Practical AI in Healthcare

In their seventh Reflections episode, and their fifty-first overall, Steve and Leon look back across six conversations: Mika Newton on interoperability that finally started working, Peter Embi on monitoring clinical AI after deployment, Vimla Patel on how clinicians actually reason, Renee Deehan on an engine built so it can't hallucinate, Christine Dymek on AI literacy, and Jeremy Harper on an adoption curve that broke. They decide against forcing a single grand lesson and find three threads anyway: keeping a human in the loop as a deliberate design decision, the national clearinghouse for AI errors that still doesn't exist, and whether literacy is even the right word for what healthcare workers need.

Every technology healthcare has adopted moved through the same curve: a bleeding-edge few went first and documented what broke, and the majority followed. Large language models skipped that entirely. Jeremy Harper, a biomedical informatician who has worked at Epic, Ohio State, and Regenstrief, and who wrote Large Language Models (LLMs) for Healthcare, explains what we gave up by going all at once. Ambient scribes are everywhere, and because most vendors discard the audio as soon as they transcribe it, nobody can say how often the notes are wrong. He offers a fix borrowed from de-identification, a one-question test for any AI vendor, and a lament for the reusable knowledge we keep paying for and deleting.

Chris Dymek came to healthcare informatics by way of philosophy, and she still thinks like a philosopher: the first question is what we actually mean. As Director of Digital Healthcare Research at AHRQ, she funded AI research, helped launch work on AI and patient safety, and wrote a request for information asking healthcare organizations how they thought about AI literacy. It was never published. She left in May 2025 and carried the work into the DCI network instead. In this conversation, she lays out the distinction at the center of her framework, knowing-that versus knowing-how, the three constituencies who each need something different, and why a literate staff and a literate patient population are what let a health system move forward without breaking things.

The internet is full of wellness advice built on a single cherry-picked study. Renee Deehan, a molecular and cell biologist who leads science and AI at InsideTracker, spent two decades building the opposite. In this episode, she explains why the core of their recommendation engine is symbolic AI, knowledge representation and reasoning, rather than a large language model: it's deterministic, fully auditable, and by design cannot hallucinate. The LLMs are fenced off to chat and summaries, while humans still write and review every recommendation against convergent clinical evidence. She and the hosts dig into a 20,000-user outcomes study, the discipline of refusing to claim causality, the MCT-oil case where the system decides not to recommend, and how a data-science team grew its own AI literacy instead of hiring it.

Dr. Vimla Patel has spent four decades studying how physicians actually reason, and what happens when technology ignores it. In this episode, she explains the difference between forward reasoning (the fast, pattern-driven hallmark of expertise) and backward reasoning (the slower, hypothesis-testing mode of novices), and why most clinical AI is built for the wrong one. Through two vivid cases, a textbook expert diagnosis and a near-fatal potassium overdose driven by a flawed order-entry system, she shows how good design preserves clinical judgment and bad design erodes it. She closes with a warning about confident AI and the quiet loss of independent thinking.