MKG Series · Part 01 of 03

Foundations —
Entities, Relationships & Metadata.

A medical knowledge graph becomes valuable through three layers working together: the entities that represent medical concepts, the relationships that connect them, and the metadata that makes every connection trustworthy.

7Core Entity Domains
6Relationship Types
5Metadata Fields per Node/Edge

Seven Domains, One Connected Map

Each domain carries its own attributes — codes, thresholds, design parameters — but the value comes from how they connect to each other.

01
Domain 01
Disease
e.g. Non-Small Cell Lung Cancer, Psoriasis, Crohn's Disease. Attributes: ICD codes, synonyms, stage, severity.
02
Domain 02
Drug
e.g. Nivolumab, Apixaban. Attributes: MoA, dosage, label indication, adverse events.
03
Domain 03
Biomarkers
e.g. PD-L1, EGFR, KRAS. Attributes: threshold, assay type, predictive/prognostic role.
04
Domain 04
Clinical Studies
e.g. CheckMate studies, Phase III trials. Attributes: design, endpoints, population, outcomes.
05
Domain 05
Guidelines
e.g. NCCN, ESMO guidelines. Attributes: recommendation strength, publication date.
06
Domain 06
Publications
e.g. journal articles, congress abstracts. Attributes: DOI, authors, publication date.

A seventh domain — HCP Concepts (line of therapy, progression, adverse event management) — captures the clinical reasoning context the other six domains feed into.

The Graph Becomes Valuable Through Edges

Entities alone are a dictionary. Relationships are what turn that dictionary into a reasoning system.

Drug Treats Disease
Nivolumab → treats → Melanoma
+
Drug Targets Biomarker
Nivolumab → acts_on → PD-1 pathway
Edge Type
Trial Evaluates Drug
CheckMate 067 → evaluates → Nivolumab
Edge Type
Guideline Recommends Therapy
NCCN → recommends → Nivolumab
Edge Type
Publication Reports Outcome
Publication → reports → OS benefit
Edge Type
Adverse Event Associated With Drug
Drug → associated_with → Pneumonitis

Trust Is Not Implicit. It's a Field.

Every node and relationship should include source, confidence score, date, version, and approval status — without this layer, a graph can be connected and still be unreliable.

Guideline Recommendation — Confidence: High · Source: NCCN 2025 · Version: 4.0

This is the layer that separates a graph an MLR team can stand behind from a graph that just looks impressive in a demo.

How Do You Actually Build This?

Explore Part 02 — the nine-phase methodology, from defining the business use case to validating and connecting the graph to AI.

See the Methodology → Business Value