PICO_Parser

Parse RCT PubMed abstracts following PICO framework to standarize PICO elements.

• Author: Tian Kang ([email protected])
• Affiliation: Department of Biomedical Informatics, Columbia Univerisity (Dr. Chunhua Weng's lab)
• Citation: "Kang, T., Zou, S. and Weng, C., 2019. Pretraining to Recognize PICO Elements from Randomized Controlled Trial Literature. Studies in health technology and informatics, 264, p.188."

UPDATE May, 2020:

1. Solved the issues with BERT-based parser.

2. Pretrained Sentence classification model for RCT abstracts is now available.

Major updates coming soon: ^_^

More modules coming soon for representing medical evidence information comprehensively from RCT abstracts.

User Guide

NEW: BlueBERT-based Parser (bugs solved, May 2020):

Adapted from NCBI-NLP BlueBERT

1. Install requirements.txt

2. If you want to use UMLS to standardize entities, please install 'UMLS' and 'QuickUMLS' locally

3. Download pretrained bluebert for PICO element recognition models (link in BERT )

4. Edit parser_config.py to customize your own diretories and BERT configuration

5. Run to start parsing (specify your input in --data_dir and output directory in -- output_dir. In the input directory, each abstract text is put in one text file with its pmid as the file name. Example data is provided in test folder.

    python run_bluebert_ner_predict.py --data_dir= --output_dir=
   

To run examples:

python run_bluebert_ner_predict.py --data_dir=test/txt --output_dir=test/json`  

Exmample

Input test/txt
Parsing results test/json

Original: LSTM Parser:

PICO Element with attributes in JSON/XML

1. Install requirements.txt
2. If you want to use UMLS to standardize entities, please install 'UMLS' and 'QuickUMLS' locally
3. Edit parser_config.py to customize your own diretories and installation
4. Run python Phase1_NER_predict.py to start parsing

Clustering parsed PICO elements to represent study design

1. Download context vector pretrained in all pubmed abstracts from 1990-2019 (downlaod link in cluster/model/download.txt)
2. Extract 3 files and put them under cluster/model
3. TO BE CONTINUED

Exmample

JSON
Input example.txt contain over 70+ abstracts with methods sections
Parsing results folder example_json_out

{
  "pmid": "11264545",
  "sentences": {
    "sent_1": {
      "Section": "METHODS",
      "text": "METHODS AND RESULTS : To determine the relative power of radiographic heart measurements for predicting outcome in dilated cardiomyopathy , we retrospectively studied 88 adult patients with chest radiographs obtained within 35 days of echocardiography .",
      "entities": {
        "entity_1": {
          "text": "radiographic heart measurements",
          "class": "Outcome",
          "negation": 0,
          "UMLS": "C0018787:heart,C1306645:radiograph,",
          "index": 1,
          "start": 10
        },
        "entity_2": {
          "text": "predicting outcome",
          "class": "Outcome",
          "negation": 0,
          "UMLS": "",
          "index": 2,
          "start": 14
        },
        "entity_3": {
          "text": "dilated cardiomyopathy",
          "class": "Participant",
          "nega    tion": 0,
          "UMLS": "C0007193:dilated cardiomyopathy,",
          "index": 3,
          "start": 17
        },
        "entity_4": {
          "text": "chest radiographs",
          "class": "Participant",
          "negation": 0,
          "UMLS": "C1306645:radiographs,C0817096:chest,",
          "index": 4,
          "start": 27
        },
        "entity_5": {
          "text": "echocardiography",
          "c    lass": "Participant",
          "negation": 0,
          "UMLS": "C0013516:echocardiography,",
          "index": 5,
          "start": 34
        }
      },
      "relations": {}
    },
    "sent_2": {
      "Section": "METHODS",
      "text": "Standard radiographic variables were measured for each patient , and the cardiothoracic ( CT ) ratio , frontal cardiac area     , and volume were calculated .",
      "entities": {
        "entity_6": {
          "text": "Standard radiographic variables",
          "class": "Outcome",
          "negation": 0,
          "UMLS": "C0038137:Standard,C1306645:radiograph,",
          "index": 1,
          "start": 0
        },
        "entity_7": {
          "text": "cardiothoracic ( CT ) ratio",
          "class": "Outcome",
          "negation": 0,
          "UMLS": "",
          "index": 2,
          "start": 11
        },
        "entity_8": {
          "text": "frontal cardiac area",
          "class": "Outcome",
          "negation": 0,
          "UMLS": "C0018787:cardiac,",
          "index": 3,
          "start": 17
        },
        "entity_9": {
          "text": "volume",
          "class": "Outcome",
          "negation": 0,
          "UMLS": "",
          "inde    x": 4,
          "start": 22
        }
      },
      "relations": {}
    }
  }
}

XML
Input test.txt
Parsing results temp.xml

A double-blind crossover comparison of pindolol , metoprolol , atenolol and labetalol in mild to moderate hypertension . 1     This study was designed to compare in a double-blind randomized crossover trial , atenolol , labetalol , metoprolol and pindolol . Considerable differences in dose ( atenolol 138 +/- 13 mg daily ; labetalol 308 +/- 34 mg daily ; metoprolol 234 +/- 22 mg daily ; and pindolol 24 +/-2 mg daily were required to produce similar antihypertensive effects . 

<abstract>
		<sent>
			<text>A double-blind crossover comparison of pindolol , metoprolol , atenolol and labetalol in mild to moderate hypertension .</text>
			<entity class='Intervention' UMLS='C0031937:pindolol' index='T1' start='5'> pindolol </entity>
			 <entity class='Intervention' UMLS='C0025859:metoprolol' index='T2' start='7'> metoprolol </entity>
			 <entity class='Intervention' UMLS='C0004147:atenolol' index='T3' start='9'> atenolol </entity>
			 <entity class='Intervention' UMLS='C0022860:labetalol' index='T4' start='11'> labetalol </entity>
			 <entity class='Participant' UMLS='C0020538:hypertension' index='T5' start='13'> mild to moderate hypertension </entity>
		</sent>
		<sent>
			<text>1 This study was designed to compare in a double-blind randomized crossover trial , atenolol , labetalol , metoprolol and pindolol .</text>
			<entity class='Intervention' UMLS='C0004147:atenolol' index='T6' start='14'> atenolol </entity>
			 <entity class='Intervention' UMLS='C0022860:labetalol' index='T7' start='16'> labetalol </entity>
			 <entity class='Intervention' UMLS='C0025859:metoprolol' index='T8' start='18'> metoprolol </entity>
			 <entity class='Intervention' UMLS='C0031937:pindolol' index='T9' start='20'> pindolol </entity>
		</sent>
		<sent>
			<text>Considerable differences in dose ( atenolol 138 +/- 13 mg daily ; labetalol 308 +/- 34 mg daily ; metoprolol 234 +/- 22 mg daily ; and pindolol 24 +/-2 mg daily were required to produce similar antihypertensive effects .</text>
			<attribute class='modifier' index='T10' start='1'> differences </attribute>
			 <entity class='Intervention' UMLS='C0004147:atenolol' index='T11' start='5'> atenolol </entity>
			 <attribute class='measure' index='T12' start='6'> 138 +/- 13 mg daily </attribute>
			 <entity class='Intervention' UMLS='C0022860:labetalol' index='T13' start='12'> labetalol </entity>
			 <attribute class='measure' index='T14' start='13'> 308 +/- 34 mg daily </attribute>
			 <entity class='Intervention' UMLS='C0025859:metoprolol' index='T15' start='19'> metoprolol </entity>
			 <attribute class='measure' index='T16' start='20'> 234 +/- 22 mg daily </attribute>
			 <entity class='Intervention' UMLS='C0031937:pindolol' index='T17' start='27'> pindolol </entity>
			 <attribute class='measure' index='T18' start='28'> 24 +/-2 mg daily </attribute>
			 <entity class='Outcome' UMLS='C0003364:antihypertensive' index='T19' start='37'> antihypertensive effects </entity>
		</sent>
</abstract>   

Reference

• Parser achitecture is adapted from my previous project of eligibility criteria parser EliIE.
• LSTM-CRF scritps modified from EBM-NLP