FDA AI/ML Device Predicate Networks

About This Dashboard

Who we are

This dashboard was developed by researchers at the Institute of Global Health Innovation, Imperial College London, as part of an ongoing programme of work examining how AI/ML-enabled medical devices are regulated by the FDA.

The dashboard covers the whole FDA AI/ML database as it stood at extraction: the primary predicate of every one of the 1,430 devices then listed has been sought in its individual FDA clearance summary and classified. 54 entered without a predicate, through the De Novo or Premarket Approval pathways, and for 8 the clearance summary was not available in the FDA archives. This yields substantially richer data than automated extraction from the FDA AI/ML device database alone, where predicate relationships are not recorded in any machine-readable field.

Alongside the whole database, we provide clinically mapped datasets for individual specialties, derived from a series of systematic scoping reviews conducted by our group. The first of these, covering cardiology AI devices, has been published, with further specialties to follow.

How the FDA approves AI devices

The FDA uses three main regulatory pathways:

1. 510(k) Premarket Clearance: the most common route for AI devices. The manufacturer demonstrates substantial equivalence to a predicate device. Does not necessarily require new clinical trials.
2. De Novo Classification: for novel, low-to-moderate risk devices with no suitable predicate. Requires more evidence than 510(k) but less than PMA. Once granted, the device itself becomes a predicate for future 510(k) submissions.
3. Premarket Approval (PMA): the most rigorous pathway, reserved for high-risk devices. Requires clinical evidence of safety and effectiveness. Rare for AI devices.

Why does this matter?

Although the FDA publishes an online database of AI/ML-enabled medical devices, the regulatory relationships between devices are not directly visible. The majority of AI devices are cleared through the 510(k) pathway, meaning each one is linked to an older predicate device that it was deemed substantially equivalent to. A new AI device may therefore be cleared based on similarity to an older device that may itself not be an AI device at all. This dashboard was built to make these chains of equivalence visible and explorable, increasing transparency around the regulatory foundations of AI medical devices.

FDA Lead Specialty vs Mapped Specialty

This dashboard classifies devices by specialty in two independent ways.

FDA Lead Specialty is the official medical specialty assigned by the FDA in their AI/ML device database, based on the device’s technology and intended use. Radiology accounts for 77% of all FDA-listed AI/ML devices under this classification.

Mapped Specialty is an additional classification from our research group. As part of our scoping reviews, we selected a number of clinical specialties and mapped devices from the FDA database to them based on their relevance to each field. Some devices may be relevant to more than one specialty, and the mapping is intended to help assess the AI device landscape from the perspective of different clinical fields rather than to replace the FDA’s own classification. For example, a radiological triage tool for detecting intracranial haemorrhage is classified by the FDA under Radiology, but may also be relevant to neurosurgical and stroke care.

Both filters are available in the dashboard, and the device detail panel displays both classifications for each individual device. In the complete database view, devices that have not yet been mapped to one of our specialties can be isolated using the Not mapped option.

Key observations

Across the database, 519 devices (36.3%) are classified as citing a non-AI predicate (terracotta dots), meaning their regulatory clearance rests on equivalence to a device that sits outside the recognised AI/ML landscape. A further 3.8% entered without a predicate at all, through the De Novo or Premarket Approval pathways.

The network is highly concentrated. 60 devices each serve as the predicate for three or more subsequent approvals, and the largest single family traces back to ContaCT (DEN170073), which anchors 116 downstream devices across multiple manufacturers and clinical applications. The tree structures show how one early approval, often a De Novo classification, can become the regulatory foundation for dozens of later devices. Predicate chains reach 8 generations deep, while 354 devices sit in isolation with no predicate link to any other device in the database.

We aim to regularly update the dashboard with additional specialties, features, and up-to-date devices as the FDA AI/ML database is periodically updated.

Hover over terms below for definitions:

How to cite this dashboard

Guni A, Finnemore Capon Diaz S, Hussain A, Gandhewar R, Davidson A, Tandon D, Dilaver N, Darzi A, Ashrafian H. FDA AI/ML Device Predicate Networks. Institute of Global Health Innovation, Imperial College London. https://predicates.app (data to December 2025; accessed [date]).

Provenance and limitations

• Only the primary predicate of each device is modelled. Submissions citing several predicates or reference devices are represented by the first cited.
• The network is self-contained within the AI/ML database. Where a device cites a predicate outside it, that predicate is named and dated but its own lineage is not traced.
• The mapped specialty datasets involve clinical judgement and are drawn from our scoping reviews; the FDA lead specialty is the FDA’s own classification.
• The vertical axis is ordered by date but not linear in time: each year is given room in proportion to how many devices it holds, so the sparse early years are compressed.
• The FDA has acknowledged that its AI/ML list may not be exhaustive, and the list itself changes: devices are added, including retrospectively. This dashboard reflects the list as extracted, so a device authorised before the cutoff but added to the FDA list afterwards will not appear until the next update.
• Predicate data were extracted manually from individual clearance summaries. Corrections are welcome: h.ashrafian@imperial.ac.uk.

Device data derive from the FDA’s public records. Reuse with attribution.