Datasets:

dylan-ma
/

EMED

annotations_creators:
  - expert-generated
language:
  - en
multilinguality:
  - monolingual
  - translation
size_categories:
  - 10K<n<100K
source_datasets:
  - original
task_categories:
  - text-classification
  - zero-shot-classification
  - feature-extraction
task_ids:
  - natural-language-inference
  - semantic-similarity-scoring
pretty_name: EMED (e-MedSolution to OMOP CDM Schema Matching dataset)
tags:
  - schema-matching
  - data-integration
  - healthcare
  - omop-cdm
  - e-medsolution
  - medical-informatics
  - tabular-schema
configs:
  - config_name: default
    default: true
    data_files:
      - split: train
        path: data/train.csv
      - split: validation
        path: data/validation.csv
      - split: test
        path: data/test.csv
dataset_info:
  features:
    - name: omop
      dtype: string
    - name: table
      dtype: string
    - name: des1
      dtype: string
    - name: des2
      dtype: string
    - name: label
      dtype: int64
    - name: d1
      dtype: string
    - name: d2
      dtype: string
    - name: d3
      dtype: string
    - name: d4
      dtype: string
  splits:
    - name: train
      num_examples: 64957
    - name: validation
      num_examples: 8126
    - name: test
      num_examples: 8121

EMED: Healthcare Schema Matching Dataset

Dataset Summary

EMED is a schema matching dataset created by mapping a clinical dataset from the e-MedSolution healthcare information system domain to the OMOP Common Data Model (CDM).

Schema matching in healthcare requires aligning complex, domain-specific clinical tables and attributes across heterogeneous standards. EMED provides 81,204 candidate attribute pairs across 55 e-MedSolution source tables (804 attributes) and 7 OMOP CDM target tables (101 attributes), with 239 verified positive matches.

Dataset Creation Pipeline

The EMED dataset was created and annotated with the following three-stage pipeline:

Data Preprocessing: The textual description of each table and attribute name originates from the database design document of the e-MedSolution system, which was translated from Hungarian to English.
Matching Pair Generation: The candidate matching pairs were generated based on the Cartesian product of the table and attribute names in the e-MedSolution data model and a core subset of tables and attributes from the OMOP CDM data model (person, visit_occurrence, visit_detail, condition_occurrence, procedure_occurrence, measurement, and specimen).
Data Annotation: Each matching pair was initially annotated manually based on the translated textual descriptions of the table and attribute names in the source and target data models, and then partially reviewed and verified by a Hungarian native speaker for consistency.

Dataset Statistics

Overall Dataset Summary

Dataset	Schema Pair	Domain	Source Tables (e-MedSolution)	Source Attributes	Target Tables (OMOP CDM)	Target Attributes	Total Pairs	Positive Matches (`label=1`)
EMED	EMED-OMOP	Healthcare	55	804	7	101	81,204	239

Split Breakdown

Split	File	Total Pairs	Positive (`label=1`)	Negative (`label=0`)
`train`	`data/train.csv`	64,957	109	64,848
`validation`	`data/validation.csv`	8,126	126	8,000
`test`	`data/test.csv`	8,121	4	8,117
Total	—	81,204	239	80,965

Data Fields

Each record in the dataset represents a candidate matching pair between an OMOP CDM <table>-<attribute> and an e-MedSolution <table>-<attribute>:

Column	Type	Description
`omop`	`string`	Target OMOP CDM table and attribute name formatted as `<omop_table>-<omop_attribute>` (e.g., `person-person_id`).
`table`	`string`	Source e-MedSolution table and attribute name formatted as `<emed_table>-<emed_attribute>` (e.g., `hun_bed-nsid`).
`des1`	`string`	Combined textual description of the target OMOP CDM table and attribute (`<d1>;<d2>`).
`des2`	`string`	Combined textual description of the source e-MedSolution table and attribute (`<d3>;<d4>`).
`label`	`int64`	Binary ground-truth match indicator (`1` = match, `0` = non-match).
`d1`	`string`	Textual description of the target OMOP CDM table.
`d2`	`string`	Textual description of the target OMOP CDM attribute.
`d3`	`string`	Textual description of the source e-MedSolution table.
`d4`	`string`	Textual description of the source e-MedSolution attribute (including foreign-key/context notes where available).

Example Instance

{
  "omop": "person-person_id",
  "table": "hun_bed-nsid",
  "des1": "the person domain contains records that uniquely identify each patient in the source data who is time at-risk to have clinical observations recorded within the source systems.;a unique identifier for each person.",
  "des2": "The hun_bed table records the basic data of patient beds, i.e., bed number, type, and its availability.;The nurses station identifier code that belongs to specific ward (patient room).;A foreign key to the nurses station identifier in hun_nurses station table recording the belonging nurses station of the bed.",
  "label": 0,
  "d1": "the person domain contains records that uniquely identify each patient in the source data who is time at-risk to have clinical observations recorded within the source systems.",
  "d2": "a unique identifier for each person.",
  "d3": "The hun_bed table records the basic data of patient beds, i.e., bed number, type, and its availability.",
  "d4": "The nurses station identifier code that belongs to specific ward (patient room).;A foreign key to the nurses station identifier in hun_nurses station table recording the belonging nurses station of the bed."
}

Usage

You can load the dataset directly using the Hugging Face datasets library:

from datasets import load_dataset

# Load the train / validation / test splits
dataset = load_dataset("dylan-ma/EMED")
print(dataset)

Repository Structure

EMED/
├── .gitattributes          # Git LFS tracking configuration
├── README.md               # Hugging Face Dataset Card & metadata
└── data/
    ├── train.csv           # Training split (64,957 rows)
    ├── validation.csv      # Validation split (8,126 rows)
    └── test.csv            # Test split (8,121 rows)

Citation

If you use the EMED dataset in your research, please cite the following work:

@article{kgrag4sm2026,
  author   = {Ma, Chuangtao and Chakrabarti, Sriom and Khan, Arijit and Moln{\'{a}}r, B{\'{a}}lint },
  title    = {Knowledge Graph-based Retrieval-Augmented Generation for Schema Matching},
  journal  = {CoRR},
  volume   = {abs/2501.08686},
  year     = {2025}
}

@article{ma2020knowledge,
  title={{Knowledge-enriched Schema Mapping: A Preliminary Case Study of e-MedSolution System}},
  author={Ma, Chuangtao},
  journal={E{\"o}tv{\"o}s Lor{\'a}nd University, Tech. Rep.},
  year={2020}
}