Data challenges with AI: Omissions, bloated problem lists and unnecessary token burn

Hallucination is not the biggest risk in clinical AI. Omission is — and it is far harder to detect. John Laursen, SVP at IMO Health, has spent his career on the layer of healthcare AI that gets the least attention: clinical terminology and the semantic data infrastructure underneath every model deployed in a hospital. IMO Health's terminology has been built and curated since 1994 and now sits behind roughly 12 billion terminology search transactions a year across US provider organisations and every major EHR. In this interview with Tjaša Zajc, Laursen makes the case that structured data was necessary but is no longer sufficient. AI reasoning across a thirty-year patient chart needs semantic continuity — an understanding that clinical language recorded in the 1990s and language recorded today can mean the same thing. Without it, health systems are investing in models that cannot reliably interpret their own records. The conversation also covers what happens when ambient AI scribes get it wrong, why accumulated clinical data has become a computational cost rather than an asset, and why clinician trust is the constraint that determines how fast clinical AI can move. Guest: John Laursen — Senior Vice President, IMO Health (Chicago, US) Host: Tjaša Zajc — Faces of Digital Health What the conversation covers: - Why omissions, not hallucinations, are the underrated risk in clinical AI - What a semantic layer does that structured data alone cannot - How clinical terminology maps to SNOMED CT and ICD-10 — and why those code sets were built for different purposes - Ambient AI scribes: what happens when a model mishears or over-infers a diagnosis - The billing and clinical consequences of an error entering the patient record - Why problem lists hundreds of entries long now cost money in token burn - Patient-generated and AI-generated content entering the EHR, and why health systems resist it - Translating lay language into clinical terminology without losing specificity - Ambient documentation, billing intensity and friction with payers - How data quality expectations differ between the US, the NHS, the Gulf states and Singapore - Who governs clinical data as coding complexity increases - Why AI performance breaks down on rare disease and the difficult 20% of cases - Knowledge graphs as a grounding source for clinical AI models - What health systems should require from AI vendors before clinical deployment Chapters: 02:20 Why the data layer decides what clinical AI can do 03:27 Inside IMO Health: 12 billion terminology searches a year 05:36 Keeping terminology current: SNOMED, ICD-10 and clinical governance 07:35 The semantic bridge: why structured data alone is not enough 10:17 Patient language versus clinical language in the record 12:23 When an ambient scribe mishears: clinical and billing consequences 14:53 Omissions, bloated problem lists and unnecessary token burn 19:12 Outside the US: the NHS, the Gulf, Singapore and coding complexity 20:46 Who governs clinical data as complexity increases 23:35 Patient-side AI recorders and resistance to external data 26:08 Ambient documentation, billing intensity and payer friction 29:21 The last 20%: rare disease, model limits and AI governance 33:38 Grounding, clinician trust and the cost of misfiring Faces of Digital Health: Website: https://www.facesofdigitalhealth.com Newsletter: https://fodh.substack.com Spotify: https://open.spotify.com/show/4cElKJHrauyP6QJQaCkvdY Apple Podcasts: https://podcasts.apple.com/gb/podcast/faces-of-digital-health/id1194284040 LinkedIn: https://www.linkedin.com/company/faces-of-digital-health #digitalhealth #healthcareAI #clinicalinformatics #EHR #ambientAI #interoperability #healthdata #SNOMED #healthIT #medicalcoding