availability of models?

[bgriffen]

I have access to UMLS via NIH but when I click the v1 link in the README I get an access denied. Is there any other way to obtain the models? I'd be happy to try and get v2 working but I can't proceed much beyond .load() at the moment :|

I see on the docs they are "coming soon" though I notice the hive of activity on the repo - is the release anytime soon? If not, how might I get the v1 models?

I just had a few synthetic EHR records I wanted to convert to snowmed or UMLS codes.

thank you and amazing work

[mart-r]

Make sure to use the API-key based link. The other URL is still not working because we've not heard back from NIH regarding changing the auth calkback URL. See CogStack/MedCAT#545 for details.

The URL to download (v1) models is currently:
https://medcat.sites.er.kcl.ac.uk/auth-callback-api

[bgriffen]

Amazing, thank you, downloading now. For the case of having som EHR strings and wanting snowcat or UMLS codes returned, I noticed most of the tutorials are around custom vocabs and fine tuning / training etc. -- are there any tutorials to just use the base models and send it EHR text and return predicted concepts/codes? Forgive me if I've missed a docs page, there's a few repos and split docs between the versions so thought I would best ask directly. Thank you again.

As a side note, I also saw on the web version to predict text, the relative probabilities were all very small, e.g. 0.02 etc. I wasn't sure if for the examples, e.g. Crohn's disease on the left sidebar were meant to be much higher/predictive?
Edit: Apologies, I'm getting wires crossed with foresight demo (https://foresight.sites.er.kcl.ac.uk/)

[mart-r]

For the case of having som EHR strings and wanting snowcat or UMLS codes returned, I noticed most of the tutorials are around custom vocabs and fine tuning / training etc.

You can just use CAT.get_entities:
https://medcat.readthedocs.io/en/latest/autoapi/medcat/cat/index.html#medcat.cat.CAT.get_entities
However, you need to mind the format of the output. And you may want to use filtering as well (refer to config.linking.filters).

As a side note, I also saw on the web version to predict text, the predicted likelihoods were all very small, e.g. 0.005 etc. I wasn't sure if for the examples, e.g. Crohn's disease on the left sidebar were meant to be much higher/predictive?

Not quite sure what "web version" you're referring to. As such, it's hard to give relevant feedback.
With that said, the confidence score refers to the cosine similarity between the context within which the concept was found and that within which it is expected (the trained/known context). MedCAT generally just takes the most similar of the concepts and returns that (if the similarity is within the threshold in the config and the concept is allowed within the filters). The confidence score is just provided as an indication of the cosine similarity. A cosine similarity of 0.005 doesn't necessarily mean there's other concepts that match better.
The other thing to point out is that the similarity / confidence score comes into play when there's disambiguation to be done. That is, if the name/entity being recognised refers to multiple different concepts, the vector context model is used to determine which of the possible concepts is most likely. However, this only works for concepts that have been trained. Because there's no "known context" to compare against for a concept the model hasn't been trained on.

PS:
There's a lot more nuance on the usage of MedCAT. And the steps you need to take are tightly tied to the exact task you're trying to accomplish. I do recommend looking through at least some of the tutorials. It may feel like they're not that useful for your use case, but they may still include details / parts that are relevant. Not to mention that for a specialised use case, you're likely to need to perform supervised training (fine tuning) anyway.

[bgriffen]

Hmm ok thank you. I basically just have some mock text about patients that I wanted to extract the snowcat and umls concept ids for and then just explore on my own the accuracy of those. Basically exactly as you have running in the browser at https://medcat.sites.er.kcl.ac.uk/. Is basically the browser version in pythonic/jupyter form in the tutorials? Thanks, I'll take a look around.

[bgriffen]

Looks like this is roughly the same thing:

from medcat.cat import CAT

# Download the model_pack from the models section in the github repo.
cat = CAT.load_model_pack('<path to downloaded zip file>')

# Test it
text = "My simple document with kidney failure"
entities = cat.get_entities(text)
print(entities)

# To run unsupervised training over documents
data_iterator = <your iterator>
cat.train(data_iterator)
#Once done, save the whole model_pack 
cat.create_model_pack(<save path>)

Thanks! https://medcat.readthedocs.io/en/latest/main.html#tutorials I'll close it out for now. Appreciate the help.

ps - I'd be curious as to what example of an iterator I should use there as I have 1,000 texts I could feed in somehow.

[bgriffen]

For medcatv2 model:

v2 Snomed 2025 MIMIC trained (This is a model based on the 2025 July release enriched with names from UMLS 2024AA. It's been trained on MIMIC-IV notes.)

are icd10, UMLS codes included? I thought from the description they would be? When I run the example below (used in medcatv1 docs which has icd10) I don't seem to get them this time around?

text = "The patient was diagnosed with leukemia."
entities = cat.get_entities(text)
import json
print(json.dumps(entities, indent=2))

Yields...

{
  "entities": {
    "0": {
      "pretty_name": "Patient",
      "cui": "116154003",
      "type_ids": [
        "31601201"
      ],
      "source_value": "patient",
      "detected_name": "patient",
      "acc": 1,
      "context_similarity": 1,
      "start": 4,
      "end": 11,
      "id": 0,
      "meta_anns": {},
      "context_left": [],
      "context_center": [],
      "context_right": []
    },
    "1": {
      "pretty_name": "Diagnosis",
      "cui": "439401001",
      "type_ids": [
        "2680757"
      ],
      "source_value": "diagnosed",
      "detected_name": "diagnosed",
      "acc": 1,
      "context_similarity": 1,
      "start": 16,
      "end": 25,
      "id": 1,
      "meta_anns": {},
      "context_left": [],
      "context_center": [],
      "context_right": []
    },
    "2": {
      "pretty_name": "Leukemia",
      "cui": "1162768007",
      "type_ids": [
        "33782986"
      ],
      "source_value": "leukemia",
      "detected_name": "leukemia",
      "acc": 0.99,
      "context_similarity": 0.99,
      "start": 31,
      "end": 39,
      "id": 2,
      "meta_anns": {},
      "context_left": [],
      "context_center": [],
      "context_right": []
    }
  },
  "tokens": []
}

[mart-r]

MedCAT v2 didn't originally ship with capabilties to map to other ontologies. We added the functionality again in #147 and #160. And these will be relased with v2.2.0. It's currently in recelease candidate stage. So it doesn't get installed automatically from PyPI. But you should be able to use if if you specifically ask for it (e.g pip install medcat==2.2.0rc1).