Musings on healthcare IT, FHIR, EHR systems, and digital health innovation

🔥 FHIR in the hole!
Back to posts

Reading 50,000 Pages of Public Comments on the Physician Fee Schedule

Original on LinkedIn

The proposed CY2027 Physician Fee Schedule drew 43,082 public comments (about 50,000 pages). That’s ~5x bigger than any docket I’ve analyzed before, and my old pipeline would have cost around $2k to run on it. So I spent some time improving the pipeline. It now groups form letters and campaign letters, works out who each comment speaks for, reads charts and tables in attachments, and publishes the whole analysis as downloadable databases. The full PFS run cost about $100.

TL;DR: Browse the results in the CY2027 PFS comment dashboard. To explore the comments with your own AI assistant, give it the skill file.

The results are in the CY2027 PFS comment dashboard. Thanks to Travis Broome and Sean Cavanaugh at Aledade for suggesting this docket.

Findings

Modifier 25 drew more comments than any other issue. CMS proposed paying 50% less for an office visit billed on the same day as a procedure, and 21,794 comments took that up — twice as many as addressed the overall cut to the conversion factor (10,368).

Most of that volume came from organized campaigns (including three campaigns about ENT care alone accounting for nearly 5,000 letters).

If you rank issues by “thoughtful submissions” (here: distinct letters longer than 1,500 words), the top three issues become: drug pricing and safety-net programs, MIPS and health IT, and remote patient monitoring.

Issue areas by number of comments

The left bars count comments, including form letters. The right bars count distinct letters longer than 1,500 words. Because a comment can raise several issues, the counts add up to more than the totals.

Loading the docket

My earlier dockets came from regulations.gov’s bulk download. For this one I added support for Mirrulations, a public mirror of regulations.gov run by Ben Coleman’s group at Moravian University. Mirrulations keeps an hourly copy of the site (about 27 million comments, attachments and text files) in a public S3 bucket. Because it’s built with many donated API keys, it isn’t held to the regulations.gov API’s limit of 1,000 items an hour. I pulled all 43,000 PFS comments and 4 GB of attachments from it in under an hour. It’s a great public service, and I’m glad the pipeline can use it now.

Form letters and campaigns

Many comments are copies of the same letter. The pipeline now spots form letters even when senders add a paragraph of their own. It ignores “See attached” stubs, and it matches scanned copies of campaign letters once they’ve been transcribed. After that, the docket comes down to about 20,000 distinct texts.

Comments by kind

Form letters are identical or nearly identical copies. Campaign letters make a campaign’s argument in the sender’s own words.

Last year I predicted that AI-personalized campaigns would make simple copy-matching obsolete. So the pipeline now also groups letters by meaning. It embeds each letter, clusters the letters that make the same argument, and asks a model whether each cluster is a real campaign. It found 284 campaigns covering 25,612 comments. About 2,500 of those are campaign letters written in the sender’s own words, which copy-matching would miss.

Next, the pipeline breaks each distinct text into its individual points, sorts the points into themes, and writes a report for each theme.

Pipeline steps

There are more points than texts because a long letter can make dozens of separate points.

Cost

The full run cost about $100 in Gemini calls. Each step uses the cheapest model that does the job well: Flash-Lite for rote work like triage, and Gemini 3.8 Flash for synthesis. Long letters are read once against the whole theme list instead of once per theme, though pulling points out of every letter was still the biggest single cost.

Cost by step

The old pipeline’s cost is an estimate based on per-comment costs measured on a sample.

On PFS the pipeline makes about 30,000 model calls with about 270 million input tokens in total. Run live, the fastest possible time depends on Gemini’s rate limits. My account is at Tier 3, which requires $1,000 of past spending and 30 days of history. At that tier Gemini 3.8 Flash allows 20 million input tokens a minute, so the whole run could finish in about 35 minutes for about $210. Getting there means keeping about 100 calls in flight, staying just under the token limit during the longest step, and retrying any calls that get rejected. At the pipeline’s default of 20 calls at a time, the run takes nearly three hours.

Batch mode avoids all of that. Thousands of requests go up as one file, and Google schedules them. Batch calls cost half as much as live ones, which is how the full run came to about $100. The batch jobs took about three hours, and that was fine for a run I left to finish on its own. Each batch job is also recorded when it’s submitted, so when my run was killed partway through, restarting it picked up the same jobs instead of paying for them again.

Before the full run I ran small experiments to choose models and settings, which cost about $50 in total. Blind comparisons showed that letting the model reason (“thinking”) helped with theme discovery and report writing but nothing else, so it’s turned off everywhere else. A 1-in-20 sample of the docket ran end to end for about $15 to make sure every step worked. I also checked whether a bigger model would write better reports, and in a blind comparison Gemini 3.8 Flash beat Gemini 3.1 Pro 8 to 3, even with Pro as the judge.

Experiments and what they decided

A few small trials weren’t metered, so the true total is a bit higher.

New features

Every docket now has two downloadable SQLite databases for your own analysis or an AI assistant’s. For PFS they’re 46 MB and 69 MB zipped. The databases document themselves: every column has a comment, and a _readme table explains where the data came from, lists caveats, and includes about ten worked example queries. Like my earlier switch from an MCP server to a skill file, this lets an AI assistant work with the data directly.

Scoped analyses let you point the analysis at one question, such as “interoperability and health IT” or “the issues raised in Johns Hopkins Medicine’s letter.” Each scope gets its own sub-site with its own themes and reports.

Entity tagging now reads the whole docket and found 819 entities (up from 128), including dQM, QRDA and TEFCA. Phrase search now downloads 0.4 MB instead of 28 MB, and the overview page has been redesigned.

Building it with Claude Code

This project used two AI systems. I steered Claude Code, which wrote and tested the pipeline, and the pipeline uses Gemini to read the comments.

How the pieces fit

Claude Code wrote the pipeline and ran the Gemini experiments. The pipeline calls Gemini to read the comments.

I built all of this in one Claude Code conversation. My messages fell into six sittings, but I was doing other things for most of that time. I estimate that writing messages and reading replies took about an hour and three quarters of my attention. Claude or the pipeline was working for about 11 hours in all, including two and a half hours while I was away.

Hours

My attention by activity

I skim quickly and write by dictation or fast typing. These ranges assume I read 575 to 1,000 words a minute and wrote 120 to 200. I didn’t track the time I spent looking at the sites, so that row is a guess.

I sent 118 messages with about 3,000 words of my own. For every word I wrote, Claude wrote about 15 and read about 250 words of tool output — mostly files, logs, query results and test runs.

Words, as areas

Each square’s area is proportional to its word count. Tool output is estimated at six characters per word, and Claude’s count leaves out its hidden reasoning.

The main conversation planned, reviewed, tested and committed the work. Eighteen subagents each took on one self-contained job, like building a Batch API backend, scaling the dashboard to 43,000 comments, profiling search with DevTools flame charts, or auditing submitter metadata. Between them, the subagents made most of the tool calls.

Timeline

Shaded bands show the sittings when I was in the conversation.

Claude wrote all 49 commits, adding 11,496 lines of code and deleting 4,326. They went in as one pull request. The deletions include an older clustering method, a superseded summary command and settings that testing showed were no longer needed.

Lines added and deleted

These are net changes across 97 files. Counting code that was written and later rewritten, Claude added 12,840 lines and deleted 5,249.

My role was to set the direction and review the results. I gave instructions like “Flash-Lite only for rote work” and “simple is good, but it can’t be so simple it’s shitty.” I also caught several problems: “See attached” stubs grouped as a campaign, top-level themes showing zero comments, an entity list missing dQM and QRDA, and an overview headline that read like AI slop. In January I wrote that subagents were getting easier to use. This project relied on them throughout, and I coordinated them just by talking with the main Claude Code conversation instead of writing orchestration code.

Conversation on LinkedIn

22 comments · 87 reactionsReply on LinkedIn
  • CMIO at Golden Valley Memorial Healthcare

    Remote patient monitoring is a great lever. The problem is patients don’t see it that way and don’t or won’t want a copay.

  • Founder of MedArise | Building AI-native healthcare operations platform

    The 50 percent same day visit and procedure proposal is the sleeper issue for independent practices. For a small group, that modifier scenario is not an edge case; it is weekly volume in dermatology, orthopedics, primary care procedures. Layer that cut on top of prior auth delays and quality reporting overhead and the margin math breaks quickly, even before staffing costs. I appreciate the SQLite approach; it lowers the barrier for practices without a policy team to actually see who is commenting and why. What did the longer letters say about RPM documentation burden versus payment?

  • Lucia SavageOct 5, 2026

    Smart ideas to fix healthcare. Board Member. Advisor. Expert advisor on public policy, AI, privacy, cybersecurity, and the future of healthcare. I answer questions for CNN and Senators. I write my own posts, not AI.

    Did you find mine, suggesting that CPT be regulated like a utility, because that’s how you manage necessary monopolies?

    4 reactions

  • Benjamin KrasinOct 5, 2026

    Founder, Krasyn. Building AI for clinicians and learners, and testing it in public.

    The step I'd most want audited is the chart and table transcription. Quoted text gets checked by whoever quotes it, but a number lifted from a scanned bar chart rarely gets checked by anyone. When we had a model read photographed handwritten lists, we settled on making it flag every value it was unsure of instead of returning a clean guess. A per-cell confidence column in those SQLite tables would tell readers which numbers to recheck.

  • Karthik ChandramouliOct 5, 2026 · edited

    Helping GI doctors achieve Peak Practice Efficiency 🔶 0️⃣➡️1️⃣ GTM Leader & Product Builder at Bookend AI 🔶 CEO 🔶 Board Member 🔶 Angel Investor

    This is so damn cool Dr. Mandel, I was futzing around with a couple of open source libraries to extract the file attachment letters and parse them.

    Just amazing, and I found my long letter on behalf of Bookend AI commenting on Gastroenterology quality measures!

    https://joshuamandel.com/regulations.gov-comment-browser/CMS-2026-2377/#/comments

    It would be helpful if commenters could see which other organizations or individuals agreed / disagreed with my comments.

    Who is aligned / not aligned with my perspective?

    Who has an interesting, but different perspective on the same issue / topic?

    For topics that might be less popular, what were the comments grouped by medical specialty or medical society?

  • Joseph S.Oct 5, 2026

    Lightening the load for seniors and their families | Board Certified Patient Advocate | Medicare-covered care navigation | Founder, SageAlly

    Josh Mandel, MD. Wow, this is a remarkable and amazing resource!

  • Physician Executive & Healthcare AI Builder | Clinical-Financial Operations, Revenue Cycle & Health Policy | AI Evaluation, Governance & Implementation | Rural Health | Founder, ClinEfficiency Pro

    very well written

  • Kajetan MazurOct 6, 2026

    AI Developer · Agent Builder · Developing end-to-end solutions on Azure

    Two hours of steering Claude Code to rebuild the pipeline, then ~50k pages read for about $100, is a striking ratio. Shipping each docket as a self-documenting SQLite database so people can query it directly is the part I'd copy.

  • Matthew HoltOct 6, 2026 · edited

    Health care curmudgeon

    How many of the comments said that fee-for-service is a completely stupid way to pay for healthcare and therefore should be abolished and changed to some kind of subscription model? Oh I forgot to submit mine.

    • Josh Mandel, MDAuthorOct 6, 2026

      Matthew Holt Yours would have brought the tally above 1/10k ;) Only four comments say fee for service should be scrapped in favour of a subscription-style payment: a flat, recurring sum per patient.

      - A health-plan analyst says per-visit pay forces doctors to carry huge patient lists to stay profitable. He argues "Primary Care needs to no longer be fee based" and points to capitated and direct-primary-care models instead. (https://joshuamandel.com/regulations.gov-comment-browser/CMS-2026-2377/#/comments/CMS-2026-2377-14244)

      - Luis Dominguez argues Medicare's payment units reward quantity, not quality. He wants Medicare to start with hybrid payments and "adjust until 100% of payments are prospective" and risk-adjusted. (CMS-2026-2377-11146)

      - An anonymous paediatrician wants primary care "paid more like a gym membership": about $200–250 per child per month for access to a paediatrician, with specialists funded by a separate payment of the same kind. (CMS-2026-2377-26009)

      - Andrew Schutzbank, an internist and early Iora Health leader, says fee-for-service pays for visits when patients need continuity. The plan he proposes ends: "Gradually fade out fee-for-service revenue." (CMS-2026-2377-30763)

    • CMIO at Golden Valley Memorial Healthcare

      Matthew Holt my problem with those models is they appear to be a money grab with no real accountability for quality. I understand doctors say they provide “higher quality” but I have also seen many doctors believe they are providing “higher quality” while never measuring or benchmarking anything. It is easy to get a bullseye if you never define the target.

  • Faiza AatifOct 6, 2026

    Director of Healthcare Technology at PureLogics | Turning Healthcare Challenges into Scalable Digital Solutions

    The proposed 50% slash to Modifier-25 reimbursement exposes the core tension in health policy: CMS views same-day E/M visits and procedures as overlapping administrative overhead, while clinicians view them as essential friction reduction for high-risk patients. When payment policy incentivizes unbundling care into separate visits to avoid haircuts, administrative efficiency directly undermines clinical access and exacerbates health inequities.

  • Chris CashwellOct 6, 2026

    Chief Revenue Officer | Chief Commercial Officer | CEO | Enterprise SaaS & B2B | AI Commercialization & Operational AI | Healthcare Technology

    Love the self-documenting SQLite drops, really useful.

  • Anu JacobOct 6, 2026

    Healthcare Commercial Executive | $250M+ P&L | Enterprise Growth Across Academic & Pediatric Health Systems | Medtech, Diagnostics & Digital Health | Kellogg MBA

    There’s an interesting implication here for the organizations submitting these comments too.

    When tens of thousands of submissions can be analyzed this quickly, volume becomes much easier to separate from substance. That could change the value of the comment itself. A hundred variations of the same argument may still demonstrate constituency, but a well-supported submission with distinctive evidence or operational insight becomes much easier to find in a dataset this large.

    The technology isn’t just making the public record easier to analyze. It may eventually change how organizations think about contributing to it in the first place.

    1 reaction

  • 1 followers

    Interesting that the modifier-25 cut drew the most comment volume — it seems to be the rare proposal that unites independent physicians and specialty societies. One question the data could answer: which specialties show the heaviest exposure? Dermatology is the canonical example of same-day E/M + procedure billing, so it would be telling to see whether derm comment volume and tone cluster differently from primary care on this one.

  • Bill LakenanOct 6, 2026

    Director of Innovation @ Bellese

    Josh Mandel, MD Thank you for this public service. Your efforts find clear signal in the noise. Truly appreciated.

    1 reaction