Medical Schema Matching using Knowledge Graph Embedding (short paper)

Chaoyu Gao, Tianxing Wu, Shengqi Jing, Yuxiang WangEEKE@JCDL 2022 · OpenReview

Introduction

Medical data integration is a critical hurdle for large-scale clinical research. Currently, patient information is often locked in "data silos" across different institutions, each utilizing unique database schemas. For instance, one hospital might identify a patient record using "Patient_ID," while another uses "PID." Traditional schema matching relies on manual expert rules or standard machine learning, both of which struggle with the specialized terminology and structural complexity of medical records.

To address these challenges, we present KGESM (Knowledge Graph Embedding for Schema Matching), a framework that automates the linking of related data fields across heterogeneous hospital databases by leveraging Knowledge Graphs (KGs) and deep language models.

Methodology

The KGESM framework consists of three primary stages: Direct Mapping, Medical Entity Description Embedding, and Schema Alignment.

1. Direct Mapping

The process begins by transforming structured relational databases into KGs. In this mapping, table names are treated as concepts, columns as properties, and individual rows as instances. This transformation preserves the structural context of the original data, which is often lost in simple text-based matching.

2. Medical Entity Description Embedding (MEDE)

To incorporate domain-specific knowledge, KGESM utilizes a Chinese medical encyclopedia. We employ a BERT encoder to generate vector representations of medical entities and their professional descriptions.

BERT-based Medical Entity Description Embedding (MEDE) architecture.

Unlike traditional models like Word2Vec, BERT’s sub-word tokenization effectively handles rare medical jargon and out-of-vocabulary terms. By encoding professional descriptions, the model gains a deeper semantic understanding of the concepts it is trying to match.

3. Schema Matching via TransE and Alignment

Finally, the KGs are embedded into a continuous vector space using the TransE algorithm. A relation alignment module then applies linear transformations to discover equivalent relations (columns) across the different graphs.

Heterogeneous knowledge graph embedding and relationship alignment process.

This module essentially learns a mapping function that "translates" the coordinates of a data property in one hospital's graph to its corresponding equivalent in another's. An iterative K-Nearest Neighbor (KNN) approach is used during training to refine these matches over time.

Experimental Evaluation

The model was evaluated using real-world data from two hospitals in the Jiangsu Province (Hospital J and Hospital K), collected in 2020. A "ground truth" dataset of 287 equivalent relation pairs was manually curated by medical experts.

KGESM was compared against several baseline methods, including SMGR (regex-based), ISMW (Word2Vec-based), and REMA (graph-based).

Results

KGESM demonstrated superior performance across all key metrics:

  • Precision: 85.9% (outperforming REMA at 83.4%)
  • Recall: 95.2% (significantly higher than REMA's 90.6%)
  • F1-Score: 90.3%

The high recall indicates that KGESM is particularly effective at identifying nearly all correct matches within complex medical datasets, representing a significant improvement over previous state-of-the-art methods.

Conclusion and Future Work

KGESM provides a robust, automated solution for medical schema matching by combining the structural strengths of Knowledge Graphs with the semantic depth of BERT-based embeddings. Future research will focus on expanding the model to handle complex mappings (1-to-n or m-to-n) and improving scalability to support the simultaneous integration of multiple regional or national health databases.

Summary updated

BibTeX
@inproceedings{gao2022medical,
  author = {Chaoyu Gao and Tianxing Wu and Shengqi Jing and Yuxiang Wang},
  title = {{Medical Schema Matching using Knowledge Graph Embedding (short paper)}},
  booktitle = {EEKE@JCDL},
  year = {2022},
  url = {https://openreview.net/forum?id=img4B3kCmM_},
}
Figure