The model is an NER LLM algorithm that can classify each word in a text into different clinical categories.

The base pretrained model is GatorTron-base which was trained on billions of words in various clinical texts (https://huggingface.co/UFNLP/gatortron-base). Then using the MedMentions dataset (https://arxiv.org/pdf/1902.09476v1.pdf), I fine-tuned the model for NER task in which the model can classify each word in a text into different clinical categories. The category system is a simplified version of UMLS concept system and consists of 21 categories: "['Living Beings', 'Virus']", "['Living Beings', 'Bacterium']", "['Anatomy', 'Anatomical Structure']", "['Anatomy', 'Body System']", "['Anatomy', 'Body Substance']", "['Disorders', 'Finding']", "['Disorders', 'Injury or Poisoning']", "['Phenomena', 'Biologic Function']", "['Procedures', 'Health Care Activity']", "['Procedures', 'Research Activity']", "['Devices', 'Medical Device']", "['Concepts & Ideas', 'Spatial Concept']", "['Occupations', 'Biomedical Occupation or Discipline']", "['Organizations', 'Organization']", "['Living Beings', 'Professional or Occupational Group']", "['Living Beings', 'Population Group']", "['Chemicals & Drugs', 'Chemical']", "['Objects', 'Food']", "['Concepts & Ideas', 'Intellectual Product']", "['Physiology', 'Clinical Attribute']", "['Living Beings', 'Eukaryote']", 'None'

The github code associated with the model can be found here: https://github.com/longluu/LLM-NER-clinical-text.

The MedMentions dataset contain 4,392 abstracts released in PubMed®1 between January 2016 and January 2017. The abstracts were manually annotated for biomedical concepts. Details are provided in https://arxiv.org/pdf/1902.09476v1.pdf and data is in https://github.com/chanzuckerberg/MedMentions.

The hyperparameters are --batch_size 4 --num_train_epochs 5 --learning_rate 5e-5 --weight_decay 0.01

The model was trained and validated on train and validation sets. Then it was tested on a separate test set. Note that some concepts in the test set were not available in the train and validatin sets.

Here we use several metrics for classification tasks including macro-average F1, precision, recall and Matthew correlation.

{'f1': 0.6271402249699903, 'precision': 0.6691625224055963, 'recall': 0.6085333637974402, 'matthews_correlation': 0.720898121696139}

Feel free to reach out to me at thelong20.4@gmail.com if you have any question or suggestion.