🔷 Epic’s next predictive tool will only open to organizations that feed Cosmos, and that gate is both a carrot and a stick.
🔷 STAT says Epic’s gatekeeping is choking innovation. Some of it rings true, but the border between choking innovation and disciplined development is not black and white.
🔷 Sharp and UCSD bought hospitals they cannot see into, and a working Carequality pipe will not fix that.
🔷 HTI-5 is a retreat to protect the certification program. Information blocking is still a paper tiger.
🔷 An AI triage model read the chart. The problem is that the model is based on what clinicians wrote.
✅️ Off-Label: autism therapy chains show what happens when billed hours become the target.
hitdoc.substack.com/p/epics-…
🔷 Epic's Connectors let outside apps plug into Emmie, Art, Penny, and Agent Factory. Somebody will build another layer that ties every health systems’ bots together. My bet is that it ends up being Epic.
🔷 MyChart links to ChatGPT rose more than 7,000% in a month. Health systems pay for the API calls. Context is king, and a well-built app inside the EHR beats a generic chatbot. There will be a hidden cost: lost ambulatory revenue, really, isn’t that the point?
🔷 EHR consolidation is a chance to clean up decades of clinical data. Most organizations move the hoarder junk into a fancier new house.
🔷 Amazon, Google, Meta, and Oracle signed the electronic prior auth pledge. Payers have already shown how to comply on paper and stay opaque. Let’s force them to remove opacity.
🔷 Stanford built an exam for the back office instead of a multiple-choice licensing test. The best agent finished about a third of the tasks. The simulation is still simpler than the real one.
✅️ Off-Label: Claude got every number right on a claims database. Some of the explanations were wrong, and that failure is harder to catch.
Epic Opened Its AI Agents to Outside Apps. What Will Connect those Apps? hitdoc.substack.com/p/epic-o…
🔷 Epic's new-customer count fell from 28 to 11. That's what a saturated market looks like, not a slowdown. Future growth needs to come from Community Connect, mergers, and OCHIN. Whether a small hospital ends up a partner or an afterthought depends on governance.
🔷 VA staff can now book veterans directly into Epic health systems' schedules. Epic will interoperate, as long as someone else carries the translation cost and nobody touches Cadence.
🔷 Medical boards are writing the rules for clinical AI. A license regulates a person, and an AI agent doesn't have one. Utah's own licensing board is already fighting Utah's own AI sandbox.
🔷 CMS wants to grade Medicare Advantage provider directories. A compliance check looks at the listing, not whether anyone can actually get an appointment.
🔷 Trinity is outsourcing 557 IT jobs. Health systems call IT strategic until the WARN notice goes out.
✅️ Off-Label: Nearly half of Texas nonprofit hospitals bill patients before checking whether they qualify for free care. The hospital policy reveals how nonprofit they want to be.
hitdoc.substack.com/p/epic-s…
Last week, I discussed how we can surface key information that used to be lost in the the text of the chart. Now what can we do with the data? Provider orgs and patients suspect that payer plans are penny wise and pound foolish. They don't pay for things but end up paying for complications and hospitalizations. This is a piece on how we can quantify that.
hitdoc.substack.com/p/the-pl…
🔷 A STAT reporter spent weeks trying to opt out of Care Everywhere and got through only by calling medical records. The workflow exists. It’s used so rarely that nobody remembers how it works. And why opt-out by default is still the right call.
🔷 Epic’s Prescription Marketplace puts cash prices next to insurance prices before the script is sent. Surescripts tried something similar in January and pulled it within weeks. The difference this time is who owns the plumbing.
🔷 Heidi’s agents are working Australia’s dominant GP software through the same screens doctors use, and Abridge just bought the same capability. Can the system tell whether a physician or a bot clicked the button?
🔷 NEJM AI says LLMs recommend more expensive care than physicians. The study’s data sampling was dubious. The headline will still drive budget decisions.
🔷 Bessemer’s ROI scorecard says healthcare AI paid back twice as fast as expected. Two big caveats: most organizations have abandoned at least half of their homegrown AI tools, and clinical AI barely registers on the ROI chart.
✅️ Off-Label: A dataset is not a business model. Health systems signing AI vendor contracts should ask the question investors are learning to ask.
hitdoc.substack.com/p/epics-…
🔷 UPMC’s Epic go live is going in waves. Vandalia Health will take 27 EHRs live on one Epic instance in a single day. Big bang is the right call if you can pay for it.
🔷 Northwestern put OpenEvidence inside Epic for nurses and pharmacists. A content tool doesn’t need deep integration. What happens when evidence and local policy don’t say the same thing?
🔷 Google’s AMIE paper didn’t evaluate its “matches clinicians on diagnosis” claim. Once the AI is in the workflow, a clean human comparator is almost impossible to find.
🔷 CMS ran AI policy through a government Slack and called itself a “sales engine.” It could be pay-to-play, or it could be the railroad land grants. The test is whether the track gets laid for everyone.
🔷 Agent errors spread like an aggressive metastasis. So can the corrections.
✅️ Off-Label: A cardiologist stopped walking into the room with the answer. Every vendor building pre-visit AI summaries should read why.
hitdoc.substack.com/p/oracle…
We are depending on population health to solve our healthcare system problems. Population health depends on acting on patterns we all see when we take care of patients. Some of them happen over and over again but we haven't been able to turn those patterns into actions. Here are my thoughts on how we can do that now.
hitdoc.substack.com/p/the-co…
🔷 Epic’s Deterioration Index is FDA-cleared, and its sensitivity varies by race in the filing. The harder problem is that you need local validation, which is far easier to say than to do.
🔷 A court says a hospital cannot contract away its breach duty. Epic’s blank audit log makes that everyone’s problem, and everyone is pointing at someone else.
🔷 AI makes Epic’s business logic cheaper to copy. The moat that survives is the 1% long tail that vibe coding can’t reach. Will Epic build doors in Agent Factory?
🔷 AKASA says its AI codes inpatient cases with no human in 90 seconds. Matching human coders proves precision, and nobody has checked accuracy.
🔷 CMS’s AI prior authorization pilot lets vendors keep a share of the savings from denials. The fix is to test the algorithm on curated cases before deployment and while it runs.
✅️ Off-Label: Americans use AI and mostly don’t like it. Your adoption dashboard has the same blind spot.
hitdoc.substack.com/p/epics-…
🔷 Hull v. Epic tells us how valuable UserWeb access is. Veeva is now helping pay for the fight. The enemy of my enemy is my friend.
🔷 You’re on Epic CareEverywhere but you still send faxes. The train tracks are there but the non-Epic stations never got configured.
🔷 RFK Jr. says AI beats doctors. Six physician groups say it doesn’t. Both are arguing about requiring a human in the loop while that human quietly turns into a rubber stamp.
🔷 HCA nurses say a Palantir-built scheduler is understaffing their units. Scheduling is an infinite game. As is usual in healthcare IT, the missing piece is a feedback loop.
🔷 CommonSpirit says AI captured $200 million. The committee that reviewed the use cases is more interesting than the number.
✅️ Off-Label: AI graders agree with each other more than humans do. That is not the same as being right.
hitdoc.substack.com/p/hull-v…
🔷 Epic’s pause has a reason: some customer configurations of MyChart could let outsiders see records without leaving an audit trail. Every EHR carries that kind of configuration risk, and AI is about to multiply it.
🔷 St. Luke’s and Penn are keeping third-party ambient over Epic’s native AI. These decisions are analog, not boolean, and clean data sometimes means you don’t need the AI at all.
🔷 Microsoft and Mayo are building a frontier model. Stuart Miller has seen this movie three times, and the clock it will be judged on is far too short.
🔷 HITECH paid health systems to implement. Nobody pays them to maintain, and a promising post-go-live AI oversight effort will have to fight that gravity.
🔷 The AHA wants AI vendors held to HIPAA-level standards. The bigger leak already runs through clinicians and patients pasting charts into chatbots.
✅️ Off-Label: AI can only work with what you’ve written down. Most of what makes an Epic build work never was.
hitdoc.substack.com/p/epics-…
🔷 A Permanente physician built a multi-agent tool at Epic’s Agent Factory Build-a-Thon and noted that was the easy part. A separate benchmark effort shows how far agents still have to go.
🔷 Epic says Hello World prevented 3.4 million no-shows. The number grades itself, but it also passes the sniff test to anyone who works on the front lines.
🔷 Prior auth APIs go live October 1. Epic’s own policy lead says they aren’t magic, and payers have incentive to keep them friction prone.
🔷 NCQA retired the hybrid quality method. Colorectal screening scores up to 9 points lower when only structured data is used. It’s painful but maybe this is how the system moves.
🔷 Billing vendors say AI coding captures what hospitals are owed. Insurers call it inflation. Both are fighting over slices of a pie that needs to be rebaked.
✅️ Off-Label: Carriers told Congress employers already get their claims data. One late sentence in a transparency bill would let them withhold nearly all of it.
hitdoc.substack.com/p/epic-m…
🔷 Modern Healthcare says Epic paused hundreds of projects for cybersecurity. Epic says the roadmap is the same. Either way, the cybersecurity tax is permanent.
🔷 A federal court ruled patients can’t stop ambient AI from recording their visits. The court answered a consent question built for paper charts.
🔷 Hospital AI files prior auth requests. Insurer AI denies them. Bots are fighting bots when they could be using the same data to actually fix healthcare.
🔷 Oracle’s VA EHR is live at 17 of 170 sites, at $27 billion and climbing. An OpenAI agent got into Australian Medicare records it wasn’t authorized to touch. Different failures, same gap: governance built for slower systems.
✅️ Off-Label: The fight over whether 100,000 or 250,000 Americans die from medical error each year is nearly a decade old. Arguing about the number misses the forest for the trees.
hitdoc.substack.com/p/epic-i…
This Andrew Huberman/Joe Rogan video is making the rounds recently. The study from Hopkins cites a death toll of ~250,000 Americans/year from "medical error." I've seen a lot of posts either citing the exact number like a tally of ICD coded deaths or poo pooing because the study did a lot of extrapolation and made assumptions.
As an emergency physician, I see the debate from both sides:
1/The paper behind that headline never meant for 250k to be a literal census count. Like the 1999 "To Err Is Human" report, the authors wanted to sound an alarm: preventable harm is real, massive, and entirely invisible on death certificates.
2/ Getting a precise national number is impossible. The study’s math relied on extrapolating a small samples and assumed 100% preventability for key groups.
3/ In acute care, separating terminal illness from human error is rarely simple. Our current data systems are simply not built to capture that.
4/ Arguing over the exact math misses the point. Preventable harm remains an urgent, unsolved problem. We already know how to stop many of these failures; we just need the systemic will and data infrastructure to do it...but I'm not holding my breath. There is not RVU or CPT attached to preventing systemic error.
nitter.gq/VigilantFox/status/210…
Joe Rogan and Andrew Huberman look up how many Americans die from medical error each year, and the answer floors them both.
ROGAN: “Okay, so medical error is responsible for 251,000 deaths a year.”
HUBERMAN: “Whoa, that’s crazy.”
ROGAN: “Do you understand how insane that is? Imagine if that was Pop-Tarts.”
That number comes from a 2016 Johns Hopkins study published in the BMJ. It estimated medical errors kill 251,454 Americans a year, which makes them the third leading cause of death, behind only heart disease and cancer.
And it gets worse. A 2023 Johns Hopkins and Harvard study found that misdiagnosis alone kills an estimated 371,000 Americans a year and leaves another 424,000 permanently disabled.
If an airline killed that many people, nobody would ever board a plane again.
🔷 Epic’s End-of-Life Care Index was built on outpatients who rarely died and is being run on inpatients who often do. The largest validation yet shows what that number looks like when you read it as a probability.
🔷 Epic is having the busiest install season in its history while HCA takes 72 hospitals to MEDITECH. The real cost of Epic was never the go-live. It’s the people you need every year after.
🔷 Medicare is considering paying AI agents 60 to 80 percent of a physician’s rate. The vendor keeps the margin, and the on-call hospitalist keeps the liability.
🔷 Congress wants to send $1.3 billion to hospitals for cybersecurity. A two-person IT shop can’t run what that money buys, and the people holding its EHR hostage are not businessmen.
🔷 ECRI is calling for one patient, one record. The faster win is one vocabulary for errors and harm, and ECRI could build it without waiting for Congress.
✅️ Off-Label: A hundred million people are chasing readiness scores nobody validated against anything. Sound familiar?
hitdoc.substack.com/p/epics-…
Ok this will get a little health IT and LLM nerdy. The concept is that we are fighting a 20th century interoperabilty war in the 21st century. It would be like Russia and Ukraine duking it out with massed tank battles instead of drones. SNOMED, ICD, LOINC et al have their place but we need to leverage more modern tools to accomplish interoperability Nirvana.
hitdoc.substack.com/p/snomed…
🔷 Epic’s Agent Factory Build-A-Thon put AI agent-building in the hands of citizen builders. What happens when those agents run on ungoverned groupers, or who owns the silent failures once the builder leaves.
🔷 A CHOP informatics leader argues the real cost of switching away from Epic’s Chart with Art isn’t the software bill. It’s the opportunity cost.
🔷 Oracle Health is bleeding tier-one health systems to Epic, and the “sell it” and “harvest the data” theories both run into the same wall: you can’t monetize clinical data too messy to be usable.
🔷 Internal documents show CMS wants to expand WISeR, into cancer care, heart procedures, and emergency air transport, with no public accounting of who’s making the calls.
🔷 Prior authorization data is exempt from information blocking rules, and EHR vendors are charging payers accordingly. A regulatory carve-out has become a business model.
✅️ Off-Label: An AI data center’s promised jobs and its actual bill rarely match, and healthcare’s real bottleneck was never knowledge. It’s delivery.
hitdoc.substack.com/p/epics-…
“A vendor scoring between 60% and 85% on quality measures only gets penalized by 5% on the payment, and any score below 60% still receives 90% of the payment.”
Seems like fee for service “quality” metrics. If an orthopedic surgeon has a lot of complications and gets penalized 5% but does double the number of surgeries that is a better than even financial trade.
Medicare's AI prior authorization pilot was rushed and full of problems, new documents reveal statnews.com/2026/09/15/medi…
After the hospital started using Epic data to understand its length of stay problems, they had to act on the information. What did they do?
hitdoc.substack.com/p/reduci…
Gwynne Shotwell talks about creating a team that has talent and getting them bandwidth to do the work…and endless meetings and administrative work isn’t doing work. (~25 min mark). How many of you in healthcare IT experience this? youtu.be/iwBOQeFPAwg?is=gc47…
🔷 Epic is fighting to keep its antitrust discovery narrow, and it is leaning on HIPAA and TEFCA to help it both protect and restrict the data.
🔷 AHA’s better heart-risk calculator just went live inside Epic. It is better but still relies on good data.
🔷 A ransomware group is holding 3.5 million Veradigm patient records, and the way in was, once again, somebody else’s stolen password.
🔷 CMS built a prior-authorization pilot that pays vendors up to twenty percent of every denial, ignored warnings that it wasn’t ready, and hid the failure rate until a FOIA lawsuit forced it out.
🔷 A physician builder wants to ship his own ideas inside Epic, and the certification-and-approval gauntlet in front of him just got both worse and better.
✅️ Off-Label: New Jersey has studied its broken health plan governance for a decade without funding a fix, and health IT is making the identical mistake at a larger scale.
hitdoc.substack.com/p/epics-…