Often, we get questions about how SNOMED CT is used in different situations.
I think the best answers will most likely come from looking at what has been communicated to us via the community in EXPO presentations.
For example, how is SNOMED CT being used with NLP and AI.
So how might you get that answer?
Here is one approach which extracts that answer from past presentations
I go to our interactive SNOMED CT YouTube Video Search
Select the category AI and Natural Language Processing (NLP) category from the drop-down
This shows currently 90 presentations
Click on Copy all filtered URLs
I am using NotebookLM, so I add these links as sources to a new notebook and then ask it questions.
The question I asked, although you could ask more specific questions, was “Give a brief overview in 500ish words on SNOMED CT, NLP and AI”
Here is the answer (Note :in the live version, this also gives references to the presentations from where the information was sourced, so you could watch the presentation if you are interested)
“The integration of SNOMED CT with Artificial Intelligence (AI) and Natural Language Processing (NLP) addresses a fundamental healthcare challenge: approximately 80% of clinical data exists as unstructured free text, making it difficult to analyze for secondary uses like research or population health.
The “Superpower” Relationship
AI and SNOMED CT are increasingly viewed as complementary forces. AI provides the speed and scalability to process vast amounts of data, while SNOMED CT provides the semantic layer and logical structure necessary to ensure that AI outputs are accurate and clinically valid. They grant each other “superpowers”: SNOMED makes AI safe and interpretable, while AI makes SNOMED CT accessible and usable at scale.
NLP and Entity Linking
NLP is the primary mechanism used to convert clinical narratives into SNOMED CT codes, a process known as Entity Linking or named entity recognition.
• Structuring the Unstructured: Tools like MedCAT and Clinithink scan documents to extract concepts, converting narrative text into structured data,.
• Context Awareness: Advanced NLP does not just match keywords; it must understand context. This includes detecting negation (e.g., “no signs of cancer”), temporality (history of vs. current), and experiencer (patient vs. family member), Without this, a patient might be incorrectly flagged as having a disease they do not have.
Generative AI and Large Language Models (LLMs)
The emergence of LLMs has shifted how SNOMED CT is utilized. While powerful, LLMs are probabilistic and can “hallucinate” by inventing incorrect codes or facts,
• Retrieval Augmented Generation (RAG): To mitigate hallucinations, developers use RAG architectures. The AI queries a trusted SNOMED CT terminology server to retrieve valid concepts before generating a response, ensuring the output is grounded in the standard.
• Vector Embeddings: Concepts are often converted into vector embeddings—mathematical representations where similar concepts appear close together in a multi-dimensional space, This allows AI to perform “semantic searches,” enabling it to understand that “kidney failure minus kidney plus heart” equals “heart failure”.
Key Applications
-
Predictive Analytics and Digital Twins: By converting patient histories into sequences of SNOMED codes, AI tools like Foresight can create “Digital Twins” to forecast future health trajectories, predicting the next likely diagnosis or medication.
-
Automated Registries: NLP pipelines are used to automatically populate clinical registries (e.g., for cardiology or cancer) from discharge summaries, reducing the burden of manual data entry.
-
Terminology Maintenance: AI assists in the authoring and quality assurance of SNOMED CT itself. It helps detect structural anomalies, automate the translation of SNOMED into other languages, and identify missing concepts.
In summary, SNOMED CT acts as the “bones” or logical foundation that AI systems rely on to reason effectively over healthcare data.”
I think this is a powerful way to use Expo Presentations to answer specific questions
Maybe have a go at the same or a different topic and let me know how you get on.