Three Decades of FDA Authorizations of AI/ML Enabled Medical Devices: Persistent Specialty Concentration and the Care Delivery Gap (1995 to 2025)

This cross-sectional analysis of 1,430 FDA authorizations from 1995 to 2025 reveals that while AI/ML-enabled medical device approvals have surged exponentially, they remain heavily concentrated in image-rich diagnostic specialties like radiology, leaving significant gaps in representation for other major clinical fields such as pathology, obstetrics, and behavioral health.

The Gist

Imagine the US Food and Drug Administration (FDA) as a massive, busy library. For the last 30 years, this library has been keeping a public guestbook of every new "smart" medical tool (devices powered by Artificial Intelligence or AI) that gets permission to be used in American hospitals.

This paper is like a librarian who finally sat down to read the entire guestbook from 1995 all the way through the end of 2025. They counted every entry, looked at who wrote them, and tried to figure out what kind of tools are being added to the shelves.

Here is what they found, explained simply:

1. The Library is Exploding in Size

For a long time, the library was quiet. Between 1995 and 2014, they only added about two new AI tools per year. But recently, the pace has turned into a firehose.

• The Analogy: Think of it like a slow drip of water turning into a rushing river.
• The Numbers: In the most recent three years (2023–2025), the FDA approved about 264 new AI tools every single year. Just in 2025 alone, they added 331 new tools. The total count reached 1,430 authorized devices.

2. The "Radiology" Room is Packed; The Rest are Empty

This is the biggest surprise. The library has many different rooms for different types of doctors (specialties), but almost everyone is crowding into just one room: Radiology (doctors who read X-rays, CT scans, and MRIs).

• The Analogy: Imagine a school with 30 different classrooms. If you walked in, you'd find 76% of all the students sitting in the "Radiology" classroom. The "Cardiovascular" (heart) and "Neurology" (brain) classrooms have a few students, but the "Pathology" (lab tests), "Microbiology" (germs), and "Obstetrics" (pregnancy) classrooms are nearly empty.
• The Reality:
  • Radiology: 76.5% of all approved tools (1,094 devices).
  • Heart & Brain: Together, they make up about 14% more.
  • The Missing Rooms: There were almost no tools approved for Psychiatry (mental health), even though mental health is a huge part of healthcare. There were also very few for Pathology, Microbiology, or Obstetrics.
• The Takeaway: The authors say this isn't a temporary glitch; it's been this way for a decade. The "Radiology" room keeps getting bigger, but the other rooms aren't catching up, even though doctors in those other fields treat millions of patients.

3. The "Super-Authors" vs. The "One-Hit Wonders"

The paper also looked at the companies making these tools.

• The Analogy: Imagine a songwriting contest. Most participants (67.8%) wrote just one song and entered it. But a tiny group of 13 "super-star" companies wrote dozens of songs each.
• The Reality:
  • The Majority: 502 different companies only have one approved AI tool.
  • The Minority: Just 13 companies (less than 2% of the total) are responsible for 15% of all the tools. One single company has 51 different tools approved.
• The Takeaway: The market is split between a huge number of small, one-time players and a small group of big companies that keep making more and more tools.

4. Why is Radiology So Dominant?

The authors suggest a few reasons why the "Radiology" room is so full:

• The Data: Radiology has a very organized way of storing pictures (called DICOM). It's like having a library where every book is already sorted, labeled, and on a shelf. Other fields (like mental health or lab work) don't have data that is as easy to organize and feed to AI.
• The Rules: The rules for approving software that looks at pictures are clearer than the rules for other types of medical AI.

What This Paper Does Not Say

It is important to stick to what the authors actually wrote:

• They do not say these tools are bad or good for patients.
• They do not say the FDA is doing a bad job.
• They do not predict exactly what will happen next.
• They do not claim that AI cannot work in mental health or other fields; they just point out that, so far, the FDA hasn't approved many tools for those areas.

The Bottom Line

This paper is a "state of the union" report for medical AI. It says: "We have approved 1,430 smart medical tools. They are growing fast, but they are almost all for reading X-rays, made by a mix of many small companies and a few big ones. If we want AI to help in other areas like mental health or pregnancy, we need to figure out how to build the data and rules to get those tools approved."

Technical Summary

Technical Summary: Three Decades of FDA Authorizations of AI/ML-Enabled Medical Devices (1995–2025)

Problem Statement
While the US Food and Drug Administration (FDA) maintains a public registry of artificial intelligence and machine learning (AI/ML)-enabled medical devices, prior analyses of this regulatory landscape have been limited by temporal scope and depth of characterization. Previous studies, such as those identifying 64 devices through 2020 or 691 by October 2023, have consistently noted a dominance of radiology applications. However, these earlier works did not characterize the full cumulative dataset through the end of 2025, nor did they systematically quantify the underrepresentation of other high-volume clinical specialties relative to their clinical burden. Furthermore, the concentration of authorizations among a small subset of manufacturers had not been systematically described, leaving gaps in understanding the market structure and innovation incentives within medical AI.

Methodology
This study performed a cross-sectional analysis of the FDA's publicly maintained list of AI/ML-enabled medical devices, accessed in a deidentified tabular format. The dataset comprised 1,430 authorizations with final-decision dates ranging from September 29, 1995, to December 30, 2025. These devices were authorized via the 510(k), De Novo, or Premarket Approval (PMA) pathways.

The analysis utilized Python 3.11 and the pandas library to compute:

• Temporal Growth: Annual and cumulative authorization counts stratified by year.
• Specialty Distribution: The proportion of authorizations attributable to each FDA lead review panel (broadly corresponding to medical specialties).
• Temporal Stability: The proportion of authorizations for the Radiology panel and the top three combined panels annually from 2015 to 2025.
• Manufacturer Concentration: The distribution of authorizations across unique manufacturers, distinguishing between single-authorization firms and repeat-authorizing entities.

The study was descriptive in nature; no formal hypothesis tests were planned. Ethical approval was not required as the analysis used a publicly available regulatory dataset containing no patient-level information.

Key Results

• Exponential Growth: The annual authorization rate increased from a mean of 1.8 per year (1995–2014) to 264.0 per year (2023–2025). The cumulative count reached 1,430 by December 2025, with a single-year peak of 331 authorizations in 2025.
• Persistent Specialty Concentration: The Radiology panel accounted for 1,094 authorizations (76.5%). The top three panels—Radiology, Cardiovascular, and Neurology—collectively represented 90.6% of all authorizations.
• Specialty Disparities: Several large clinical specialties showed minimal representation. Pathology had 9 authorizations (0.6%), Microbiology had 6 (0.4%), and Obstetrics and Gynecology had 4 (0.3%). Notably, zero authorizations were recorded under a psychiatry or behavioral health review panel.
• Temporal Stability: The share of annual authorizations reviewed by the Radiology panel remained stable between 61.5% and 86.8% from 2018 to 2025, with no sustained downward trend despite the massive increase in total volume.
• Regulatory Pathways: The 510(k) pathway was the predominant mechanism, clearing 1,376 devices (96.2%). De Novo and PMA pathways accounted for 2.6% and 1.2%, respectively.
• Manufacturer Concentration: Among 740 unique manufacturers, 502 (67.8%) had only a single authorized device. Conversely, 13 manufacturers (1.8%) accounted for 217 devices (15.2%), with the top manufacturer holding 51 authorizations.

Significance and Claims
The paper claims that the cumulative regulatory record demonstrates rapid growth that is structurally concentrated in image-rich diagnostic specialties, specifically radiology. The stability of this concentration over a decade, despite a roughly 50-fold increase in annual authorization rates, suggests that the drivers are structural—likely related to the availability of large, labeled, standardized imaging datasets (via DICOM infrastructure), well-defined regulatory frameworks for image analysis, and established commercial workflows—rather than transitional.

The authors highlight the "care-delivery gap," noting that the underrepresentation of specialties with substantial clinical activity (e.g., psychiatry, pathology, OB/GYN) contrasts sharply with the volume of clinical work performed in those fields. The findings are intended to inform policy discussions regarding where regulatory, infrastructure, and dataset investments are most needed to broaden the clinical scope of medical AI, rather than to advocate for specific policy interventions. The study also notes that public-facing aggregators, such as PhysicianAITools.com, can support broader landscape analyses by linking FDA records with discovery-based metadata.