app.py

import torch
import nltk
import re
import numpy as np
import gradio as gr
from typing import Dict
from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler
from transformers import BertForSequenceClassification,BertTokenizer


#Loading Model and Tokenizer from Hugging Face Spaces
model = BertForSequenceClassification.from_pretrained("owaiskha9654/PICO_Evidence_Classification_Model", num_labels=4)
tokenizer = BertTokenizer.from_pretrained('owaiskha9654/PICO_Evidence_Classification_Model', do_lower_case=True) 
nltk.download('punkt')

examples = [ """In addition to their essential catalytic role in protein biosynthesis, 
                aminoacyl tRNA synthetases participate into numerous other functions, 
                including regulation of gene expression and amino acid biosynthesis via 
                transamidation pathways. labor 48 labor induced amniotomy oxytocin.
                Herein, we describe a class of aminoacyl tRNA 
                synthetase like HisZ proteins based on the catalytic core of the contemporary 
                class II histidyl tRNA synthetase whose members lack aminoacylation activity 
                but are instead essential components of the first enzyme in histidine biosynthesis 
                ATP phosphoribosyltransferase (HisG).""",
            """total 164 female breast cancer patients treated anthracyclines received anthracycline-based 
                chemotherapy = 85 = 79 dexrazoxane maximum cycles. participants received amlodipine 5 daily lisinopril 5 daily
                secondary outcomes death noncardiac repeat revascularization procedures pci cabg"""
            """ Prediction of the function of HisZ in Lactococcus 
                lactis was assisted by comparative genomics, a technique that revealed a link between the 
                presence or the absence of HisZ and a systematic variation in the length of the HisG polypeptide.""",
                """HisZ is required for histidine prototrophy, and three other lines of evidence support the direct 
                involvement of HisZ in the transferase function. 1) Genetic experiments demonstrate that 
                complementation of an in-frame deletion of HisG from Escherichia coli (which does not possess HisZ) 
                requires both HisG and HisZ from L. lactis. 2) Coelution of HisG and HisZ during affinity chromatography 
                provides evidence of direct physical interaction. 3) Both HisG and HisZ are required for catalysis of 
                the ATP phosphoribosyltransferase reaction. This observation of a common protein domain 
                linking amino acid biosynthesis and protein synthesis implies an early connection between 
                the biosynthesis of amino acids and proteins."""]


def PICO_Tagger(text):
    evidence_dict = {'Not Relevant to the Evidence (Label N)': 0,
                     'Population/Problem (element P)': 1,
                     'Intervention and Comparison (element I and C)': 2,
                     'Outcome (element O)': 3}

    evidence_dict_inverse = {v: k for k, v in evidence_dict.items()}

    text = re.sub(r"(@\[A-Za-z0-9]+)|([^.0-9A-Za-z \t])|(\w+:\/\/\S+)|^rt|http.+?", "", text)
    sent_text = nltk.sent_tokenize(text)
    inputs = tokenizer(sent_text, return_tensors="pt",padding=True)
    inputs = inputs.to('cpu')
    outputs = model(**inputs)
    pred_logits = outputs.logits
    pred_label = torch.sigmoid(pred_logits)
    out = torch.argmax(pred_label, dim=1).detach().cpu().numpy().tolist()
    output=[]
    for i in out:
        output.append(evidence_dict_inverse[i])

    sent_text_len = [len(i) for i in sent_text]

    rendered_output =[]
    temp =0
    for index,values in enumerate(sent_text_len):
        dict1={}
        dict1['start'] = temp
        temp=temp+values+1
        dict1['end']=temp
        dict1['entity'] = output[index]
        rendered_output.append(dict1)

    return {"text": text, "entities": rendered_output} 
    
       
title = "Utilizing BioBERT for PICO Evidence Summarization"
description = "The traditional machine learning models give a lot of pain when we do not have sufficient labeled data for the specific task or domain we care about to train a reliable model. Transfer learning allows us to deal with these scenarios by leveraging the already existing labeled data of some related task or domain. We try to store this knowledge gained in solving the source task in the source domain and apply it to our problem of interest.In this work, I have utilized the dataset from the paper published in IEEE Journal of Biomedical and Health Informatics <a href=\"https://ieeexplore.ieee.org/document/9056501\">Aceso: PICO-Guided Evidence Summarization on Medical Literature</a>"

text1 = (
    "<center> Author: Owais Ahmad Data Scientist at <b> Thoucentric </b> <a href=\"https://www.linkedin.com/in/owaiskhan9654/\">Visit Profile</a> <br></center>"
    
    "<center> Model Trained Kaggle Kernel <a href=\"https://www.kaggle.com/code/owaiskhan9654/utilizing-biobert-for-pico-evidence-summarization\">Link</a> <br></center>"
        
    "<center> Kaggle Profile <a href=\"https://www.kaggle.com/owaiskhan9654\">Link</a> <br> </center>"
    
    "<center> PICO Evidence Model Deployed Repository <a href=\"https://huggingface.co/owaiskha9654/PICO_Evidence_Classification_Model\">Link</a> <br></center>"
)   

PICO_APP = gr.Interface(PICO_Tagger,
             gr.Textbox(placeholder="Enter Article sentence here..."), 
             gr.HighlightedText(),
             examples=examples,allow_flagging='never',
    analytics_enabled=False,    
    title=title,
	description=description,
    article=text1,)

PICO_APP.launch()