Snomed CT Entity Linking Challenge | Hacker News Reader

Snomed CT Entity Linking Challenge

drivendata.org

2 pointsnradov13 comments

Tangentially, but also to the point,

Software Development IDEs have (optional) autocomplete;

Unstructured Medical Coding interfaces could also have autocomplete,

such that when you type `icd:` or `snomed:` it presents a search interface for that particular medical terminology / vocabulary / system of classification / categories.

GNU Health > Issue tracker > "Freetext ICD-10 references as URIs (e.g. icd10:A01)" (2013) https://lists.gnu.org/archive/html/health-dev/2013-12/msg000...

I'm not convinced that auto complete would add much value to clinical workflows. Clinicians typically don't enter the codes themselves. In order to write those unstructured narrative blocks they might type them in manually, fill in the blanks on a template, or dictate. But in any case the clinician typically isn't entering billing or clinical codes. Usually certified medical coders add codes afterwards using specialized interfaces, often with some kind of NLP assistance.

Yeah, so it's at that workflow where they're not getting linked data edges out of their own unstructured data input.

To prompt them at that time for whether or not this referenced thing in a dictated .txt file is actually a thing with a URI and have them confirm that those are the correct annotations for their input would save a lot of time and money.

Requisite study for Linked Data clinical coding / informatics:

- HIPAA and unstructured notes, HIPAA and Linked Data; Informed Consent; Precision Medicine

- Tokens; NLP, stemming, conceptual entity recognition

- The LODcloud; a great big graph of Linked Open Data (that our data does not yet link to, is siloed separately from, does not yet have references to existing URIs in)

- RDFa: RDF-in-html-Attributes

- JSON-LD: JSON Linked Data

- Schema.org/MedicalEntity ; /docs/schemas.html -> "Health and medical types" https://schema.org/docs/meddocs.html

- TTS Text-to-Speech, Speech-to-Text; Multimodal models with RAG, LORA, RLHF, Transformer Self Attention tensor Networks, Benchmarking OpenAI Whisper with relative performance metrics, ONNX

- FHIR for EHR data portability (JSON-LD,)

UX Improvements:

- More IDE-like unstructured data entry

- Auto-annotate this text field in place

- Auto-annotate this pasted text

- Auto-annotate this text field and display the linked data annotations

- Display the revisions of the unstructured note -> annotated linked data document

- Autocomplete with search from e.g. icd:

- Frequently used annotations per Clinician/Provider

- Recently used annotations per Clinician/Provider

- Hover over visually distinguished auto-recognized entities and approve/clarify/reject

- Indicate which prov:Agent added which annotations; human (name(s)), AI (signed git revision of repo URI)

- Indicate degree of confidence in annotation (note that AGI hypergraph systems have TruthValue and also AttentionValue, like attention networks; and that AGI systems may someday also support CDSS: Clinical Decision Support Systems and Health Decision Support Systems)

Right, but physicians currently don't have time to do that and don't get paid for it. For provider organizations it ends up being more profitable to hire certified coders to do the coding so that physicians can focus on treating patients.

Why aren't they trained in the above, and who will ask large EHR vendors for those features, who would do the asking and is there a forms committee yet?

It doesn't make sense to train most physicians in coding. We already have a shortage of doctors. They need to focus on directly treating patients while lower paid staff handle other tasks.

Most EHRs already have features for building custom data entry template forms for common situations, which can allow for selecting specific codes. Large provider organizations typically have internal committees to design these forms in order to ensure some level of consistency in patient charts.

Hot take: Most clinicians should code at time of data entry in order to not play telephone with patients' medical information.

Transmission chain method > Application in studies of memory: https://en.wikipedia.org/wiki/Transmission_chain_method

https://en.wikipedia.org/wiki/False_memory_syndrome#Evidence... :

> Human memory is created and highly suggestible

Though nowhere is it shown that clinicians are even competent enough to add RDFa linked data annotations to their TXT unstructured notes with an encabulator.

It's clear that there's value in disambiguation; did they say "biomemetic" or "liquid metal" days ago on a dictaphone in a loud environment?

> Though nowhere is it shown that clinicians are even competent enough to add RDFa linked data annotations to their TXT unstructured notes with an encabulator.

You can add Linked Data (JSON-LD RDF) annotations with threaded markdown to document Trump vectors with W3C Web Annotations.

This [SNMOED-CT, ICD10-] competition could be scored with Web Annotations as the schema for the output.

W3C Web Annotations open source implementations:

- hypothesis/h: https://github.com/hypothesis/h

- https://github.com/linkeddata/dokieli

- https://www.w3.org/annotation/wiki/Implementations

* You can add Linked Data (JSON-LD RDF) annotations with threaded markdown to document term vectors with W3C Web Annotations.

ElasticSearch's Term Vector optionally returns with_positions_offsets, with_positions_offsets_payloads: https://www.elastic.co/guide/en/elasticsearch/reference/curr...

Meilisearch has showMatchesPosition: https://www.meilisearch.com/docs/reference/api/search#show-m... :

> `showMatchesPosition` returns the location of matched query terms within all attributes, even attributes that are not set as searchableAttributes.

Meilisearch > Comparison to Alternatives; Algolia, ElasticSearch, Meilisearch, Typesense: https://www.meilisearch.com/docs/learn/what_is_meilisearch/c...

And then now instead, Vector search engines: TODO

Vector database: https://en.wikipedia.org/wiki/Vector_database

You're missing the point. We have a shortage of clinicians. Any extra time spent doing data entry is time not spent treating patients. As long as claims are getting paid, coded clinical data quality is a secondary concern. Changing this would require major structural realignments in the entire healthcare system, which obviously won't happen any time soon.

When you realize that you need better than unstructured notes to do Clinical Decision Support ("AI"), which passes a Cost-Benefit Analysis.

> The objective of this competition is to link spans of text in clinical notes with specific topics in the SNOMED CT clinical terminology. Participants will train models based on real-world doctor's notes which have been de-identified and annotated with SNOMED CT concepts by medically trained professionals. This is the largest publicly available dataset of labelled clinical notes, and you can be one of the first to use it!

NER: Named Entity Recognition: https://en.wikipedia.org/wiki/Named-entity_recognition

awsome-medical-coding-nlp: https://github.com/acadTags/Awesome-medical-coding-NLP

awesome-ehr-deep-learning: https://github.com/hurcy/awesome-ehr-deeplearning

awesome-ner: https://github.com/smiyawaki0820/awesome-ner

awesome-bioie > Research groups: https://github.com/caufieldjh/awesome-bioie#groups-active-in...

SNOMED-CT as RDF: https://sphn-semantic-framework.readthedocs.io/en/latest/ext...

...

SNOMED-CT is a Medical Terminology.

SNOMED-CT: https://en.wikipedia.org/wiki/SNOMED_CT

Traditionally, a Terminology Service provides a query endpoint over a schema like SNOMED-CT (which is a nonredistributable medical XML schema, which is challenging for Open Linked Data):

https://github.com/NCIP/lexevs

https://github.com/OpenConceptLab

https://openconceptlab.org/terminology-service/

https://github.com/MedevaKnowledgeSystems/pymedtermino/blob/... :

> For SNOMED CT, ICD10 and MedDRA, the data are not included (because they are not freely redistribuable) but they can be downloaded in XML format. PyMedTermino includes scripts for exporting these data into SQLite3 databases.

Reply on news.ycombinator.com