Ground the answer.
Retrieve a compact set of relevant clinical relationships instead of repeatedly interpreting long passages. Keep the supporting evidence within reach.
Medical knowledge representation
Clinical knowledge.
Connected.
Reasoning, grounded.
Building a rich, connected knowledge graph across medicine—designed to grow, improve and ground medical AI in traceable knowledge.
Clinical meaning. Traceable connections.
A foundation for medical AI
Give the answer
something to stand on.
LLMs can generate fluent answers while inventing facts or reaching inconsistent conclusions. Clinical reasoning needs an explicit connection to knowledge and its sources.
Our goal is a rich clinical knowledge graph spanning all of medicine, built and improved at scale. Its explicit relationships and supporting evidence can ground and guide LLM reasoning, providing guardrails against hallucinations and non-deterministic variation.
Keep the context.
Carry route, formulation, population and other applicable qualifications with the knowledge. A connection alone is not the whole story.
Make reasoning inspectable.
Show which relationships support an explanation—and where the graph does not supply an answer.
See all advantages
Why structure knowledge when an LLM can already retrieve text? Prose remains the source. The graph makes its relationships reusable, directly queryable and easier to inspect.
Compact knowledge. Focused context.
A relevant subgraph can represent the facts needed for a question in far less text than complete labels or textbook chapters. Retrieve connected facts and applicable qualifiers first, then include the supporting passages as needed—reducing repeated interpretation and the amount of context sent to a model.
Explainability
Follow an explicit path from a clinical concept through a relationship to its source evidence. Clinicians and developers can inspect the knowledge supporting an answer instead of relying on an explanation alone.
Maintainability
Revise affected relationships as evidence changes, without retraining the answering model. Versioned sources make changes traceable and support reproducible retrieval.
Interoperability
Standard terminology annotations help connect knowledge to clinical systems. Our starting dataset uses RxNorm and SNOMED CT, with UMLS concept identifiers linking terminology across vocabularies.
Efficient retrieval
Graph database indexes and traversal find specific relationships and follow connected facts across sources. The same structured knowledge can serve many questions, including on-device applications, without asking an LLM to reconstruct those connections each time.
Multi-morbidity support
Query several documented conditions against treatment and contraindication relationships together. Retain the route, population and other qualifications that determine whether a relationship applies.
Quality and scalable review
Automated checks and targeted expert review help identify extraction errors, questionable mappings and apparent contradictions. Explicit assertions make it possible to inspect and improve individual pieces of knowledge as coverage grows.
Auditability
Source references and versioned assertions provide evidence for review, reproducibility and governance. They give teams an inspectable knowledge layer for evaluating clinical AI systems.
Research applications
Explore treatment relationships, compare evidence, identify knowledge gaps and generate hypotheses. A broader graph could support pharmacovigilance and treatment-pattern research when linked to suitable data.
Explore the possibilities
Clinical applications.
Here are a few ways this knowledge graph could power better healthcare. These two demos illustrate potential applications of the knowledge—not products we are developing.
For the clinician
Multi-Morbidity Safety Check
“The routine renewal”
A familiar medication. Kidney results recorded elsewhere. See how a source-backed contraindication brings the two together.
Explore the safety checkFor the patient
Pre-Visit History Builder
Remember the medicine.
Rediscover the history.
Turn a medication list into useful questions. Let the patient confirm, deny or flag uncertainty before the visit.
Explore the history builderDetails for potential clinical applications
Patient-facing
Pre-Visit History Builder
“Remember your medications, rediscover your conditions”
A patient enters their medications. A graph-backed app looks up conditions those medications may treat, then asks simple confirmation questions: “Do you have high blood pressure?” “Do you have diabetes?” The patient can confirm, deny or say they are not sure.
The app could prepare a visual summary connecting the patient’s confirmed history to their medications, ready to share with their physician. Local graph retrieval and an on-device model could support a privacy-focused implementation.
Graph value: Medication-to-condition links provide a grounded starting point for questions. Patient confirmation turns possibilities into a useful history, helping focus the visit on treatment planning and unanswered questions.
Try the history builder demo→Patient + physician
Visual Treatment Map
“See how your treatments connect to your health”
A patient’s “My Health Map” could connect documented conditions to medications, using color to distinguish curative treatment, chronic management, prevention and symptom relief. Tap a medication to highlight its uses; tap a condition to explore its treatment connections.
A clinician’s summary could place conditions and medications side by side, with evidence attached to each link and questions flagged where an indication needs clarification.
Graph value: Explicit, auditable connections make complex medication lists easier to review and explain. The same graph supports two interfaces for different audiences.
Physician-facing
Multi-Morbidity Safety Check
“Safe prescribing for complex patients”
Before a prescription or renewal, an application could check a patient’s documented conditions and relevant findings against contraindication relationships. It could also support review after an office visit, bringing together facts recorded in different parts of the history.
Our demo follows a metformin renewal with a kidney-function result that meets the source label’s severe-renal-impairment contraindication. The graph connects the medication, the applicable threshold and the supporting evidence.
Graph value: Structured queries can check multiple patient facts against explicit relationships and return the reason for a flag, with its source and qualifications, for clinician review.
Try the safety check demo→Physician-facing
Medication Reconciliation Assistant
“Simplify complex medication lists”
During admission or a care transition, an application could map medications to documented conditions and help organize the questions that need review: Which indication explains each medication? Was a treatment for an acute problem intended to continue? Do overlapping treatment roles warrant a closer look?
It could highlight medications whose indication is not documented and surface relevant evidence alongside the history. Decisions about duplication, continued need or deprescribing would also require patient-specific assessment.
Graph value: Finding what is connected—and where documentation needs clarification—becomes a structured query, giving the clinician a focused starting point for reconciliation.
The connection, with its context
Medication
Ingredient · route · formulation
Clinical relationship
Clinical concept
Condition · finding · population
Source evidence Original assertion · supporting passage · source version
Conceptual illustration of the knowledge represented.
Our approach
Across medicine.
Starting with medications.
The destination is a connected graph for all of medicine. Our starting point is medication–disease and symptom knowledge from DailyMed Structured Product Labels (SPLs).
Connecting medications, clinical concepts and the evidence behind them is the first step. An extensive proprietary automated pipeline uses an LLM to extract relationships, enrich their source context and add terminology annotations.
Terminologies such as UMLS provide a foundation for identifying and linking concepts. We aim to build a richer layer of qualified clinical relationships from medical sources than terminology links alone provide. The sample includes SNOMED CT and RxNorm annotations with UMLS concept identifiers.
The approach is designed to scale to additional sources and areas of medicine, with iterative improvement through automated checks and expert review.
From knowledge to an answer
A graph to guide
the generation.
GraphRAG—graph-based retrieval-augmented generation—can give an LLM relevant relationships and evidence at the time it answers a question.
- 01
Find the relevant knowledge
Use graph database indexes and traversal to retrieve a focused set of connected facts. Compared with retrieving prose alone, the relationships are already explicit; the LLM need not reconstruct them for every question.
- 02
Check applicability
Keep patient context and the assertion’s qualifications in view.
- 03
Explain with evidence
Ground the response in retrieved knowledge and expose the supporting sources.
The graph and source text work together: structured relationships support focused retrieval, while the original passages provide the detail needed to interpret and verify them.
Current work · September 2026
A working demonstration.
A foundation in progress.
This example data package contains SPL extractions with source-scope enrichment and terminology annotations. It demonstrates representation and retrieval—not comprehensive clinical coverage.
- 808
- Source-grounded assertions
in the sample and demo graph - 5
- Selected medication labels
in the example cohort - 2
- Interactive scenarios
for inspection
Built for inspection
A persistent browser graph store, linked source evidence, applicability context and terminology details. The demos query the converted KG locally.
Validation continues
Clinical validation is ongoing. Expert review and improvements to extraction, terminology mapping and coverage guide the next stage of work.
Growing toward broader access
A future scaled version of the knowledge graph will be published, free for non-commercial use. The sample below offers a starting point for exploring the representation.
Example data
A sample of what can be done.
Explore extractions from five DailyMed labels: levothyroxine, metformin, lisinopril, atorvastatin and amoxicillin. The JSON includes clinical assertions, product context, source evidence and RxNorm / SNOMED CT annotations.
See the reasoning behind the connection