Routine Health Data Is Fueling AI Devices Faster Than FDA Can Define the Standard

The evidence base just shifted under sponsors’ feet

For a decade, AI-enabled device makers built validation packages around curated trial data because that was the evidence FDA trusted. That assumption is now outdated. In December 2025, FDA updated its Real-World Evidence guidance for medical devices, expanding how both agency staff and sponsors can use real-world data to support regulatory decisions, including cases where RWE can substitute for evidence that once required a dedicated trial arm, according to IQVIA’s analysis covered by Clinical Trial Vanguard. For AI training specifically, this matters more than it sounds. Routine health data, the messy, high-volume exhaust of everyday clinical encounters, is proving to be a better training substrate than the smaller, cleaner datasets trials produce.

That is a genuine capability upgrade. It is also a governance vacuum. The guidance tells sponsors they can use this data. It does not yet tell them how to document its provenance, representativeness, or bias profile in a way that will hold up under premarket review or postmarket surveillance.

The rest of the agency hasn’t caught up either

This gap is not isolated to one guidance document. FDA has issued a run of draft guidances on AI device and software evaluation even as the current administration’s deregulatory posture pushes the opposite direction, per Medical Device Network’s review of governance tenets for AI imaging. Sponsors are getting expanded permission to use real-world data at the same time the review infrastructure meant to evaluate that data is still being drafted piece by piece.

Capacity is the other half of the problem. The draft MDUFA VI commitment letter, now in negotiation, sets the user-fee funded staffing and review timelines that will determine how fast FDA can actually process RWE-trained AI submissions, according to the Bipartisan Policy Center’s breakdown of the draft letter. A sponsor building a submission strategy around routine health data today is making a bet on review capacity that won’t be finalized for months.

What this means for regulated buyers

Sponsors developing or procuring AI-enabled devices should not wait for FDA to formalize a data-quality standard for RWE-trained models. The prudent move is to build the documentation FDA will eventually require anyway: dataset lineage, demographic representativeness, known gaps in the routine data source, and a defensible rationale for why that data supports the specific claim being made. Treating RWE guidance as a green light without that documentation is the fastest way to end up with a device that trained well but cannot get through review.

There is a parallel worth watching for multinational programs. The EU’s MDCG 2025-9 guidance introduces a Breakthrough Devices framework under MDR and IVDR, offering expedited pathways for qualifying devices, per Jones Day’s digital health law update. Sponsors running parallel US and EU strategies now face two regulators moving toward more flexible evidentiary standards on different timelines, with different documentation expectations attached.

The capability curve for AI training data has bent sharply toward routine health data. The compliance curve has not caught up, and it will not catch up on a sponsor’s schedule. Build the provenance file now, not after the first RWE-based submission gets a deficiency letter.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. Core argument is coherent and the regulatory timing mismatch is real, but the central claim that routine health data is ‘outperforming’ trial data for AI training rests entirely on a single source (Cl
Source & Claim Verification	Qwen · local	cleared. Most factual claims are supported by citations, but a few lines lack specific references, such as the assertion about the capability upgrade of routine health data and the parallel worth watching for
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately reflects FDA and EU MDR/IVDR developments but lacks explicit alignment with ISO 42001’s structured risk management and documentation requirements for AI systems.
Technical Accuracy	Llama	cleared. The article accurately conveys the shift towards using real-world data for AI device validation and the challenges it poses for FDA regulation, but could be improved with more technical depth on AI an
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively identifies and addresses potential vendor hype by focusing on the regulatory gaps and challenges despite the perceived ‘capability upgrade’ of routine health data for AI train
Novelty & Non-Duplication	Grok	held. Core thesis, framing, and even title language closely duplicate the Clinical Trial Vanguard/IQVIA wire piece it cites, with the rest a standard multi-source synthesis rather than net-new reporting or
Validation	DeepSeek	cleared. The central claim that routine health data is outperforming trial data for AI training is presented as a given by the cited analysis but lacks direct, adversarial validation against contradictory evid

Sources cited: 8. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.