If your medical AI procurement framework treats FDA authorisation as a reliable signal of clinical evidence quality (the assumption being that FDA-cleared devices have crossed a clinical-evidence threshold meaningful for clinical adoption), the 2025 FDA authorisation data and the December 2024 peer-reviewed analysis show the assumption needs adjustment. The FDA authorised 258 AI/ML-enabled medical devices in 2025, the highest annual total ever. Cumulative authorisations crossed 1,357 devices from 693 companies. But only 2.4% of devices with clinical studies were supported by randomised controlled trial data, according to a peer-reviewed analysis of all 1,016 authorisations through December 2024. The vast majority entered via 510(k) pathways that rely on existing safety and efficacy evidence rather than new clinical trials. The regulatory authorisation count is a market-formation metric, not a clinical-evidence metric.

The 510(k) pathway is structurally the bottleneck. Under FDA 510(k), manufacturers demonstrate that a new device is "substantially equivalent" to one already on the market. The pathway does not require new clinical trials: it relies on the evidence base of the predicate device. For software, this means a new AI/ML algorithm with a similar intended use can be authorised based on the clinical evidence supporting an earlier algorithm, even when the underlying model architecture, training data, and operational characteristics differ substantially.

The Singh et al. 2025 peer-reviewed analysis examined all 1,016 FDA authorisations through December 2024 and found that:

  • The vast majority entered via 510(k) pathways
  • Only 2.4% of devices with clinical studies were supported by randomised controlled trial data
  • Most cleared devices rely on retrospective performance evaluation rather than prospective clinical validation
  • Predetermined Change Control Plans, a mechanism permitting iterative updates after initial authorisation, were used in approximately 10% of 2025 clearances

The prescriptive implication is direct: hospital systems and clinical procurement teams cannot rely on FDA authorisation alone as a signal of clinical evidence quality. The authorisation pathway is appropriate for the regulatory mission (preventing harmful devices from reaching market) but does not validate clinical effectiveness in the deploying setting.

Three structural drivers shape the current authorisation environment.

The first driver: regulatory pathway design predates the rapid expansion of AI/ML medical devices. The 510(k) framework was developed when device innovation moved at the pace of mechanical and electrical engineering iteration. AI/ML algorithms iterate at substantially faster pace. The pathway has been adapted (Predetermined Change Control Plans, January 2025 draft guidance on AI-enabled device software) but the underlying substantial-equivalence framework remains. The pathway throughput has scaled (258 devices in 2025) but the evidence base per device has not strengthened proportionally.

The second driver: the market is concentrated. By December 2025, 1,357 cumulative authorisations were held by 693 companies, meaning many companies hold one or two devices each. GE Healthcare leads with 93 devices, Siemens Healthineers 82, Shanghai United Imaging 38, Philips 36, Canon 35, Aidoc 30. The concentration at the top reflects deep regulatory expertise; the dispersion across many small holders reflects broad ecosystem participation but limited evidence-generation capacity at each individual firm.

The third driver: prospective clinical trials are increasing but from a low base. The number of papers reporting prospective trials of clinical imaging ML/AI models grew from 417 in 2024 to 536 in 2025, a 28.5% increase. The field is moving toward more prospective validation, but the volume of trials is still substantially smaller than the volume of authorisations. The trial backlog is unlikely to close at current pace.

The prescriptive moves for organisations procuring medical AI in 2026:

The first prescriptive move: separate regulatory authorisation review from clinical evidence review. The procurement process should require both. Regulatory authorisation confirms the device is legally available; clinical evidence review (peer-reviewed studies, real-world deployment data, randomised trial results where available) confirms the device is likely to produce intended clinical outcomes in the deploying setting. The two reviews are different and should not be conflated.

The second prescriptive move: prioritise devices with prospective clinical trial evidence specifically. The 28.5% growth in prospective trials means there are now hundreds of devices with trial-grade evidence. Procurement that filters for prospective evidence as a baseline criterion will select among a smaller but better-evidenced subset of authorised devices.

The third prescriptive move: build internal post-deployment monitoring that operates independently of pre-market regulatory pathways. The FDA 510(k) pathway accepts evidence from earlier devices; the deploying institution should generate its own evidence from current deployment. Outcomes measurement that compares pre- and post-deployment clinical metrics provides the institution-specific evidence base that regulatory authorisation does not provide. The peer-reviewed publications from Sharp HealthCare, UChicago Medicine, and Northwestern Medicine (cited in article #121) demonstrate the methodology for this.

The fourth prescriptive move: engage with the FDA AI-enabled device guidance (January 2025) and Predetermined Change Control Plans (used in ~10% of 2025 clearances) as the regulatory direction. The mechanisms allow iterative algorithm updates after initial authorisation, which can support real-world learning. But they also mean the device that goes into deployment may differ materially from the device that was originally authorised. Procurement and clinical oversight should engage with the update mechanism as part of ongoing governance, not a one-time pre-market check.

The fifth prescriptive move: contribute to the prospective trial evidence base where deployment provides the opportunity. Large health systems deploying AI medical devices at enterprise scale have a structural opportunity to contribute prospective trial-grade evidence through deployment design. The University of Chicago, Northwestern, and similar large systems have done this for ambient AI scribes. The same approach can be applied to FDA-cleared diagnostic and decision-support devices. The contribution strengthens the field's evidence base and improves the deploying institution's evidence on the specific deployment.

The trajectory: FDA authorisation volume will likely continue to grow through 2026-2028 as the AI/ML device ecosystem expands. The proportion of authorisations with RCT-grade evidence will likely improve from 2.4% but slowly. The gap between regulatory authorisation and clinical evidence will remain a procurement consideration for the foreseeable future. Strategic plans that engage with this reality through 2028 will be operationally aligned; plans that rely on FDA authorisation alone as evidence quality signal will be exposed to deployment-evidence gaps that affect clinical outcomes.