AI in Medical Devices: Separating Hype from Evidence

← Back to All Insights

"AI-powered" has become a common line in medtech pitches, and the label says little on its own. What matters in diligence is what a device's regulatory record, validation data and deployment show, and whether the business around it can last.

The AI Medical Device Landscape

FDA keeps a public list of AI-enabled medical devices, and it is long: a 2025 study in npj Digital Medicine built a taxonomy across 1,016 authorizations, and one industry tally counted 295 AI/ML 510(k) clearances in 2025 alone. Most of these devices use medical images as their main input. Authorization volume says little about any one company, though. A rough way to group the landscape:

Category	Examples	What to Check
Diagnostic AI	Radiology triage, pathology analysis	Multi-site validation, reader studies, fit with the reading workflow
Clinical Decision Support	Risk scores, treatment recommendations	Whether it is a device at all under FDA's January 2026 CDS guidance, and how clinicians review its basis
Monitoring and Prediction	Early warning systems, deterioration detection	Alert burden, prospective performance, who acts on the output
Administrative AI	Documentation, coding assistance	Whether it is regulated as a device, and what the buyer pays for

What Makes an AI Medical Device Defensible?

1. Training Data That Is Hard to Replicate

Durable AI companies often have training data competitors can't easily assemble, for example from:

• Exclusive partnerships with health systems
• Unique hardware that generates novel data types
• Historical datasets with outcomes data
• Domain-specific annotation expertise

2. Clinical Validation Beyond FDA Authorization

Most AI devices reach the market through a 510(k), which shows substantial equivalence to a predicate rather than independent proof of clinical benefit. A 2025 systematic review in JAMA Network Open of 717 FDA-cleared radiology AI devices found that 97% were cleared through the 510(k) pathway, 29% reported clinical testing and 5% reported prospective testing. Commercial success usually needs more:

• Peer-reviewed publications demonstrating clinical impact
• Real-world evidence from deployment at scale
• Economic validation showing value for health systems
• Comparative studies against standard of care

3. Workflow Integration

A technically strong algorithm that doesn't fit clinical workflow tends to go unused. Signs of a better fit include:

• Deep integration with EHR systems
• Few added clicks or steps for users
• Clear triggers for when the AI should be consulted
• Actionable outputs, not just information

4. Sustainable Reimbursement

Payment for AI is still developing. Stronger positions tend to include:

• Category I CPT codes, not only Category III
• Coverage decisions from major payers
• Economic models that show value to payers and health systems
• Revenue models, such as licensing or subscription, that don't depend on a single payment code

Red Flags We Watch For

• "Our AI is better": without rigorous comparative studies, this is marketing, not evidence
• An indication chosen to avoid clinical data: an intended use narrowed mainly to limit the evidence required
• Single-site validation: performance at one institution doesn't reliably carry over to others, and a Stanford analysis of FDA-cleared AI devices found many had been tested at only one or two sites
• A crowded application: many cleared products already doing the same task
• No clinical owner: no specialty or use case that owns the product

What Stronger Opportunities Tend to Share

• A problem where human performance is genuinely limited
• Value that comes from speed or scale
• An established or emerging payment pathway
• A team with clinical domain expertise, not only machine-learning expertise
• Evidence that could support adoption into standard practice

Regulatory Changes to Track

FDA's final guidance on predetermined change control plans for AI-enabled device software functions, first issued in December 2024, lets planned model updates be reviewed with the original submission. Its January 2025 draft guidance on AI-enabled device software functions covers lifecycle management and what a marketing submission should include. And its January 2026 clinical decision support guidance changed which software functions fall outside device regulation.