What happens when the rules for getting AI into clinical care are written by someone who has spent his career inside both the advocacy world and the government? In this episode, we talk with Jeff Smith of ONC at HHS, the first government official on Practical AI in Healthcare. Smith walks us through ONC's proposed HTI-5 rule, including a striking move to treat AI agents as "users" with the same data-access rights as clinicians, and a new question about whether blocking data from being written back into the EHR is itself information blocking. We also dig into the limits of what a regulator can actually do, and why the real work is coordination across agencies rather than control from any one of them.

https://practicalaiinhealthcare.com/

https://www.youtube.com/@PracticalAIinHealthcare

On National Nurses Day, Practical AI in Healthcare welcomes its first nurse: Sarah Rossetti, RN, PhD, of Columbia University. Her CONCERN early warning system takes an unusual approach to predicting patient deterioration. Instead of modeling a patient's vital signs and labs, it models the nurse's documentation behavior, since the frequency and timing of charting reflect clinical concern long before the numbers move. In a 74-unit randomized trial of more than 60,000 patients, published in Nature Medicine, CONCERN was associated with a 35.6% reduction in instantaneous mortality risk. Rossetti and the hosts unpack the method, the counterintuitive rise in ICU transfers, equity safeguards, and what ambient AI means for the signal.

https://practicalaiinhealthcare.com/episodes/#S1E39

More on Sarah Rossetti's work: https://www.dbmi.columbia.edu/profile/sarah-collins-rossetti/

In their fifth Reflections episode, Steve and Leon look back across six conversations (Matt Truppo at Sanofi, Ted Shortliffe, Barry Chaiken, David Hidalgo-Gato, and Danny van Leeuwen) to ask a sharper question: how specialized does AI have to be to actually work? The throughline is depth. The LLM is a commodity, and so, increasingly, is the generalist agent. What stays scarce is specialization in a workflow, the revival of symbolic methods like knowledge graphs, the literacy that separates an AI's ~95% solo accuracy from the under-35% people get using it themselves, and leaders willing to use themselves as the test rig. After 37 episodes, the technology is no longer the question. The specificity of the work around it is.