NabzGraph · A Roshan AI productFrom 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.
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.
Signal evidence
Build the graph, then select any concept node to trace it back to the raw signal windows that activated it.
Edge types measured, not imposed.
The critique that kills most clinical knowledge-graph papers is “why these types?” NabzGraph answers it by deriving every edge from a measurable property of the signal evidence. Nothing is hand-designed by an ontology author.
Temporal
A directed edge when one concept reliably precedes another. The lag and correlation are read straight off the activation timelines: the trajectory, not just the snapshot.
Co-occurs
An undirected edge when two concepts fire in the same windows. The Jaccard score quantifies how tightly their evidence overlaps in time.
Granger
A directed edge suggesting one signal drives another. It gives a causal direction a threshold alarm cannot, the kind of structure that survives peer review.
Three panels, one reasoning chain.
Graph view
A force-directed graph of concept nodes, colour-coded by persistence tier, with labelled temporal, co-occurrence, and Granger edges. Hover any edge for its statistical evidence.
Signal evidence
Select a node and the raw 30s windows that activated it open on the right: the waveform strip, the window in context, and the concept probability. Scrub every activating window chronologically.
Patient timeline
Concept activation density across the full ICU stay as a per-concept heatmap. Persistence tiers read as colour bands; jump the graph to any point in time.
Interpretability you can measure.
NabzGraph is evaluated against published protocols for concept fidelity, explanation faithfulness, and causal replication, not just downstream accuracy.
Concept fidelity
Mean held-out AUROC of the concept bottleneck against rule-defined ground truth, across the seven concepts. High fidelity means the node labels are earned by the evidence, not guessed.
Faithfulness
Comprehensiveness of the cited signal windows: removing the evidence behind a node drops its activation by 0.64 on average, with near-zero sufficiency loss. The explanation is the cause, not a post-hoc rationalisation.
Granger replication
Held-out replication of the derived causal edges. Granger edges fit on the first half of each stay recur in the second half 70% of the time (151 of 217), evidence the structure is stable, not an artefact of a single window.
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.
Coverage across modalities, concepts, and edge types.
- ECGLive
- PPGLive
- Arterial BPLive
- RespirationLive
- SpO₂Roadmap
- Tachycardia · bradycardia (ECG)Live
- Tachycardia (PPG)Live
- Tachypnea · bradypnea (RESP)Live
- Hypotension · hypertension (ABP)Live
- ICD-10 concept extensionRoadmap
- Temporal edgesLive
- Co-occurrence edgesLive
- Granger edgesLive
- Retrospective analysisLive
- Real-time streamingRoadmap
The graph as a JSON API.
A FastAPI service exposes every node, edge, and piece of evidence. Evidence endpoints return signal-window metadata plus base64 waveform arrays for direct rendering.
Developer overviewPOST /patients/{id}/kg Build / refresh the KG from latest signals
GET /patients/{id}/kg Retrieve the current KG as JSON
GET /patients/{id}/kg/nodes Concept nodes with tiers + scores
GET /patients/{id}/kg/edges Edges with evidence metrics
GET /patients/{id}/evidence/{node} Signal windows behind a node
GET /patients/{id}/timeline Concept activation density over time{
"nodes": [
{
"id": "hypotension::PERSISTENT",
"concept": "hypotension",
"snomed": "45007003",
"modality": "ABP",
"tier": "PERSISTENT",
"activation": 0.88,
"windows": ["08:00", "09:30"]
}
],
"edges": [
{
"type": "TEMPORAL",
"source": "tachycardia::EPISODIC",
"target": "hypotension::PERSISTENT",
"lag_min": 18,
"cross_corr": 0.71
}
]
}Clinical AI you can deploy without holding your breath.
HIPAA-ready
BAA-eligible deployments. PHI never leaves the customer perimeter without explicit consent.
Encryption everywhere
TLS 1.3 in transit. AES-256 at rest. Customer-managed keys available on enterprise plans.
Auditable by design
Every prediction logs its activated concepts and evidence. No opaque inference.
No training on customer data
Customer notes are never used to train base models. Opt-in only, contract-bound.
Questions we get a lot.
Bring your signals. See the graph behind the patient.
The dashboard runs against MIMIC-IV-WDB today. Talk to us about a retrospective cohort or a research collaboration and we'll fit the path.