MIMIC-SPARQL

This repository provides the official mimic-sparql dataset implementation of the following paper: Knowledge Graph-based Question Answering with Electronic Health Records accepted at Machine Learning in Health Care (MLHC) 2021.

Example

NLQ: how many patients were born before the year 2060?

SQL: select count ( distinct patients."subject_id" ) from patients  where patients."dob_year" < "2060"

SPARQL: select ( count ( distinct ?subject_id ) as ?agg ) where { ?subject_id </dob_year> ?dob_year. filter( ?dob_year < 2060 ).

Prerequisites

1. MIMIC-III
   https://mimic.physionet.org/
2. MIMICSQL
   Paper title: Text-to-SQL Generation for Question Answering on Electronic Medical Records.
   Dataset and codes: https://github.com/wangpinggl/TREQS
3. ENV

python 3.6
networkx
rdflib
pandas
numpy
sqlite3  # sqlite is a built-in library from python 2.5. so there is no need to install manually.
requests

​ Set up ENV using pip

pip install networkx rdflib pandas numpy requests

Datasets

1. MIMICSQL*
   MIMICSQL* is extended version of MIMICSQL. The database consists of 9 table of MIMIC-III.
2. MIMIC-SPARQL
   MIMIC-SPARQL is a graph-based counterpart of MIMICSQL*. The knowledge graph of this dataset has 173,096 triples and the max hop is 5.

Guide for creating the MIMICSQL* and MIMIC-SPARQL

0. Prepare MIMIC-III and make mimic.db from MIMICSQL
1. Build mimicsql* database from mimicsql database
2. Build mimic-sparql knowlege graph from mimicsql* database
3. Convert mimicsql SQL query to mimicsql* SQL query
4. Convert mimicsql* SQL query to mimic-sparql SPARQL query

0. Prepare MIMIC-III and mimic.db from MIMICSQL

First, you need to access the MIMIC-III data. This requires certification from https://mimic.physionet.org/ And then, mimic.db is necessary to go to the next step following the https://github.com/wangpinggl/TREQS README.md

1. Build mimicsql* database from mimicsql database

First, you need to save mimic.db under mimicsql/evaluation/mimic_db path. And then, set the current directory in the project root folder, mimic-sparql.

python build_mimicsqlstar_db/build_mimicstar_db_from_mimicsql_db.py

This is to build MIMICSQL* DB and mimicsqlstar.db is made.

2. Build mimic-sparql knowlege graph from mimicsql* database

Set the current directory in the project root folder, mimic-sparql. For building mimic-sparql* from mimicsql*,

python build_mimicsparql_kg/build_complex_kg_from_mimicsqlstar_db.py

For building mimic-sparql from mimicsql,

python build_mimicsparql_kg/build_simple_kg_from_mimicsql_db.py

This is to build MIMIC-SPARQL KG and mimic_sparqlstar_kg.xml and mimic_sparql_kg.xml are made.

3. Convert mimicsql SQL query to mimicsql* SQL query

Set the current directory in the project root folder, mimic-sparql.

python convert_mimicsql2sql_dataset.py --dataset_type natural --execution False
python convert_mimicsql2sql_dataset.py --dataset_type template --execution False

If set execution as True, the execution results of both queries are compared with each other.

4. Convert mimicsql* SQL query to mimic-sparql SPARQL query

Set the current directory as the project root folder, mimic-sparql.

python convert_sql2sparql_dataset.py --dataset_type natural --complex True --execution False
python convert_sql2sparql_dataset.py --dataset_type natural --complex False --execution False

python convert_sql2sparql_dataset.py --dataset_type template --complex True --execution False
python convert_sql2sparql_dataset.py --dataset_type template --complex False --execution False

Complex option is for selecting simplied schema (mimic-sparql from mimicsql) or original schema (mimic-sparql* from mimicsql*)

Citation

@inproceedings{pmlr-v149-park21a,
  title = 	 {Knowledge Graph-based Question Answering with Electronic Health Records},
  author =       {Park, Junwoo and Cho, Youngwoo and Lee, Haneol and Choo, Jaegul and Choi, Edward},
  booktitle = 	 {Proceedings of the 6th Machine Learning for Healthcare Conference (MLHC)},
  pages = 	 {36--53},
  year = 	 {2021},
  volume = 	 {149},
  publisher =    {PMLR}
}

This bibtex will be changed after being published on PMLR