# Moving from umls to snomed-ct

**URL:** <https://discourse.cogstack.org/t/moving-from-umls-to-snomed-ct/177>\
**Category:** MedCAT\
**Created:** [March 24, 2023, 9:06am UTC](https://discourse.cogstack.org/t/moving-from-umls-to-snomed-ct/177 "2023-03-24T09:06:26Z")\
**Posts on this page:** 5\
**Page:** 1

<div class="post-metadata">

**Author:** ![sangeetabose](https://avatars.discourse-cdn.com/v4/letter/s/eb8c5e/32.png) [@sangeetabose](https://discourse.cogstack.org/u/sangeetabose)\
**Post date:** [March 24, 2023, 9:06am UTC](https://discourse.cogstack.org/t/moving-from-umls-to-snomed-ct/177/1 "2023-03-24T09:06:26Z")

</div>

HI… how easy or difficult is to move from a umls trained medcat to snomed trained medcat. Would i need to start afresh and retrain on the snomed medcat.

thanks  
Sangeeta

---

<div class="post-metadata">

**Author:** ![tomolopolis](https://avatars.discourse-cdn.com/v4/letter/t/6de8d8/32.png) [@tomolopolis](https://discourse.cogstack.org/u/tomolopolis)\
**Post date:** [March 27, 2023, 2:47pm UTC](https://discourse.cogstack.org/t/moving-from-umls-to-snomed-ct/177/2 "2023-03-27T14:47:09Z")

</div>

hi @sangeetabose - depends what your downstream use case is. There are mappings from UMLS to SNOMED, but the mapping can be of one-to-many and many-to-many mappings, so I wouldn’t recommend converting UMLS to SNOMED or vice versa.

I’d recommend starting fresh, MetaCAT models will be the same

---

<div class="post-metadata">

**Author:** ![sangeetabose](https://avatars.discourse-cdn.com/v4/letter/s/eb8c5e/32.png) [@sangeetabose](https://discourse.cogstack.org/u/sangeetabose)\
**Post date:** [March 27, 2023, 3:36pm UTC](https://discourse.cogstack.org/t/moving-from-umls-to-snomed-ct/177/3 "2023-03-27T15:36:06Z")

</div>

ok… I have done multiple rounds of annotation using the trainer with the UMLS . To start afresh i would need to use Medcat trainer and reannotate with SNOMED code isnt it. Could you please confirm that.

Or is there any way to use my UMLS training jsons towards creating a SNOMED Medcat - for a onetime migration.

---

<div class="post-metadata">

**Author:** ![anthony.shek](https://avatars.discourse-cdn.com/v4/letter/a/b782af/32.png) [@anthony.shek](https://discourse.cogstack.org/u/anthony.shek)\
**Post date:** [March 28, 2023, 10:28am UTC](https://discourse.cogstack.org/t/moving-from-umls-to-snomed-ct/177/4 "2023-03-28T10:28:37Z")

</div>

So there are two possible options: the latter is what I would recommend and what you proposed.

1. UMLS model with only snomed ct concepts

- Filter your existing UMLS model for only the CUIs which have SNOMED CT mapping. (you’ll technically have a subset of UMLS which only includes those with SNOMED mapping). You then always been annotation with UMLS and then remapping the output to snomed.

1. SNOMED model trained on your UMLS annotations

> [@sangeetabose](#):
>
> use my UMLS training jsons towards creating a SNOMED Medcat - for a onetime migration.

- Create a new model from SNOMED
- Perform the unsupervised training steps.
- Then to use all your hard work! (in the form of UMLS annotations on MCT) you will need to rename the UMLS CUIs present in the MCT, with the mapped SNOMED SCTID wherever a mapping is possible (not all UMLS codes will have a mapping).
- Conduct the supervised training step
- Now you will have SNOMED model trained on work that you’ve already done!

To do both of these steps you will certainly need the UMLS-\>SNOMED CT mapping which can be found in the UMLS files provided by NIH. Let me know if you have any questions along the way

---

<div class="post-metadata">

**Author:** ![sangeetabose](https://avatars.discourse-cdn.com/v4/letter/s/eb8c5e/32.png) [@sangeetabose](https://discourse.cogstack.org/u/sangeetabose)\
**Post date:** [March 28, 2023, 10:56am UTC](https://discourse.cogstack.org/t/moving-from-umls-to-snomed-ct/177/5 "2023-03-28T10:56:42Z")

</div>

Understood… thank you so much @tomolopolis and @anthony.shek Appreciate the response