This is a fine-tuned BioBERT model for extracting reported outcomes (i.e. those for which results are presented) from articles reporting clinical trials.
This is the second version of the model; the original model development was reported in:
Anna Koroleva, Sanjay Kamath, Patrick Paroubek. Extracting primary and reported outcomes from articles reporting randomized controlled trials using pre-trained deep language representations. Preprint: https://easychair.org/publications/preprint/qpml
The original work was conducted within the scope of the Assisted authoring for avoiding inadequate claims in scientific reporting PhD project of the Methods for Research on Research (MiRoR, http://miror-ejd.eu/) program.
Model creator: Anna Koroleva
Intended uses & limitations
The model is intended to be used for extracting reported outcomes from texts of clinical trials.
The main limitation is that the model was trained on a fairly small sample of data annotated by a single annotator. Annotating more data or involvig more annotators was not possiblw within the PhD project.
How to use
The model should be used with the BioBERT tokeniser. A sample code for getting model predictions is below:
import numpy as np
from transformers import AutoTokenizer
from transformers import AutoModelForTokenClassification
tokenizer = AutoTokenizer.from_pretrained('dmis-lab/biobert-v1.1')
model = AutoModelForTokenClassification.from_pretrained(r'aakorolyova/reported_outcome_extraction')
text = """Compared with placebo plus chemotherapy, pembrolizumab plus chemotherapy improved overall survival in patients with previously untreated, advanced oesophageal squamous cell carcinoma and PD-L1 CPS of 10 or more, and overall survival and progression-free survival in patients with oesophageal squamous cell carcinoma, PD-L1 CPS of 10 or more, and in all randomised patients regardless of histology, and had a manageable safety profile in the total as-treated population."""
encoded_input = tokenizer(text, padding=True, truncation=True, max_length=2000, return_tensors='pt')
output = model(**encoded_input)['logits']
output = np.argmax(output.detach().numpy(), axis=2)
print(output)
