device.  Both companies above earned FDA warning letters—adulterated devices lacking PMA/IDE or 510(k)/De Novo. (SeniorLife details: FDA)

Evidence light → recalls and mistrust.
Analyses of FDA records and trade-press data keep landing on the same conclusion: AI devices cleared via 510(k) often lack prospective human validation at the time of clearance; those with sparser premarket evidence are more likely to be recalled later. That corrodes clinician confidence and slows uptake even for good tools. 

• Developers read this as a warning: If you can run a pragmatic, prospective study, do it. It pays for itself in adoption velocity. (AI devices and FDA clearance details: Radiology Business)

A cautionary AI-monitoring example outside imaging: sepsis alerts.
Hospitals deployed sepsis prediction/alert systems for years with mixed results. External validations found poor discrimination and calibration for widely used proprietary models, while newer evidence suggests mortality benefits may flow more from broad workflow effects than from per-patient model precision. Regardless of which camp you favor, regulators and buyers now expect transparent evaluation, monitored drift and guardrails against alert fatigue. That expectation is bleeding back into diagnostic AI. (Validation of sepsis prediction model details: JAMA)

Success Patterns vs. Stalled Efforts

We’ve alluded to the differences, but let’s get specific. What separates the liver-ultrasound and stroke-triage successes from the apps that took warning letters—or the imaging tools that clear but never scale?

Intended use clarity vs. hand-wavy claims.
Winners define a narrow, clinical intended use, tie it to guideline-recognized tasks, and avoid implication creep. Velacur ONE: quantify liver stiffness and fat fraction in chronic liver disease patients at point of care. Viz LVO: triage suspected large-vessel occlusion to accelerate treatment. 

• Contrast that with “wellness” products that still claim diagnostic-grade insights; the former were approved, the latter get enforcement.

Evidence that matters to payers vs. retrospective convenience.
Retrospective AUCs get you marketing copy. Prospective or quasi-experimental designs that show time-to-treatment, length-of-stay or hard outcomes unlock coverage (or NTAP). 

• Reimbursement pathways reward impact on care delivery, not just image-level accuracy. Expect AMA and CMS coding to keep nudging in that direction (Category III codes proliferating; a slow pipeline to Category I).

Workflow first vs. “API in search of a clinic.”
POCUS devices with AI overlays are useful because they shorten training curves and let non-specialists hit standard planes quickly. Likewise, stroke triage runs in the background and pings teams—no extra clicks, immediate routing decisions. 

• Tools that force new screens or demand bespoke staffing struggle post-pilot. The FDA’s list shows growth in embedded AI within existing scanners and viewers—a sign vendors learned this lesson.

Lifecycle plan vs. one-and-done model.
Teams that file with a PCCP (or at least architect for one) can update models safely as populations or scanners shift. Teams that treat clearance as “we’re done” will drift and eventually underperform, triggering distrust or recall risk. 

• The regulator is telegraphing this: show your monitoring and update plan. Buyers are starting to demand it in RFPs.

Economic story vs. accuracy story.
Velacur ONE’s emphasis on B-mode (higher reimbursement) and MRI-PDFF correlation is not an academic flourish—it’s a business narrative clinics can underwrite. Viz’s time-to-treatment gains mapped to an NTAP request. 

• Precision is table stakes; economic utility closes deals. (Latest news: Diagnostic Imaging)

At the end of the day, technological adoption is about trust and familiarity. Improved oversight and changes in regulatory behavior address trust issues, but the real trust component relates to business use cases—and just like clinical outcomes, economic outcomes need data and proof as well. With that comes use and with use comes familiarity.

Concrete Contrasts (Company vs. Company)

A note on risk and recall:
Developers sometimes race to market with retrospective studies and a friendly predicate. It clears. It demos well. Then real-world variability (scanner configs, patient mix, site workflows) exposes edge cases. Recalls follow, and the entire category pays the price. The literature linking lighter premarket evaluation to higher recall risk isn’t perfect, but it is persistent enough that hospital committees have absorbed the message. (Validation data details: Radiology Business)

Real Talk, No Cheerleading

Diagnostic AI is maturing, but adulthood looks less like an AI revolution and more like a thousand narrow, regulated tools that quietly make clinicians faster or more consistent—and only the ones with proof and payment stick. If you want to pick winners, follow the evidence, the workflow and the money; if any of those three are hand-waved, assume the adoption curve bends to zero.