Of the 174,226 unique 510(k) clearances in the FDA's premarket notification archive we queried directly, 98.4% (1,491,956 of 1,515,722 indexed records) were decided "SESE" — substantially equivalent to a device already on the market. Denials ("DENG") account for less than 0.5%. The 510(k) pathway isn't really a gate that devices sometimes fail; it's a pathway that, once entered, almost never rejects. The interesting regulatory question isn't "will FDA say no," it's "what device gets compared to what predecessor, and for how many years that comparison keeps working."
The volume curve tells a story about administrative capacity, not just innovation
Annual clearance counts (by decision year, from FDA's own submission dates in the index) roughly tripled from 1996 (12,909) to 2025 (87,758) — but the shape is not smooth. There's a sharp jump starting exactly in 1996, coinciding with the FDA Modernization Act of 1997, which authorized third-party (accredited-lab) review for lower-risk devices and streamlined 510(k) processing. Volume then plateaus through the 2000s before climbing again after 2011, tracking the growth of combination and software-driven devices. The mechanism is procedural, not purely inventive: each expansion of eligible review tracks (special 510(k), abbreviated 510(k), and now the Safety and Performance-based pathway) adds throughput without changing the substantial-equivalence standard itself.
Where the volume concentrates: a long tail of narrow product codes
We ran a terms aggregation on product_code across 4,744 distinct codes. The single largest, IYN (dental — CAD/CAM-milled restorative and abutment components, per FDA's Product Classification Database), accounts for 48,224 chunked records on its own — more than triple the next largest code. That concentration reflects how digital dentistry (CEREC-style chairside milling — our sample includes numerous clearances in this family, such as K160519, "Link Abutment for CEREC," cleared 2016-10-28) generates enormous numbers of near-identical component clearances, each a minor iteration substantially equivalent to the last.
The predicate-chain caveat
This is also where the "always equivalent" statistic becomes a policy story rather than a compliance one. FDA's own count is that nearly 20% of current 510(k)s cite a predicate more than 10 years old, and a 2023 JAMA-published analysis found 44.1% of the 127 devices subject to the most severe Class I recalls (2017–2021) had themselves been cleared against a predicate that was also later recalled. Former FDA Commissioner Scott Gottlieb publicly proposed retiring reliance on decades-old predicates in 2018 — a proposal industry pushed back on, and which has not been enacted. Public 510(k) summary data doesn't itself trace individual predicate lineage, so this figure comes from FDA's public statements, not the clearance data itself — a limitation worth being explicit about.
Data limitations: the public 510(k) database yields 1,515,722 text chunks (average ~8.7 per submission, via a chunk_index/total_chunks pair) against 174,226 unique K-numbers — so raw document counts overstate submission volume roughly 8-9x; all figures above are stated as chunk-level shares (valid for rate comparisons) or de-duplicated by k_number cardinality where noted. The 2026 bucket (27,982) is a partial year and excluded from trend commentary.
Why this matters: as FDA leans further into AI/ML-enabled device clearances and predetermined change control plans, the predicate-chain question stops being an academic critique and becomes the actual mechanism by which software-driven risk propagates through a system built for incremental hardware equivalence.