2025 MedCAT v2 models missing MetaCAT; adding from v1 works but saved model cannot be re-loaded

[LeJudith]

Hi!

The pretrained 2025 MedCAT v2 SNOMED models (e.g., SNOMED 2025 trained on MIMIC-IV) do not include a MetaCAT negation component, whereas the older SNOMED-Int v1 models do.

To add negation detection to the 2025 model, I converted the v1 SNOMED-Int model into a v2 model using the migration tutorial and transferred its MetaCAT addon into the 2025 v2 model:

cat_snomed_int_v2 = CAT.load_model_pack("./mc_modelpack_snomed_int_16_mar_2022_25be3857ba34bdd5v2_85293d2448666153")
mc = cat_snomed_int_v2.get_addons()[0]

cat_2025.add_addon(mc)

This approach works, but it requires loading two models every time. To avoid that, I tried saving the 2025 model after adding the MetaCAT component. However, loading the updated model back in fails:

File "/home/user/venvs/prism/lib/python3.10/site-packages/medcat/cat.py", line 816, in CAT.load_model_pack
    cat = deserialise(
        model_pack_path,
        model_load_path=model_pack_path,
        ignore_folders_prefix={
            AddonComponent.NAME_PREFIX,
            AbstractCoreComponent.NAME_PREFIX,
            TOKENIZER_PREFIX,
            COMPONENTS_FOLDER,
            "."
        },
        config_dict=config_dict,
        addon_config_dict=addon_config_dict,
    )

File "/home/user/venvs/prism/lib/python3.10/site-packages/medcat/utils/serialisation.py", line 79, in deserialise
    with open(file_path) as f:
        contents = f.read()

FileNotFoundError: [Errno 2] No such file or directory:
    "/data/.../v2_Snomed2025_MIMIC_IV_meta_status/.serialised_by"

Questions:

• Is there an issue with adding the MetaCAT component taken from the older (v1 → v2 converted) model to a newer model that I’m not aware of?
• Related to this, is the error expected or a potential a bug in the v2 model-pack save/load process for addons?

Thanks!

[mart-r]

From a quick glance what you've described should work. By converting the v1 model (with MetaCATs) to v2, you've now got v2 versions of these MetaCATs (they do get converted alongside the entire model). So you should then be able to take them out and add them to another model.

I'll need to look at this in more detail. I'll try and get back to you tomorrow.

PS/EDIT:
Thanks for using MedCAT v2 and reporting issue! Only be people using it and pointing out issue can we polish this fully!

[mart-r]

Hi @LeJudith

I've tried to replicate this issue by using the following script:
https://gist.github.com/mart-r/11addb8e8f1e7f2f67044f36063b7a73
However, for me this does work as expected:

% python replicate.py models/mc_modelpack_snomed_int_16_mar_2022_25be3857ba34bdd5v2_4fd4e2d2d0e3cfd9.zip models/v2_Snomed2025_MIMIC_IV_bbe806e192df009f.zip models
Loading model with meta cats: models/mc_modelpack_snomed_int_16_mar_2022_25be3857ba34bdd5v2_4fd4e2d2d0e3cfd9.zip
Got 1 of which 1 are MetaCATs
Loading model WITHOUT meta cats: models/v2_Snomed2025_MIMIC_IV_bbe806e192df009f.zip
Loading a model with incorrectly saved tokenizer internals path. Was saved as '.temp/v2_Snomed2025_MIMIC_IV/v1_Snomed2025_MIMIC_IVv2_bbe806e192df009f/tokenizer_internals_en_core_web_md' whereas should have had just 'tokenizer_internals_en_core_web_md'. This is an automated fix - no further action is needed
Got 0 of which 0 are MetaCATs
Adding MetaCAT to new other one
Adding {
  "Category Name": "Status",
  "Description": "Detects is a concept Affirmed or Negated/Hypothetical",
  "Classes": {
    "Affirmed": 0,
    "Other": 1
  },
  "Model": "lstm"
}
Saving to models
DONE initial!
Loading again from models/medcat2_model_pack_23228014de5d7d12
Got 1 of which 1 are MetaCATs
Trying out new model pack with meta cat
Training was enabled during inference. It was automatically disabled.
{'entities': {1: {'acc': 1,
                  'context_center': [],
                  'context_left': [],
                  'context_right': [],
                  'context_similarity': 1,
                  'cui': '25609006',
                  'detected_name': 'patient~has',
                  'end': 11,
                  'icd10': [],
                  'id': 1,
                  'meta_anns': {'Status': {'confidence': 0.9980130195617676,
                                           'name': 'Status',
                                           'value': 'Affirmed'}},
                  'opcs4': [],
                  'pretty_name': 'Has patient',
                  'source_value': 'Patient has',
                  'start': 0,
                  'type_ids': ['43039974']},
              2: {'acc': 0.99,
                  'context_center': [],
                  'context_left': [],
                  'context_right': [],
                  'context_similarity': 0.99,
                  'cui': '50177009',
                  'detected_name': 'fever',
                  'end': 19,
                  'icd10': ['R509'],
                  'id': 2,
                  'meta_anns': {'Status': {'confidence': 0.9999513626098633,
                                           'name': 'Status',
                                           'value': 'Affirmed'}},
                  'opcs4': [],
                  'pretty_name': 'Body temperature above reference range',
                  'source_value': 'fever',
                  'start': 14,
                  'type_ids': ['67667581']},
              3: {'acc': 0.9092692183006501,
                  'context_center': [],
                  'context_left': [],
                  'context_right': [],
                  'context_similarity': 0.9092692183006501,
                  'cui': '249497008',
                  'detected_name': 'vomit',
                  'end': 36,
                  'icd10': ['#NC'],
                  'id': 3,
                  'meta_anns': {'Status': {'confidence': 0.9999985694885254,
                                           'name': 'Status',
                                           'value': 'Other'}},
                  'opcs4': [],
                  'pretty_name': 'Vomiting symptom',
                  'source_value': 'vomitting',
                  'start': 27,
                  'type_ids': ['67667581']}},
 'tokens': []}

Does using the script I used work on your side? (obviously provide your own paths when calling it)

Could you descibe your environment in some detail, please? I.e the pythono version you're running, the OS you're using, the medcat version you're using.
Could you also be a bit more specific regarding what it is exactly you're doing? I.e if there's a diferent workflow that you're using to get to the model that's failing to load, what is it, exactly?