After deploying an AI platform with 50+ algorithms across 2,000 hospitals, you learn a lot. This week in NEJM AI I was fortunate to team up with my colleague Meghana Valluri from our Product team at Viz.ai to share our learnings over decade of putting AI into the clinical workflow. "Unfortunately, this story is not unique to sub-types of heart failure, or even cardiology more broadly. You can find it in every corner of the health care system: a condition presents with symptoms easily mistaken for something far more common, the clinician reasonably treats what it looks like, the patient does not quite get better, and the search quietly stops at a diagnosis that was good enough to move on from. Every clinician — the pulmonologist, the neurologist, the generalist — face their own version of the same challenge. It reflects the modern state of medicine — we have far more diagnoses to choose from, and more inputs to integrate to choose the right one. And it is precisely where AI tools can deliver. But getting them adopted goes far beyond a successful algorithm." https://lnkd.in/gsh4eJwU
The line about the search stopping at a diagnosis that was good enough to move on from is the part I keep rereading. We build on the consumer side rather than inside the hospital, and the adoption problem shows up there too, just wearing different clothes. Ours is that people stop reading anything that cries wolf. We ended up keeping our red flag list deliberately tight, so a high fever on its own comes back calm with things to watch for, because an assistant that escalates constantly trains people to ignore the one time it matters. Different setting, same tax. The algorithm is rarely what decides whether the thing gets used.
The expertise from professional and experienced cliicians gets even more critical as a layer above AL algorithms.
Lessons from deploying 50+ algorithms across 2,000 hospitals are exactly what the field needs more of. We appreciate you and Meghana sharing a decade of real-world clinical workflow learnings.
Exactly, Dr. Andrew. In clinical AI, a strong algorithm is only the starting point. The real challenge is integrating it into diagnostic pathways in a way that changes clinician behavior, reduces missed or delayed diagnoses, and still preserves clear clinical accountability.