Karin Verspoor's first day at Los Alamos National Laboratory, a colleague walked up and said he had a project that could use her skills — natural language processing. She asked what the project was. Protein function prediction, he said. She looked at him and asked what a protein was. She hadn't done biology since year nine.

That's not a self-deprecating anecdote she tells for effect. It's the actual starting point of a career that's spent two decades since applying computational linguistics to some of healthcare's hardest data problems, and it says something true about how expertise transfers across domains — you don't need to know the field first, you need to know how to make sense of the language the field uses to describe itself.

Verspoor is Executive Dean of Computer Science at RMIT University. Her path there ran through cognitive science as an undergraduate — an "aha moment" realizing that psychology-of-language classes and AI classes were describing the same underlying problem — then a PhD at the University of Edinburgh on how words connect to represent meaning, then an AI startup that didn't survive the dot-com crash, then Los Alamos, where the protein question redirected everything.

From proteins to patients

The protein work turned out to be a rehearsal for something bigger. Biomedical research findings live in scientific papers — relationships between proteins, genes, diseases and chemicals, expressed in language rather than structured data. Verspoor's team built tools to machine-read that literature and extract the relationships buried inside it. The vocabulary was brutal at first — phosphorylation, methylation, ubiquitination — "is this English?" she remembers thinking. But a little domain knowledge went a long way once she had language-processing tools to apply it.

Electronic health records posed the same problem at a different scale. Clinical documentation is language, because language is how people naturally communicate what happened in an encounter. The challenge she'd already solved for scientific literature — turning free text into structured, usable information — applied directly to hospitals and clinics.

Forty years of unglamorous infrastructure

The part of the conversation worth sitting with longest isn't about large language models at all. It's about the Unified Medical Language System, built by the US National Library of Medicine roughly forty years ago. UMLS ties together standards that evolved independently, for different reasons: the International Classification of Disease, organized hierarchically by body system and now the backbone of billing systems worldwide, and SNOMED, built by clinicians to standardize the language of signs and symptoms. UMLS is the map that says this SNOMED term connects to that ICD code, because they share underlying meaning.

None of this was built with AI in mind. It grew out of a library needing to catalog things — the same instinct behind the Dewey Decimal System, applied to medicine. But when digital health records arrived, that decades-old cataloguing work became the anchoring framework everything else could hang off. Verspoor's point, made almost in passing, deserves more weight than it got in the moment: the unglamorous work of building shared vocabulary is exactly what makes the flashy AI layer usable later. You don't get the second without the first.

Why dictionary lookup breaks on "no evidence of infection"

Early clinical NLP leaned on that vocabulary directly — dictionary lookup, matching terms in the record to terms in UMLS. Simple, and it works as a starting point. But medicine runs on ruling things out. "No evidence of infection" and "infection" share a keyword and mean opposite things. A system that just matches terms will flag both as positive findings — which, as I put it to Verspoor, is like telling a child not to do something and having them hear only the instruction to do it. Negation isn't an edge case in clinical text. It's a huge share of the diagnostic process, and it's exactly the kind of context-dependent meaning that keyword matching can't touch.

That's the gap large language models were supposed to close, and to a real degree they have — they hold context in a way a lookup table never could. But Verspoor is blunt about where they still fall short: the big models haven't been trained on the full diversity of clinical documentation, so there's still distance between what they know in general and what a specific clinical context requires. Closing that distance still runs through domain-specific solutions built on the structured vocabulary work — not around it.

The box-ticking problem

Where clinical AI has actually landed in practice today is AI scribes — tools that listen to a consultation and produce a summary, with the clinician required to confirm it's accurate before it's filed. It's human-in-the-loop by design. Verspoor's worry is what happens to that design under repetition: a reviewer who checks carefully nine times and finds the summary correct nine times will, on the tenth, be tempted to tick the box without reading closely. It was right the last nine times — why would this one be different? That's not hypothetical. It's the standard failure mode of any quality-by-inspection process, and it's why she expects more tooling built specifically to flag when something looks off, rather than relying on human vigilance to catch every case cold.

The same logic pushes her toward blending generative AI with older, deterministic methods — information extraction, normalization, narrow classification models that predict rather than generate — with a human always positioned to make the final call rather than have it made for them.

Know more than the model

What she said last is worth keeping. You can only catch an AI system's errors if you know enough about the domain to recognize when something's wrong. Large language models are good at aggregating and summarizing what's been said thousands of times before. What they don't have access to is your specific situation — the particulars of this patient, this customer, this deployment. "What's not in your data?" is the question Verspoor asks before trusting any model's output.

Subscribe & Review:

If you found this episode valuable, please subscribe and leave a review on Apple Podcasts, Spotify, or YouTube. It helps other technology leaders discover these conversations.

Sound Bites

• “What’s not in your data?”
• “I literally looked at him and said, what’s a protein?”
• "Every situation is unique — and that’s what a human can bring that the LLM doesn’t have access to.”
• “People don’t always use the terminology correctly.”
• “I checked it nine times and it was right… the tenth time, they just tick the box.”

Chapters

00:00  —  Introduction

00:54  —  Karin’s journey: from cognitive science to NLP

05:25  —  “What’s a protein?” — How molecular biology led to healthcare

09:43  —  Electronic health records and the unstructured data challenge

11:20  —  UMLS, ICD, and SNOMED: structuring medical terminology

14:58  —  How organizations large and small use these standards

17:32  —  The negation problem: why dictionary lookup isn’t enough

18:48  —  Where healthcare analytics is heading in the next 3–5 years

20:33  —  Guardrails: preventing AI hallucinations in clinical settings

23:36  —  Staying human-in-the-loop: the role of domain knowledge

25:12  —  Closing thoughts