DrBERT-CASM2
Model description
DrBERT-CASM2 is a French Named Entity Recognition model that was fine-tuned from DrBERT: A PreTrained model in French for biomedical and clinical domains. It has been trained to detect the following type of entities: problem, treatment and test using the medkit Trainer.
Fine-tuned using medkit GitHub Repo
Developed by @camila-ud, medkit, HeKA Research team
Dataset source
Annotated version from @aneuraz called 'corpusCasM2: A corpus of annotated clinical texts'
The annotation was performed collaborativelly by the students of masters students from Université Paris Cité.
The corpus contains documents from CAS:
Natalia Grabar, Vincent Claveau, and Clément Dalloux. 2018. CAS: French Corpus with Clinical Cases. In Proceedings of the Ninth International Workshop on Health Text Mining and Information Analysis, pages 122–128, Brussels, Belgium. Association for Computational Linguistics.
Intended uses & limitations
Limitations and bias
This model was trained for development and test phases. This model is limited by its training dataset, and it should be used with caution. The results are not guaranteed, and the model should be used only in data exploration stages. The model may be able to detect entities in the early stages of the analysis of medical documents in French.
The maximum token size was reduced to 128 tokens to minimize training time.
How to use
Install medkit
First of all, please install medkit with the following command:
pip install 'medkit-lib[optional]'
Please check the documentation for more info and examples.
Using the model
from medkit.core.text import TextDocument
from medkit.text.ner.hf_entity_matcher import HFEntityMatcher
matcher = HFEntityMatcher(model="medkit/DrBERT-CASM2")
test_doc = TextDocument("Elle souffre d'asthme mais n'a pas besoin d'Allegra")
detected_entities = matcher.run([test_doc.raw_segment])
# show information
msg = "|".join(f"'{entity.label}':{entity.text}" for entity in detected_entities)
print(f"Text: '{test_doc.text}'\n{msg}")
Text: "Elle souffre d'asthme mais n'a pas besoin d'Allegra"
'problem':asthme|'treatment':Allegra
Training data
This model was fine-tuned on CASM2, an internal corpus with clinical cases (in french) annotated by master students. The corpus contains more than 5000 medkit documents (~ phrases) with entities to detect.
Number of documents (~ phrases) by split
Split | # medkit docs |
---|---|
Train | 5824 |
Validation | 1457 |
Test | 1821 |
Number of examples per entity type
Split | treatment | test | problem |
---|---|---|---|
Train | 3258 | 3990 | 6808 |
Validation | 842 | 1007 | 1745 |
Test | 994 | 1289 | 2113 |
Training procedure
This model was fine-tuned using the medkit trainer on CPU, it takes about 3h.
Model perfomances
Model performances computes on CASM2 test dataset (using medkit seqeval evaluator)
Entity | precision | recall | f1 |
---|---|---|---|
treatment | 0.7492 | 0.7666 | 0.7578 |
test | 0.7449 | 0.8240 | 0.7824 |
problem | 0.6884 | 0.7304 | 0.7088 |
Overall | 0.7188 | 0.7660 | 0.7416 |
How to evaluate using medkit
from medkit.text.metrics.ner import SeqEvalEvaluator
# load the matcher and get predicted entities by document
matcher = HFEntityMatcher(model="medkit/DrBERT-CASM2")
predicted_entities = [matcher.run([doc.raw_segment]) for doc in test_documents]
evaluator = SeqEvalEvaluator(tagging_scheme="iob2")
evaluator.compute(test_documents,predicted_entities=predicted_entities)
You can use the tokenizer from HF to evaluate by tokens instead of characters
from transformers import AutoTokenizer
tokenizer_drbert = AutoTokenizer.from_pretrained("medkit/DrBERT-CASM2", use_fast=True)
evaluator = SeqEvalEvaluator(tokenizer=tokenizer_drbert,tagging_scheme="iob2")
evaluator.compute(test_documents,predicted_entities=predicted_entities)
Citation
@online{medkit-lib,
author={HeKA Research Team},
title={medkit, A Python library for a learning health system},
url={https://pypi.org/project/medkit-lib/},
urldate = {2023-07-24},
}
HeKA Research Team, “medkit, a Python library for a learning health system.” https://pypi.org/project/medkit-lib/ (accessed Jul. 24, 2023).
- Downloads last month
- 209