Clinical adoption of medical AI devices through US insurance claims (Wu et al., 2024)

Kevin Wu, Eric Wu, Brandon Theodorou, Weixin Liang, Christina Mack, Lucas Glass, Jimeng Sun and James Zou, Characterizing the Clinical Adoption of Medical AI Devices through U.S. Insurance Claims, NEJM AI 1(1), AIoa2300030 (January 2024), doi:10.1056/AIoa2300030; preprint medRxiv 10.1101/2023.08.26.23294629 (2023-08-28). The research pass on 2026-09-27 brought it in for growth edge 6c.

Read at abstract level only. NEJM AI returned 403 and medRxiv served a bot check, so the figures below come from the journal and preprint abstracts (the preprint's through Europe PMC, PPR704293). The per-device table was not read.

What it counts

The FDA had authorised more than 500 AI devices, and nobody knew how often any of them ran. The authors used billing to find out. They tracked the Current Procedural Terminology (CPT) codes created specifically for medical AI through a nationwide claims database:

• Journal version: 11 billion CPT claims, 2018-01-01 to 2023-06-01.
• Preprint: 16 billion CPT claims, 2015-01-01 to 2023-06-01 (a different extract of the data).

They picked CPT over Medicare's new-technology add-on payment because the add-on exists only inside Medicare and would show part of national use.

What it found

• Adoption is early and concentrated. A handful of leading devices account for most use.
• Only two device classes pass 10,000 claims: AI for assessing coronary artery disease and AI for diagnosing diabetic retinopathy.
• Use follows money and cities. Zip codes with higher income, metropolitan zip codes, and zip codes with academic medical centres were "much more likely" to show AI use, in the journal abstract. The preprint abstract said "moderately over-represented". The two versions used different data, and the wording changed with it.

Why it matters here

This is the first count of uses in the corpus, not a count of things authorised. Set it beside ai-ml-device-adverse-events: Babic et al. (2025) found 943 adverse-event reports across 950 authorised AI/ML devices and said a rate was impossible without a usage figure. Wu et al. supply a usage figure for two device classes. For the other authorised devices the abstract gives nothing per device, only that use is concentrated in a handful of them.

The source of the count still matters. The count is a by-product of billing, compiled by researchers from a commercial claims database. The FDA did not collect it, publish it, or require it. That matches what ai-act-post-market-monitoring found for the EU: the regulation asks for performance data and never for volume. The health route produced a denominator only because payers need one to pay.

Limits

• Only billed use is visible. A device used without a dedicated CPT code, or bundled into another payment, leaves no trace here.
• Two classes cross the threshold, so the corpus can now divide harm by use for at most two kinds of device, and has not done it.
• No outside replication. A 2026 follow-up on commercially insured adults (Zhang et al., JMIR Formative Research, 2026-07-23, doi:10.2196/92726) cites it and reports rapid growth with variable use. We read only its one-sentence abstract.

Related

ai-ml-device-adverse-events · ai-act-post-market-monitoring · risk-based-regulation · ai-incident-database · synthesis