Healthcare Skipped Its Early Adopters

Every technology healthcare ever adopted arrived on the same schedule. A bleeding-edge 5% went first, broke things, and told everyone what they learned. The majority followed, and the laggards came when they had no choice. It is slow, and easy to resent. It is also how a field learns what a new tool actually does.

Jeremy Harper’s point, on this week’s episode, is that it did not happen this time.

“We didn’t have that this time,” he says. “This time everyone implemented all of these solutions at the same time. And when you have that, it really creates a very different beast.”

Harper is a biomedical informatician, a former Chief Research Information Officer at Regenstrief and Eskenazi Health, and the author of Large Language Models (LLMs) for Healthcare.

The number nobody has

Ambient documentation is the clearest case. It went everywhere, fast, because clinicians wanted it and the marketing ran ahead of the evidence. The architecture is simple: record the voice, transcribe it, hand the text and a prompt to a language model, get a note.

And nobody can tell you how often that note is wrong.

Not because the research is hard, but because the evidence is destroyed on purpose. “The only way we do that is if we record and keep those recordings,” Harper says, “and most vendors are throwing away those recordings as soon as they’re transcribed.” No ground truth survives to check the note against, so the error rate is not a number anyone possesses.

A physician at one of his workshops told him the tradeoff was fine: “It doesn’t really matter to me if 5% of them have errors. It gets my job done faster.” Harper’s reply is the sentence that has stayed with us. “The errors won’t matter until they do matter. And it’s hard to predict what errors are going to matter ahead of time.”

He knows the shape of that personally. His own chart once carried a diagnosis of clinical depression he did not have, filed to his insurance company, and it took him years of fighting to get it removed. Some errors are an annoyance. Others follow you into an underwriting decision.

An attack surface nobody planned for

There is a second failure mode hiding in the same architecture. Most deployments skip speaker diarization, because working out who is talking costs more, so the note never really knows whose words it is recording. Anyone in the room can supply them. Harper’s example is not hypothetical: “If you can slip into the conversation, we’re going to prescribe GLP-1s for a patient, or you’re going to prescribe a type 1 narcotic, the system will automatically show up for the physician as, hey, you said that you’re going to prescribe this.” His summary is the part that should keep a security team up at night. “Large language models have opened up vectors of attack for a healthcare system that we’ve never thought about in the past.”

Whose problem is it?

The health system’s answer is that the physician signs the note, so the physician owns the error, and a bad enough one costs a license. Harper does not accept the arrangement. “It’s not about the physician, it’s about the patient.”

His fix is a playbook healthcare already owns, from research informatics. We never claimed a de-identified dataset was 100% de-identified. As he puts it, “we’re going to measure what level of risk that we’re putting forward, and then we’re going to get a trained professional to sign off on this.” Do the same for ambient notes. Which requires keeping the recordings, exactly what the industry decided not to do.

He also hands buyers a one-question maturity test. Ask the vendor what prompt they use to turn the transcript into the note. It is the load-bearing component, and the one nobody audits.

The knowledge we keep paying for and deleting

Then there is the part that visibly stings him. The National COVID Cohort Collaborative pooled data from academic medical centers nationwide, and informaticians built up a body of code, workbooks that knew how to ask real questions of the OMOP data model. The enclave was rebuilt on a new Palantir instance. The old code workbooks did not come with it.

The timing is what hurts. Language models are extraordinarily good at generalizing from a corpus of worked examples, and we threw the corpus away exactly when it became most valuable. Leon’s reframe on air: it is the tragedy we complain about with real-world data, happening now to our operational data, which nobody recognizes as a training asset, so it gets deleted.

Which points at where the real wins are, and they are aggressively boring. Harper built one: a language model reads an incoming data-request ticket, works out which of the questions analysts always ask are missing, and emails the requester for them before an analyst has burned an hour. “The type of thing we’re not spending energy on,” he says, “but it’s going to be the most transformative.”

That is the throughline. The value is not on the frontier. It is in the plumbing: the prompt nobody audits, the recordings nobody keeps, the knowledge we already paid for and threw away.

The dumbest model you can get away with

The frontier may matter even less than the vendors are betting on. Harper already runs open-weight models like Gemma and Qwen on his own machine, doing work that needed a data center a year ago; a current iPhone, he points out, outclasses a supercomputer from twenty years ago. The hosted providers are subsidizing tokens to win the market, and that subsidy will not last. Healthcare never gave up its own servers, so when the bill comes due, the intelligence moves back in-house.

Which sets up his parting line, offered almost as a gift. “Every healthcare organization is going to choose the dumbest large language model they can get away with to do any task. Because it’s the cheapest way to do it. You don’t want to throw more intelligence than you have to.”

Leon’s answer was the old NASA joke: a billion pounds of explosive, two million parts, every one built by the lowest bidder. But it does make it to orbit once in a while.

Listen to the full conversation: https://practicalaiinhealthcare.com/episodes/#harper

You Might Also Enjoy

• S1E46 with Peter Embi — The host of this same worry from the other side: Embi coined “algorithmovigilance” and is the guest who told us this is the first time in his career he has tried to slow adoption down.
• S1E27 with Charlie Harp — On healthcare data quality and the PIQI framework, the backbone that any reusable, queryable knowledge base depends on.
• S1E43 with Zak Kohane — Who controls healthcare data and how to measure what clinical AI is actually optimizing for, the governance counterpart to “show me your prompt.”