Applying Knowledge Graphs for Clinical VTE Risk Assessment: A Case Study from HuiMei Technology

The presentation introduces HuiMei Technology, a company of about 200 employees focused on improving clinical diagnosis quality through AI, serving over 200 major Chinese hospitals with products such as clinical assistance systems and VTE prevention tools.

It outlines the business need for VTE risk assessment, highlighting VTE’s high mortality, multi‑department occurrence, and evolving nature, and explains why traditional manual scoring is burdensome for clinicians.

A typical VTE assessment workflow is described, including patient admission, nursing data entry, physician verification, and multiple evaluation time points, emphasizing the lack of clinician time for manual scoring.

The core system automatically extracts information from electronic medical records using NLP, links extracted entities to a proprietary medical knowledge graph, and generates risk scores with real‑time alerts and intervention recommendations.

Technical architecture is presented: hospital data sources (HIS, EMR, LIS) are ingested, cleaned, and structured; entities are recognized and linked to graph nodes via Elasticsearch‑based text search and ranking; scoring rules are applied based on hierarchical graph relationships.

Implementation details cover entity extraction, triple generation (e.g., <foot swelling, location, ankle> ), entity‑to‑scorecard matching, and the use of deep‑learning models for NLP.

The knowledge‑graph construction process is explained, including the use of proprietary data, manual curation by medical editors, and alignment with standards such as SNOMED‑CT, ICD‑10/11, with a simple two‑table schema (descriptions and relationships) supporting flexible updates.

Applications beyond risk scoring are discussed, such as advanced patient search and clinical decision support, illustrating the limitations of keyword‑based SQL queries.

The article concludes with a summary of three key factors for successful graph deployment—business‑driven use cases, flexible graph design with expert validation, and massive high‑quality data—and outlines ongoing work on disease‑specific datasets and research‑grade risk models.