JMIR Publications

ABSTRACT

Background:

Diabetes is a prevalent chronic disease that imposes significant burdens on individuals and healthcare systems worldwide. With the growing volume of electronic medical records (EMRs), there is a pressing need to leverage these real-world datasets to uncover meaningful insights for clinical decision-making and research.

Objective:

This study aims to build a high-throughput knowledge graph from medical records on diabetes via Bi-LSTM-CRF model and the Neo4j database and then perform analytics based on such phenomics graph.

Methods:

The experimental dataset consists of de-identified medical record data of 2002 patients from The First Affiliated Hospital of Zhengzhou University. The Bi-LSTM-CRF model was trained using annotated medical record data. After verifying the accuracy of the model, the entities are automatically extracted based on the trained model, and the relationships between nodes are constructed to create triples, which are then imported into the graph database Neo4j for visualization and topology analysis. In addition, the extracted entities are deduplicated and disambiguated to calculate the Jaccard coefficient matrix of disease and symptom entities and further network analysed.

Results:

The Bi-LSTM-CRF model in this experiment achieved state-of-art performances. Based on which, 45541 entities and 325325 triples were extracted. The recognized entities were saved in the Neo4j for visualization and topological analysis. The networks of co-occurrence morbidities and symptoms analytics were performed and the meaningful scientific discovery was approved.

Conclusions:

The framework can automatically extract entities from real-world data and can be used for further graph construction and network analysis, laying the feasibility for other data-driven phenomics analysis.