Close Menu

    Subscribe to Updates

    Get the latest creative news from Healthradar about News,Health and Gadgets.

    Bitte aktiviere JavaScript in deinem Browser, um dieses Formular fertigzustellen.
    Wird geladen
    What's Hot

    Why LLMs Cannot Replace Deterministic Data Infrastructure

    18. September 2026

    Neptune Medical secures FDA nod for endoscopy robot

    18. September 2026

    Cognita Imaging wins FDA contract to test LLMs in evaluating radiology AI

    18. September 2026
    Facebook X (Twitter) Instagram
    Facebook X (Twitter) Instagram Pinterest Vimeo
    healthradar.nethealthradar.net
    • Home
    • Ai
    • Gadgets
    • Health
    • News
    • Contact Us
    Contact
    healthradar.nethealthradar.net
    Home»News»Why LLMs Cannot Replace Deterministic Data Infrastructure
    News

    Why LLMs Cannot Replace Deterministic Data Infrastructure

    HealthradarBy Healthradar18. September 2026Keine Kommentare5 Mins Read
    Share Facebook Twitter Pinterest LinkedIn Tumblr Reddit Telegram Email
    Why LLMs Cannot Replace Deterministic Data Infrastructure
    Share
    Facebook Twitter LinkedIn Pinterest Email


    Why LLMs Cannot Replace Deterministic Data Infrastructure
    Glynn Dennis, PhD, Chief Science Officer, Kythera Labs

    Healthcare AI has reached an inflection point. AI is rapidly moving from isolated demonstrations to production systems embedded in clinical, operational, and research workflows. A new generation of AI applications is emerging to support nearly every aspect of healthcare delivery, operations, and research.

    As organizations deploy these capabilities at scale, they are exposing an invisible semantic challenge that has quietly existed for decades. Clinical information is translated repeatedly as it moves through the healthcare system. Each translation can subtly change its meaning, even though we usually treat the result as if nothing has changed. Few are discussing this problem, and even fewer are solving it.

    Maintaining semantic continuity requires preserving clinical meaning across many forms, not just one. As clinical information moves through the healthcare system, it is continually interpreted and expressed in new forms, from conversations to clinical documentation, documentation to codes, codes to data, and increasingly, data to AI. Every transition is a transformation, and every transformation creates another opportunity for meaning to be lost, altered, or misrepresented. As these transformations accumulate, AI is increasingly asked to infer what is no longer explicit, no longer preserved, or no longer traceable to its original clinical meaning.

    This raises a fundamental question. Where should healthcare semantic understanding live? Should every AI model reconstruct it independently from increasingly transformed data, or should it exist as a shared semantic infrastructure that every model can rely on?

    The prevailing view is that increasingly capable models will eventually learn healthcare semantics. More clinical data, larger context windows, fine-tuning, and agentic workflows will progressively reduce the problem until semantic understanding simply emerges from the models themselves.

    I believe the opposite. Healthcare semantics should not emerge independently inside every model. It should be preserved explicitly as shared semantic infrastructure that is traceable, reusable, and governed independently of whichever LLM or agent happens to consume it. Every AI model should reason from the same trusted semantic foundation rather than reconstructing it independently.

    Even if future foundation models perfectly understood clinical meaning and were fluent in healthcare semantics able to translate between all vocabularies and ontologies, they would still face an important practical limitation. Healthcare depends on hundreds of shared vocabularies and coding systems for procedures, diagnosis, prescriptions and more. A model may correctly recognize methotrexate as a drug and return several associated National Drug Codes (NDCs). Yet herein lies the problem, methotrexate has more than six hundred NDC codes. Production systems for, say, patient finding or comparative effectiveness studies, often require every valid NDC, every descendant SNOMED concept, every applicable ICD code, or every relevant CPT code. That is not a reasoning problem. It is a fidelity and completeness problem. Healthcare AI requires both reasoning and completeness. LLMs increasingly provide the former. Semantic infrastructure must provide the latter.

    This is not a limitation of AI. It is a separation of duty. Models reason over clinical meaning. Semantic infrastructure translates, preserves, and governs it. As foundation models continue to improve and agents become increasingly autonomous, that separation becomes more important, not less.

    What does semantic infrastructure actually do?

    Consider the seemingly straightforward task of identifying patients with idiopathic pulmonary fibrosis (IPF). To a clinician, the diagnosis may seem obvious. To an AI model, the clinical concepts are recognizable. Yet accurately identifying an IPF population across real-world data is anything but straightforward.

    An IPF patient may be represented through pulmonology notes, radiology reports describing a usual interstitial pneumonia (UIP) pattern, pathology findings, pulmonary function tests, medications, ICD diagnosis codes, SNOMED concepts, and longitudinal patterns of care. Other patients may carry similar diagnoses but ultimately have connective tissue disease-associated interstitial lung disease, chronic hypersensitivity pneumonitis, or another fibrotic lung disease. No single representation tells the whole story.

    A capable language model can recognize these concepts and reason about them. Production healthcare systems must first translate clinical meaning across disparate clinical representations and coding systems while preserving semantic fidelity and ensuring completeness.  Diagnoses become ICD or SNOMED codes. Medications become RxNorm or NDC identifiers. Laboratory observations become LOINC codes. Procedures become CPT or HCPCS codes. Every translation must be comprehensive, traceable, governed, and reproducible before AI can reliably reason and act upon them.

    That is the role of semantic infrastructure. It establishes a trusted semantic foundation by resolving concepts across clinical language, coding systems, and data sources into a consistent, traceable understanding. It preserves semantic continuity as information changes form, supports deterministic retrieval of complete patient populations and concept sets, and provides every downstream model and agent with the same semantic foundation from which to reason.

    Why does this matter now?

    The stakes change dramatically once AI begins acting in healthcare workflows rather than simply answering questions in a chat interface. Historically, researchers and practitioners have compensated for gaps in clinical meaning. Clinicians recognize ambiguous documentation. Medical coders resolve inconsistencies. Analysts reconcile disparate coding systems. Real-world data scientists spend months translating, harmonizing, validating, and governing clinical concepts before they ever become evidence. Those activities are not incidental. They embody the necessary semantic work that makes healthcare data usable and trustworthy.

    Agentic AI changes that equation. Increasingly, AI will not simply summarize information or answer questions. It will trigger workflows, recommend treatments, generate evidence, draft regulatory submissions, and coordinate decisions across healthcare systems. Every one of those actions depends on accurate and comprehensive semantic understanding of the underlying clinical information.

    As AI assumes greater responsibility, the cost of semantic ambiguity changes. What was once an analyst’s inconvenience or a manual reconciliation becomes an operational dependency. The future of healthcare AI depends on two fundamentally different capabilities: 1) probabilistic reasoning and 2) deterministic semantic infrastructure. Confusing one for the other is becoming one of the industry’s biggest blind spots.


    About Glynn Dennis

    Glynn Dennis leads Kythera Labs Life Sciences and BioPharma initiatives, including research and product development. Glynn brings over two decades of expertise in Data & AI to Kythera Labs, having served in numerous scientific, data, and AI leadership roles across the industry, including at NIAID, Genentech, Bio-Rad Laboratories, and AstraZeneca.



    Source link

    Artificial Intelligence data Deterministic Infrastructure LLMs replace
    Share. Facebook Twitter Pinterest LinkedIn Tumblr Email
    Previous ArticleNeptune Medical secures FDA nod for endoscopy robot
    ekass777x
    Healthradar
    • Website

    Related Posts

    News

    Neptune Medical secures FDA nod for endoscopy robot

    18. September 2026
    News

    Cognita Imaging wins FDA contract to test LLMs in evaluating radiology AI

    18. September 2026
    News

    Healthcare’s agentic AI boom is outpacing governance: report

    18. September 2026
    Add A Comment
    Leave A Reply Cancel Reply

    Top Posts

    Marvel’s Dyasonic: A Superhero Powered by Sound—and Diabetes Tech

    13. Juni 2025750 Views

    Tombot Secures $6.1M to Bring Lifelike Robotic Puppy to Seniors with Dementia –

    19. Juni 2025725 Views

    Luna ring review | TechRadar

    26. Dezember 2025183 Views

    Hims & Hers Acquires Australian Digital Health Eucalyptus for Up to $1.15B

    20. Februar 2026165 Views
    Stay In Touch
    • Facebook
    • YouTube
    • TikTok
    • WhatsApp
    • Twitter
    • Instagram
    Latest Reviews

    Subscribe to Updates

    Bitte aktiviere JavaScript in deinem Browser, um dieses Formular fertigzustellen.
    Wird geladen
    About Us

    Welcome to HealthRadar.net — your trusted destination for discovering the latest innovations in digital health. We are dedicated to connecting individuals, healthcare professionals, and organizations with cutting-edge tools, applications

    Most Popular

    Marvel’s Dyasonic: A Superhero Powered by Sound—and Diabetes Tech

    13. Juni 2025750 Views

    Tombot Secures $6.1M to Bring Lifelike Robotic Puppy to Seniors with Dementia –

    19. Juni 2025725 Views
    USEFULL LINK
    • About Us
    • Contact Us
    • Disclaimer
    • Privacy Policy
    QUICK LINKS
    • Ai
    • Gadgets
    • Health
    • News
    • About Us
    • Contact Us
    • Disclaimer
    • Privacy Policy
    Copyright© 2025 Healthradar All Rights Reserved.

    Type above and press Enter to search. Press Esc to cancel.