How to find AI device predicates in the FDA 510(k) database

Every 510(k) clearance depends on a predicate. For AI-enabled medical devices, finding the right predicate is both the most important step in your clearance strategy and, in practice, one of the hardest pieces of research to do well. FDA's public database gives you the raw material. What it doesn't give you is a working method for using it.

This post covers how to actually find AI device predicates — what the database contains, where it falls short, and how to reconstruct predicate lineage that FDA never explicitly publishes.

What a predicate actually does

A 510(k) clearance requires you to demonstrate substantial equivalence to a previously cleared device — the predicate. For an AI device, the predicate establishes three things: the intended use, the technological characteristics (including the algorithm type), and implicitly, the performance standards FDA has already accepted for that device type.

Choosing the right predicate is not just a paperwork exercise. A predicate that closely matches your intended use and algorithm type makes the substantial equivalence argument straightforward. A predicate that's a poor match creates exactly the kind of substantial equivalence questions that generate additional information requests and extend review timelines.

Key AI device product codes for predicate search

The four-step method for predicate research

Step 1: Identify your product code

Start with product code, not keyword search. FDA's product codes define device categories, and clearances within the same product code share a regulatory history that makes them the strongest predicate candidates.

If your device doesn't fit cleanly into an existing product code, that's important information — it may indicate a De Novo pathway is more appropriate, or that you're working in an indication where FDA has accepted clearances under a more general code.

Step 2: Pull all clearances in your product code and filter by indication

Once you have a product code, filter the 510(k) database for all clearances in that code. For high-volume codes like QMF, this will return hundreds of results. The next filter is indication — you want devices cleared for the same intended use as your device, not just the same broad product category.

A practical shortcut: sort by date descending and read the 20 most recent clearances in your product code. Recent clearances are stronger predicates because they reflect FDA's current expectations.

Step 3: Read the 510(k) summary documents for your top candidates

For your top five to ten predicate candidates, read the 510(k) summary documents. You're looking for the full predicate list, the substantial equivalence discussion, the performance data FDA accepted, and whether a PCCP was included.

Step 4: Trace the predicate chain back

Once you've identified a strong predicate candidate, trace its own predicate chain. FDA's clearance history for any given device type is a lineage — each cleared device cited a predicate, which itself cited a predicate, going back to the device type's first clearance or De Novo grant.

Where the database falls short

FDA's database has three significant gaps for AI predicate research: no indication taxonomy (indication filtering requires reading, not filtering), no algorithm type tagging in the main database, and no predicate relationship data — the graph of what cites what isn't published. Going forward in the lineage requires searching full text of summary documents.

For serious predicate research in competitive indications, manual database work is time-intensive enough that most teams either underinvest or spend significant consultant hours on research that should be automatable.