Conceptual hero image for healthcare AI safety featuring Dr. Peter Embi’s silhouette next to an open, malfunctioning control panel with a red alarm light, representing the monitoring of clinical algorithms and medical AI drift.

The Doctor Who Diagnosed Himself  — and Now Wants to Watch the Algorithms

07/19/2026 | Practical AI in Healthcare |

Peter Embí opens his own keynotes with a patient case. The patient is him.

For nearly two decades, he had symptoms that kept getting explained away: headaches, sweating, tremors, a racing heart. As he puts it, “Like a good doctor, I explained most of these things away in my head.” In 2017 it escalated to hypertensive crises, and one night a top hospital’s ER discharged him with “atypical migraine and hypertension” while his systolic pressure read 220. It took a sixth attack, mid-way through Grand Rounds, for Embi to work the problem like the internist he is: headache, sweating, hypertension, the classic triad. He called his neurologist and said he thought he had a pheochromocytoma, a rare adrenal tumor. He was right. He was cured, and he’s now eight years out. It is a tumor that, historically, was as likely to be found at autopsy as in life.

Zebras are what AI is good at

“If you hear hoofbeats, think horses, not zebras,” Embi told us, citing the line every clinician learns. Rare diseases get dismissed precisely because they’re rare. But the pattern-matching that humans skip is exactly what well-built AI does well. As a rheumatologist, Embi watched patients carry undiagnosed lupus for two or three years; his group is now showing that AI-assisted routing can compress that to months or even weeks, which also frees specialist slots for the people who genuinely need them.

The villain: unmonitored AI

Here’s the catch that animates his work. We train clinical models carefully, test them, deploy them, and then mostly stop watching. “Once they’re actually in practice, we don’t really have good mechanisms to monitor and understand whether they’re continuing to have the effect we expect, unless we happen to do a study or something catastrophic happens.” Models drift. A tool validated at Vanderbilt can fail across the street, or even floor to floor.

On a 2019 panel, asked for his biggest worry, Embi named it and coined a word: algorithmovigilance. The analogy is pharmacovigilance, the safety surveillance that follows a drug to market. Co-host Steve Labkoff has lived that culture and has argued, with Dean Sittig, that there’s no “MedWatch for AI.” Embi built the missing piece: VAMOS (Vanderbilt’s Algorithmovigilance Monitoring and Operations System), an “air traffic control tower” that maps each model to real outcomes (a sepsis model to sepsis rates and mortality), now being open-sourced and networked across institutions so the evidence we lack can accumulate.

Brakes let you go faster

Embi is no skeptic. He notes the unsustainable costs and access gaps that make AI necessary, and warns against letting fear freeze adoption: “If there’s not a significant positive gain to be had, then no amount of risk is worthwhile.” The reframe we kept coming back to: why do cars have brakes? So you can go faster. Monitoring isn’t the brake on AI in healthcare. It’s what lets us accelerate without flying blind.

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

You Might Also Enjoy

  • S1E43 with Isaac “Zak” Kohane — A Harvard informatics chair on who controls healthcare data and the values baked into clinical AI, the governance backdrop to Embi’s monitoring work.
  • S1E33 with Ted Shortliffe — Fifty years of clinical AI from a decision-support pioneer, the lineage behind the “what’s actually different about AI” question.
  • S1E41 with Hugo Campos — A patient advocate building his own tools, the empowerment-versus-peril tension Embi weighs for patient-facing AI.