Knowledge Graph Construction, Reasoning, and QA for Intelligent Hypertension Diagnosis

The presentation introduces knowledge‑graph construction, reasoning, and question‑answering techniques applied to intelligent hypertension diagnosis, using the precise medication decision problem as a concrete example.

It highlights the complexity of medical decision tasks that require extensive domain knowledge, explainable inference, and multi‑step reasoning, emphasizing challenges such as heterogeneous knowledge modeling and the need for interpretable outcomes.

Six key topics are covered: (1) medical decision‑making tasks, (2) hierarchical hyper‑relational knowledge modeling, (3) neural‑symbolic multi‑hop reasoning, (4) large‑model‑driven intelligent QA, (5) practical deployment in clinical settings, and (6) a Q&A session.

Evaluations of GPT‑4 and several domain‑specific medical large models show that, while they can identify basic hypertension grades, they fail to provide personalized treatment plans or accurate risk stratification, indicating current limitations of LLMs in this vertical.

The proposed hyper‑relational knowledge‑graph engine correctly infers blood‑pressure grade, risk level, and a combined CCB + β‑blocker medication regimen, while also providing the logical reasoning path that justifies each decision.

Hierarchical hyper‑relational modeling contributions include HAHE (global‑local attention embedding), CDSS (hyper‑relational rule‑based decision support), DHGE (dual‑view embedding), and THH‑KG (instance‑concept‑reasoning three‑layer architecture), with HAHE achieving SOTA link‑prediction results on ACL‑2023 benchmarks.

Neural‑symbolic multi‑hop reasoning frameworks such as FLEX (feature‑logic embedding), NQE (n‑ary query embedding), and TFLEX (temporal extension) are introduced, each attaining state‑of‑the‑art performance on static, hyper‑relational, and temporal reasoning tasks across multiple benchmark datasets.

The ChatKBQA framework combines large‑model generation of logical forms with unsupervised entity‑relation retrieval to produce SPARQL queries, achieving superior accuracy on WebQSP and CWQ datasets compared with traditional KBQA pipelines.

Real‑world deployment spans 18 hospitals and community health centers, serving over 4,000 hypertension patients with doctor‑assist decision support, personalized medication recommendations, and chronic‑disease management via a patient‑facing mini‑program.

Future work focuses on (1) efficient large‑model‑driven construction of hierarchical hyper‑relational KGs, (2) precise and explainable multi‑hop reasoning on incomplete graphs, and (3) integrating LLM agents with domain‑specific KGs to realize AI digital doctors for broader medical decision‑making scenarios.