Patient graphs that know who acted.
Typed relations keep a drug’s effect apart from the clinician’s decision, and every edge points to its evidence.
See NabzGraph build a patient graph.
Pick a patient, build the graph, then select a node to see its evidence. Prerendered example.
Evidence
Build the graph, then select any node to see the events behind it.
Every relation knows its type.
So a clinician’s decision and a drug’s effect are tested apart.
ResponsePropofol ↑ → MAP ↓
What a drug or ventilator change does to the patient.
PolicyRising MAP → remifentanil ↑
What the clinician does in response to the patient.
AssociationHypotension ↔ bradycardia
Findings that travel together, with no causal claim.
MechanisticPulse arrival time ↔ systolic pressure
Beat-by-beat couplings, tested per patient.
Three panels, one reasoning chain.
From the graph down to the events behind each edge.
Graph view
Concept, action and beat nodes, joined by typed relations.
Evidence
Open the episodes, actions or beats behind any edge.
Patient timeline
Concept episodes and clinician actions across the case.
Reliability you can measure.
False relations counted after swapping in another patient’s streams.
Research results on retrospective data. Not validated for clinical use.
Coverage across data, concepts, and relations.
- VitalDB surgical casesBuilt
- MIMIC-IV ICU recordsBuilt
- Drug & ventilator actionsBuilt
- Retrospective analysisBuilt
- Real-time streamingRoadmap
- Tachycardia · bradycardiaBuilt
- Hypotension · hypertensionBuilt
- Tachypnea · bradypneaBuilt
- Hypercapnia · hypocapniaBuilt
- Beat-level conceptsBuilt
- ResponseBuilt
- PolicyBuilt
- AssociationBuilt
- MechanisticBuilt
- Evidence on every edgeBuilt