In healthcare, knowledge is power – but historically, that knowledge has been locked in rigid guidelines, medical ontologies, and static decision trees. Today, large language models (LLMs) are starting to shift how we represent and use medical knowledge. These AI systems can interpret natural language and provide recommendations in real-time. It appears this evolution is set to revolutionise clinical decision support, making it more flexible, up-to-date, and context-aware than ever before.
In this post, we’ll explore how LLMs are transforming knowledge representation in healthcare, and what the ideal medical knowledge system might look like.
Traditional medical knowledge representation has relied on deterministic systems – think of rule-based expert systems or large ontologies like SNOMED CT. In these systems, facts are encoded explicitly: for example, “diabetes is a risk factor for heart disease” might be a link in a knowledge graph, or a clinical rule might fire an alert if a diabetic patient’s blood sugar is high.
Such approaches require human experts to manually curate relationships and rules. They are powerful but inherently rigid and labor-intensive.
Large Language Models, by contrast, absorb knowledge from free text. Instead of needing every medical fact pre-coded, an LLM is trained on millions of clinical notes, textbooks, and journal articles. As a result, much of medicine’s knowledge - causes, symptoms, treatments, nuances - ends up implicitly stored in the model’s parameters. An LLM doesn’t memorise guidelines verbatim, but it forms a statistical understanding of language that often aligns with medical truth. For example, a domain-tuned model can learn that “myocardial infarction” and “heart attack” are synonyms simply by observing how they are used interchangeably in text.
Deterministic vs. Probabilistic: A Paradigm Shift
It’s critical to recognise that LLMs represent a probabilistic approach to knowledge. They don’t follow a single hard-coded path to an answer; they weigh many possible interpretations and facts at once. At their core, these tools are extremely sophisticated “next word predictors,” with trillions of parameters finely tuned for this task.
This distinction - deterministic versus probabilistic - is essential for clinicians to understand as they begin using LLMs for medical knowledge retrieval and decision support.
CDSS have evolved over decades. Early systems were essentially static checklists and simple algorithms: if a patient has X, do Y. These were often derived from clinical practice guidelines, which themselves are distilled expert knowledge. Over time, CDSS became more sophisticated: drug-interaction checkers, allergy alerts, rule-based logic suggesting tests based on patient risk factors. These systems have saved lives and standardised care - but they’ve also introduced significant limitations.
Challenges of Rigidity and Obsolescence
A key challenge is rigidity. Traditional CDSS operate on fixed rules that don’t account for the messy reality of patient care. Guidelines usually address a single disease in an idealised patient, but real patients often have multiple comorbidities and atypical presentations. As a result, guideline-based systems can oversimplify or provide no help at all in complex scenarios. Sometimes, they give incorrect guidance - leading to frustration or even harm.
Clinicians frequently override these systems. Alert override rates in hospitals are high, and doctors often dismiss pop-ups that don’t feel relevant. This leads to alert fatigue, where important warnings are lost in the noise. The cognitive load of clicking through irrelevant alerts has been linked to physician burnout. Instead of being helpful, CDSS can feel like a hindrance.
Keeping knowledge current is another limitation. Medical knowledge doubles rapidly, and new evidence can render guidelines obsolete. Yet updating a CDSS rule is slow: it may require guideline committee revisions, software engineering changes, regulatory review, and clinical deployment - a process that can take months or years (depending on the EHR vendor).
Many decision aids end up out of date, reflecting older standards of care. Even their underlying knowledge bases often rely on manual curation. Ontologies like SNOMED CT and the Unified Medical Language System (UMLS) are curated maps of medicine, but they’re only as comprehensive as their curators make them.
All of this leads to clinician pushback. A tool perceived as inflexible or burdensome quickly loses trust. The story of IBM Watson for Oncology illustrates this well. Without reliability or context awareness, AI can become more of a burden than a benefit.
A New Model: Probabilistic Reasoning
Deterministic systems work on fixed logic: if condition X and Y are met, do Z. If not, they go silent. LLMs always attempt an answer. Even if information is missing, they can infer, express uncertainty, or provide options: “This could be due to A or B, but information is incomplete.”
In many ways, this is closer to how humans think under uncertainty. But it introduces new risks - like hallucinations, where the model generates confident but incorrect responses.
Our relationship with computers has historically been deterministic.
If I do X on a computer, I expect Y to happen every time. But this is changing. From “vibe coding” to AI-generated design prompts, we’re becoming more comfortable with probabilistic outputs. This shift is happening across every industry - and healthcare is no exception.
Across diagnostics and therapeutics we’re already very accepting and training very well in probabilistic reasoning, medicine already embraces high degrees of uncertainty in almost every facet. A very quick example, we trust high-sensitivity troponin tests with ~98-99% sensitivity, accepting a 1% margin of error when discharging chest pain patients. If clinicians can accept probabilistic diagnostics, shouldn’t we apply similar reasoning to AI-based decision support - so long as we understand the risks and use the tools appropriately?
Imagine working up a chest pain patient without a troponin in our arsenal, what makes a probabilistic decision support system different?
A New Paradigm: Blueprint of an Ideal Healthcare Knowledge System
Healthcare is often discussed in terms of incremental improvements - but what would great actually look like, across any facet of healthcare, this is very rarely defined and strived for. Based on everything we’ve learned, the ideal medical knowledge system would blend the structured reliability of traditional systems with the flexible intelligence of LLMs. Here are its key attributes:
Scalable and Real-Time Updating
The system must update seamlessly, with human oversight. New research or guideline updates should be integrated immediately. This could involve:
• Regular fine-tuning of models on new medical literature, or
• Retrieval-Augmented Generation (RAG): A semantic search system retrieves expert-approved content on demand and feeds it into the LLM as context.
RAG offers human oversight, reduces the need for expensive fine-tuning, and allows institutions to customise their knowledge base - for example, to reflect different antimicrobial resistance patterns or local operational policies.
System-Agnostic and Distributable
Medical knowledge should not be locked inside one hospital’s EHR or one vendor’s product. The ideal system functions like an internet of medical knowledge - accessible via open APIs or standardised protocols.
Hospitals, ambient AI tools, or point-of-care systems should all be able to tap into a shared medical intelligence layer, while still overlaying their institution-specific data and preferences.
Machine-Readable and Human-Readable
Knowledge outputs should serve both clinicians and machines. For example, a recommendation might include a structured plan, orders, codes, EHR flags as well as a plain-language explanation and rationale for the decision.
This duality enables both automation and transparency. A single source of truth should support safe clinical action and human understanding.
Conclusion
LLMs are ushering in a new era of knowledge representation in healthcare - one that’s probabilistic, natural-language-based, and vastly more adaptive than the rigid systems of the past. While risks like hallucination, uncertainty and systems being “confidently wrong” must be acknowledged, the benefits of faster updates, context-aware reasoning, and improved usability are compelling. It raises the age old debate of - at what point is it unethical not to use these tools for our patients?
%201.svg)
%201.svg)
