import streamlit as st import time from annotated_text import annotated_text from io import StringIO from transformers import AutoTokenizer, AutoModelForTokenClassification import os os.environ['KMP_DUPLICATE_LIB_OK']='True' import plotly.express as px from streamlit_option_menu import option_menu st. set_page_config(layout="wide") from transformers import pipeline import pandas as pd @st.cache(allow_output_mutation = True) def init_text_summarization_model(): MODEL = 'facebook/bart-large-cnn' pipe = pipeline("summarization", model=MODEL) return pipe @st.cache(allow_output_mutation = True) def init_zsl_topic_classification(): MODEL = 'facebook/bart-large-mnli' pipe = pipeline("zero-shot-classification", model=MODEL) template = "This text is about {}." return pipe, template @st.cache(allow_output_mutation = True) def init_zsl_topic_classification(): MODEL = 'facebook/bart-large-mnli' pipe = pipeline("zero-shot-classification", model=MODEL) template = "This text is about {}." return pipe, template @st.cache(allow_output_mutation = True) def init_ner_pipeline(): tokenizer = AutoTokenizer.from_pretrained("d4data/biomedical-ner-all") model = AutoModelForTokenClassification.from_pretrained("d4data/biomedical-ner-all") pipe = pipeline("ner", model=model, tokenizer=tokenizer, aggregation_strategy="simple") # pass device=0 if using gpu return pipe @st.cache(allow_output_mutation = True) def init_qa_pipeline(): question_answerer_pipe = pipeline("question-answering", model='deepset/roberta-base-squad2') return question_answerer_pipe def get_formatted_text_for_annotation(output): colour_map = {'Coreference': '#29D93B', 'Severity':'#FCF3CF', 'Sex': '#E9F7EF', 'Sign_symptom': '#EAF2F8', 'Detailed_description': '#078E8B', 'Date': '#F5EEF8', 'History': '#FDEDEC', 'Medication': '#F4F6F6', 'Therapeutic_procedure': '#A3E4D7', 'Age': '#85C1E9', 'Subject': '#D7BDE2', 'Biological_structure': '#AF7AC5', 'Activity': '#B2BABB', 'Lab_value': '#E6B0AA', 'Family_history': '#2471A3', 'Diagnostic_procedure': '#CCD1D1', 'Other_event': '#239B56', 'Occupation': '#B3B6B7'} annotated_texts = [] next_index = 0 for entity in output: if entity['start'] == next_index: # print("found entity") extracted_text = text[entity['start']:entity['end']] # print("annotated",annotated_text) annotated_texts.append((extracted_text ,entity['entity_group'],colour_map[entity['entity_group']])) else: unannotated_text = text[next_index:entity['start']-1] annotated_texts.append(unannotated_text) extracted_text = text[entity['start']:entity['end']] annotated_texts.append((extracted_text ,entity['entity_group'],colour_map[entity['entity_group']])) next_index =entity['end'] +1 if next_index < len(text): annotated_texts.append(text[next_index-1:len(text)-1]) return tuple(annotated_texts) # Model initialization pipeline_summarization = init_text_summarization_model() pipeline_zsl, template = init_zsl_topic_classification() pipeline_ner =init_ner_pipeline() pipeline_qa = init_qa_pipeline() st.header("Intelligent Document Automation") def get_paragraphs_for_summaries(): paras =[] paras.append("""Stephen is a 53 year old gentleman who does general duties police work. He is married and has an 18 year old daughter at home. He is right hand dominant. Cigarettes nil, alcohol rare, allergies nil. DVTe nil. Past medical history hyperlipidemia and reflux testicular cancer in 2000 and right knee reconstruction in 1987. Medications include Nexium and Crestor.""") paras.append("""History presenting complaint: Right knee and right elbow injuries. On 12 January 2020, while at work he was trying to apprehend a stolen vehicle. The deployed some road spikes onto the road. The stolen vehicle went over the road spikes. He was attempting to retrieve the spikes to prevent damage to civilian and police vehicles and while he was doing that, he was hit by a police vehicle coming from behind. The police vehicle was approximately doing 50 km/hr. The headlight of the police car stuck him on the lateral aspect of the right knee. He jumped in the air and flicked in the air with his right elbow also hitting the police car before he flung over a barrier. He was able to mobilise afterwards when the adrenaline was taking effect. After he cooled down that night he developed significant pain in both the elbow and the knee. The elbow seemed to have settled with time but he has got slight discomfort on the lateral epicondyle of the right elbow but otherwise no instability, clicking, locking or catching of the elbow.""") paras.append("""His knee has medial sided pain. It is an annoying type of pain that is present at rear and with activity. It feels like it is getting slightly better but his main problem is that he cannot still fully flex his knee as he used to do before the accident. He has noticed some clicking ad a little bit of swelling in the knee. He has had no instability but he is still a little bit stiff when he first gets up. He has been treated with 4 sessions of physiotherapy. He has had two operations on this knee before when it was reconstructed which was an open procedure and an anthroscopy later to shave off the cartilage. He never got back to playing rugby after his reconstruction but he was able to do martial arts and all his policing duties without any problem.""") paras.append("""He is mildly overweight, normal gait and no effusion in the knee and a range of motion of 0 to 110 degrees which cannot be improved passively. His right elbow has full range of motion and is stable. He has had an Xray on right elbow showing no fracture and an xray of knee showed early medial compartment osteoarthritis. He has got new clicking in his knee and has been advised to get an MRI scan.""") paras.append("""As far as his right elbow is concerned, he should reduce weight lifting activity and any repetitive right upper limb activity that causes any discomfort. His right knee requires an MRI scan to look for any loose bodies and I will see him with the results of the scan. In the meantime, he can continue with his general policing duties which he has anyways been doing since the accident""") return paras def get_paragraphs_for_entities(): paras =[] paras.append("""18 February2020 Dr Christine Fowler""") paras.append("""Dear Christine, Re: Stephen Parrot""") paras.append("""Stephen is a 53 year old gentleman who does general duties police work. He is married and has an 18 year old daughter at home. He is right hand dominant. Cigarettes nil, alcohol rare, allergies nil. DVTe nil. Past medical history hyperlipidemia and reflux testicular cancer in 2000 and right knee reconstruction in 1987. Medications include Nexium and Crestor.""") paras.append("""History presenting complaint: Right knee and right elbow injuries. On 12 January 2020, while at work he was trying to apprehend a stolen vehicle. The deployed some road spikes onto the road. The stolen vehicle went over the road spikes. He was attempting to retrieve the spikes to prevent damage to civilian and police vehicles and while he was doing that, he was hit by a police vehicle coming from behind. The police vehicle was approximately doing 50 km/hr. The headlight of the police car stuck him on the lateral aspect of the right knee. He jumped in the air and flicked in the air with his right elbow also hitting the police car before he flung over a barrier. He was able to mobilise afterwards when the adrenaline was taking effect. After he cooled down that night he developed significant pain in both the elbow and the knee. The elbow seemed to have settled with time but he has got slight discomfort on the lateral epicondyle of the right elbow but otherwise no instability, clicking, locking or catching of the elbow.""") paras.append("""His knee has medial sided pain. It is an annoying type of pain that is present at rear and with activity. It feels like it is getting slightly better but his main problem is that he cannot still fully flex his knee as he used to do before the accident. He has noticed some clicking ad a little bit of swelling in the knee. He has had no instability but he is still a little bit stiff when he first gets up. He has been treated with 4 sessions of physiotherapy. He has had two operations on this knee before when it was reconstructed which was an open procedure and an anthroscopy later to shave off the cartilage. He never got back to playing rugby after his reconstruction but he was able to do martial arts and all his policing duties without any problem.""") paras.append("""He is mildly overweight, normal gait and no effusion in the knee and a range of motion of 0 to 110 degrees which cannot be improved passively. His right elbow has full range of motion and is stable. He has had an Xray on right elbow showing no fracture and an xray of knee showed early medial compartment osteoarthritis. He has got new clicking in his knee and has been advised to get an MRI scan.""") paras.append("""As far as his right elbow is concerned, he should reduce weight lifting activity and any repetitive right upper limb activity that causes any discomfort. His right knee requires an MRI scan to look for any loose bodies and I will see him with the results of the scan. In the meantime, he can continue with his general policing duties which he has anyways been doing since the accident""") paras.append("""Kind regards, Dr Jason Sanders""") return paras def get_text_from_ocr_engine(): return """18 February2020 Dr Christine Fowler Dear Christine, Re: Stephen Parrot Stephen is a 53 year old gentleman who does general duties police work. He is married and has an 18 year old daughter at home. He is right hand dominant. Cigarettes nil, alcohol rare, allergies nil. DVTe nil. Past medical history hyperlipidemia and reflux testicular cancer in 2000 and right knee reconstruction in 1987. Medications include Nexium and Crestor. History presenting complaint: Right knee and right elbow injuries. On 12 January 2020, while at work he was trying to apprehend a stolen vehicle. The deployed some road spikes onto the road. The stolen vehicle went over the road spikes. He was attempting to retrieve the spikes to prevent damage to civilian and police vehicles and while he was doing that, he was hit by a police vehicle coming from behind. The police vehicle was approximately doing 50 km/hr. The headlight of the police car stuck him on the lateral aspect of the right knee. He jumped in the air and flicked in the air with his right elbow also hitting the police car before he flung over a barrier. He was able to mobilise afterwards when the adrenaline was taking effect. After he cooled down that night he developed significant pain in both the elbow and the knee. The elbow seemed to have settled with time but he has got slight discomfort on the lateral epicondyle of the right elbow but otherwise no instability, clicking, locking or catching of the elbow. His knee has medial sided pain. It is an annoying type of pain that is present at rear and with activity. It feels like it is getting slightly better but his main problem is that he cannot still fully flex his knee as he used to do before the accident. He has noticed some clicking ad a little bit of swelling in the knee. He has had no instability but he is still a little bit stiff when he first gets up. He has been treated with 4 sessions of physiotherapy. He has had two operations on this knee before when it was reconstructed which was an open procedure and an anthroscopy later to shave off the cartilage. He never got back to playing rugby after his reconstruction but he was able to do martial arts and all his policing duties without any problem. He is mildly overweight, normal gait and no effusion in the knee and a range of motion of 0 to 110 degrees which cannot be improved passively. His right elbow has full range of motion and is stable. He has had an Xray on right elbow showing no fracture and an xray of knee showed early medial compartment osteoarthritis. He has got new clicking in his knee and has been advised to get an MRI scan. As far as his right elbow is concerned, he should reduce weight lifting activity and any repetitive right upper limb activity that causes any discomfort. His right knee requires an MRI scan to look for any loose bodies and I will see him with the results of the scan. In the meantime, he can continue with his general policing duties which he has anyways been doing since the accident. Kind regards, Dr Jason Sanders""" with st.sidebar: selected_menu = option_menu("Select Option", ["Upload Document", "Extract Text", "Summarize Document", "Extract Entities","Detected Barriers","Get Answers"], menu_icon="cast", default_index=0) if selected_menu == "Upload Document": uploaded_file = st.file_uploader("Choose a file") if uploaded_file is not None: ocr_text = get_text_from_ocr_engine() st.write("Upload Successful") elif selected_menu == "Extract Text": with st.spinner("Extracting Text..."): time.sleep(6) st.write(get_text_from_ocr_engine()) elif selected_menu == "Summarize Document": paragraphs= get_paragraphs_for_summaries() with st.spinner("Finding Topics..."): tags_found = ["Injury Details", "Past Medical Conditions", "Injury Management Plan", "GP Correspondence"] time.sleep(5) st.write("This document is about:") st.markdown(";".join(["#" + tag + " " for tag in tags_found]) + "**") st.markdown("""---""") with st.spinner("Summarizing Document..."): for text in paragraphs: summary_text = pipeline_summarization(text, max_length=130, min_length=30, do_sample=False) # Show output st.write(summary_text[0]['summary_text']) st.markdown("""---""") elif selected_menu == "Extract Entities": paragraphs= get_paragraphs_for_entities() with st.spinner("Extracting Entities..."): for text in paragraphs: output = pipeline_ner (text) entities_text =get_formatted_text_for_annotation(output) annotated_text(*entities_text) st.markdown("""---""") elif selected_menu == "Detected Barriers": #st.subheader('Barriers Detected') barriers_to_detect = {"Chronic Pain":"Is the patint experiencing chronic pain?", "Mental Health Issues":"Does he have any mental issues?", "Prior History":"What is prior medical history?", "Smoking":"Does he smoke?", "Drinking":"Does he drink?", "Comorbidities":"Does he have any comorbidities?"} with st.spinner("Detecting Barriers..."): for barrier,question_text in barriers_to_detect.items(): context = get_text_from_ocr_engine() if question_text: result = pipeline_qa(question=question_text, context=context) st.subheader(barrier) #st.text(result) if result['score'] < 0.3: st.text("Not Found") else: st.text(result['answer']) elif selected_menu == "Get Answers": st.subheader('Question') question_text = st.text_input("Type your question") context = get_text_from_ocr_engine() if question_text: with st.spinner("Finding Answer(s)..."): result = pipeline_qa(question=question_text, context=context) st.subheader('Answer') st.text(result['answer'])