What Did Fifty-Two Conversations About Healthcare AI Actually Teach Us?

A year ago this was an experiment neither of us was sure about. Fifty-two episodes about healthcare AI later, we put the interviews down and asked what they added up to. Two things came out, and the second surprised us more.

The show started talking to itself. Mid-season, guests began answering each other. Ted Shortliffe answered a question Bob Wachter had raised earlier, having gone back and listened to it. Dave deBronkart named Hugo Campos thirty-one episodes before Campos was a guest, and Chris Dymek credited Campos from a completely different direction. Zak Kohane and Wachter converged on the same sentence with no contact between them. The deflationary read is that it’s a small field, and we considered that. We don’t think it’s right: the field is converging, and a year of conversations is finally enough to see it happening.

That changes the status of everything below. These aren’t five opinions we assembled. They’re five conclusions that arrived independently.

The AI was never the hard part. The data is, and it still is. Charlie Harp measured lab data quality at roughly 70% against USCDI, which is to say about a third of it falls short, and found a facility that had mapped every hematology test to the LOINC code for glucose. Aaron Kamauu showed one code choice costing upwards of 30% of a cohort. Given thirty seconds to advise the rest of pharma, Matt Truppo spent all of it on data foundations. Mika Newton gave it the sharpest form: records travel, but they don’t translate. The models became infrastructure while we argued about them, and it’s boring the specific way electricity is boring. Nobody’s impressed that the socket works. They’re impressed when the washing machine returns clean clothes.

Almost nobody’s watching after go-live. Peter Embi coined algorithmovigilance and built VAMOS at Vanderbilt, an air traffic control tower for clinical AI. Across fifty-two episodes he’s the only guest who brought a system rather than a complaint. Yuri Quintana supplied the reason it matters: sepsis models trained at a few hospitals and deployed at a hundred, performing no better than a coin toss.

An interface built for a novice is destructive to an expert. Vimla Patel’s research shows experts reason by pattern recognition rather than dissecting a problem backwards, so a system that scaffolds a beginner degrades someone with ten thousand hours. Scott Snyder got there from teaching prompt engineering. Most tools are built as though expertise were a quantity rather than a direction.

And the question we keep skipping is “compared to what.” Hold a system against perfect safety and efficacy and everything looks terrible. Hold it against real clinicians working real days, where medications get given wrong and diagnoses missed, and the arithmetic changes. Wachter argued it, Kohane arrived at it independently nineteen episodes later, and Adam Rodman pointed at urgent-care telehealth as the first viable slice.

Underneath all of it sits money, said quietly nearly every time. Alvin Liu named the second valley of death, and put the question plainly: we know how to do it, but who’s going to pay? Amar Mandavia and his colleague chased private insurers before working out that people who develop schizophrenia become Medicaid and Medicare’s problem, which answers the investment question in the worst way. The payer and the beneficiary are rarely the same party, and that’s the oldest problem in health IT wearing new clothes.

We closed on what we couldn’t resolve. Dymek argues AI literacy needs an institutional answer, and reading eventually got schools and libraries. Neither of us could say what the equivalent would be. And there’s a real tension between deBronkart, who points out you can’t stop patients using this technology, and Campos, who went and built patient-facing tools himself: the same system doesn’t perform equally well for everyone holding it. Leon put himself on the wrong side of that line rather than talking about patients in the third person, which is the right place to stand.

Season 2 starts next Sunday with Raihan Faroqui on the front and back office, then Randi Seigel of Manatt on state regulation. Tell us what you want to hear: info@practicalAIinHealthcare.com

Listen to the full conversation: https://practicalaiinhealthcare.com/episodes/#reflections-8

You Might Also Enjoy

• S1E27 with Charlie Harp — The data-quality argument in full, including the lab that mapped every hematology test to the code for glucose.
• S1E46 with Peter Embi — Algorithmovigilance and VAMOS, the one working answer anyone brought us on monitoring deployed clinical AI.
• S1E24 with Bob Wachter — Where “compared to what” came from, and the episode Ted Shortliffe later answered on air.