NabzGraph · A Roshan AI product

From ICU sensor streams to an interpretable clinical graph.

NabzGraph turns continuous multi-modal signals (ECG, PPG, arterial pressure, respiration) into a patient-specific knowledge graph of SNOMED concepts and their measured relationships. Every node and edge traces back to the exact signal windows behind it.

0.92
Concept fidelity
Held-out AUROC
120+
Relationships / graph
All signal-derived
< 2s
KG per patient
CPU, post-encoder
Try it

See NabzGraph build a patient graph.

Pick a patient, build the graph, then click any concept node to trace it to the raw signal. Prerendered for the demo; the platform runs against real waveform data.

patientEarly-warning trajectory, pneumonia admission
knowledge graph11 nodes · 23 edges
PersistentEpisodicTransientfilter edges

Signal evidence

Build the graph, then select any concept node to trace it back to the raw signal windows that activated it.

The research contribution

Edge types measured, not imposed.

Every relationship is derived from signal evidence rather than imposed by an ontology author.

TEMPORAL
Temporal
Lagged cross-correlation of concept activation series
A directed relationship when one concept reliably precedes another.
CO_OCCURS
Co-occurs
Jaccard overlap of two concepts’ source-window sets
An undirected relationship when evidence activates in the same windows.
GRANGER
Granger
Granger causality on the underlying signal features
A directed relationship that supplies a measured causal direction.
The dashboard

Three panels, one reasoning chain.

Move from clinical structure to the source signals that substantiate it.

01

Graph view

Patient-specific concepts, persistence tiers, and measured edge types in one view.

02

Signal evidence

Open the exact waveform windows that activated any concept node.

03

Patient timeline

Trace concept density across the ICU stay and jump to a moment in time.

Evaluation · preprint AAAI 2027

Interpretability you can measure.

Published protocols test concept fidelity, explanation faithfulness, and causal replication.

CF
0.92
Concept fidelity
Koh et al. 2020

Held-out AUROC for the concept bottleneck against rule-defined ground truth.

ERASER
0.64
Faithfulness
DeYoung et al. 2020

Removing cited signal evidence reduces node activation, testing whether the explanation is causal.

GR
0.70
Granger replication
Seth 2010

Derived causal edges recur in held-out portions of the ICU stay.

Figures are from the single-seed held-out test split, pending the camera-ready preprint. Cohort: 200 MIMIC-IV-WDB patients (167 with built graphs); V2 targets full MIMIC-IV with external validation.

Breadth

Coverage across modalities, concepts, and edge types.

Modalities
  • ECGLive
  • PPGLive
  • Arterial BPLive
  • RespirationLive
  • SpO₂Roadmap
SNOMED concepts
  • Tachycardia · bradycardia (ECG)Live
  • Tachycardia (PPG)Live
  • Tachypnea · bradypnea (RESP)Live
  • Hypotension · hypertension (ABP)Live
  • ICD-10 concept extensionRoadmap
Edge & deployment
  • Temporal edgesLive
  • Co-occurrence edgesLive
  • Granger edgesLive
  • Retrospective analysisLive
  • Real-time streamingRoadmap
Security & compliance

Clinical AI you can deploy with confidence.

Enterprise safeguards designed for sensitive clinical workflows.

HIPAA-ready
BAA-eligible deployments
Encrypted
TLS 1.3 · AES-256
Auditable
Concepts and evidence logged
Your data stays yours
No training without opt-in
FAQ

Questions we get a lot.

NabzGraph · ready to evaluate

Bring your signals. See the graph behind the patient.

Talk to us about a retrospective cohort, research collaboration, or the next signal modality.