Explore the FDA predicate relationships of AI medical devices
This dashboard was developed by researchers at the Institute of Global Health Innovation, Imperial College London, as part of an ongoing programme examining how AI medical devices are regulated by the FDA.
It brings together 1,614 devices from the dated FDA AI/ML list, with decisions through 29 June 2026, and the primary predicate relationships identified from public FDA documents. Alongside the full list, the Surgery, Cardiology and Neurology datasets reflect clinical scoping reviews conducted by our group. Their coverage differs, as explained below. Devices are counted by FDA submission number; separate entries can represent versions or supplements of the same commercial product.
The devices in this dashboard entered through three regulatory pathways:
FDA guidance: 510(k) · De Novo · PMA.
The FDA publishes a list of AI medical devices, but their predicate relationships are not directly visible in that list. This dashboard makes those connections explorable. Predicate comparisons can support proportionate evaluation of incremental innovation, allowing new devices to build on established uses. For clinicians, the useful question is how a device and its supporting evidence relate to the patients and clinical task for which it will be used.
Find a device, compare its description with its direct predicate, and open both FDA sources. Consider the intended patients, clinical task, inputs and performance evidence. Network shows the wider connections; Lineage follows a family or the selected device’s ancestry. A chain’s length alone does not establish safety, benefit or the adequacy of evidence.
FDA lead specialty is the classification assigned in the FDA AI/ML list. Radiology accounts for 76% of devices under this classification.
Mapped specialty reflects our research group’s clinical reviews. The original reviews cover 661 devices, of which 659 occur in the current list. Some devices are relevant to more than one specialty. For example, a radiological tool for detecting intracranial haemorrhage may also be relevant to neurosurgical and stroke care.
Both classifications appear in the filters and device details. The specialty datasets are historical subsets with different coverage, and the retained same/other specialty labels do not establish similarity of clinical application. Not mapped identifies devices without a clinical mapping here.
Across all listed devices, 555 (34.4%) cite a predicate outside the FDA AI/ML list. This describes list membership; it does not establish whether the predicate uses AI. Predicate identity could not be established for 10 devices (0.6%). A further 61 devices (3.8%) are listed under De Novo or PMA, including PMA supplements, for which a 510(k) predicate is not required.
The graph records one selected primary predicate per submission. Shared ancestry does not establish equivalence between every pair of devices, and relationships beyond the dataset boundary are not fully reconstructed. The Devices and Trends views provide the underlying data and filtered results for download.
Read the dataset coverage, extraction methods and suggested citation, or download the data with its field dictionary. We intend to update this resource regularly as new FDA records become available. For corrections, contact Ahmad Guni.
Hover over a term, select it with the keyboard, or tap it for a definition.