Who it is for

What the platform does

HealthOS ICU works at the bedside: it assesses patient state and issues equipment-control commands. Behind it a knowledge cycle — the domain is formalised, stabilised into an approved graph snapshot, compiled into deterministic modules, embedded into equipment and refined by application materials.

• A formal domain description for state assessment and command synthesis;
• Deterministic supervisor compute modules for discrete and continuous patient-state assessment;
• Supporting materials — outputs from use that improve the system.

Confirmed. Rospatent, patents granted: a method of machine-readable representation of medical knowledge (26 May 2026) · a method of semantically correct ECG interpretation (3 August 2026).

Bedside loop

In the OR and ICU this becomes a loop over immutable components and the compiled module: data from monitors, pumps and equipment passes through adapters, the module forms a recommendation, and an approved command reaches execution through the safety kernel. Every recommendation, check, decision and command is logged.

Knowledge graph

The knowledge graph is a formal, machine-readable domain description: rules, triggers, parameters, evidence, device capabilities, constraints and governance roles. Compilation uses a stable snapshot; operational materials feed the triage path that prepares a refined one.

Object type	Quantity
Clinical rule	142
Condition / trigger	89
Evidence reference	310
Device capability profile	24
Monitor parameter	58
SHACL constraint	19
Governance role and policy	6
Bundle version and signature	12
MedBench criteria (acceptance flags)	26 groups, 202 flags
Triplestore composition	10 dialect graphs, 11 general-purpose graphs; ~73.5k triples
Test ECGs with pathology labels	9 nosology groups; 3,620 labelled cases of 4,133

Where each number comes from — on the sources page →

Place of trusted AI in the platform

Trusted AI works not at the bedside but in the knowledge-preparation path: agents analyse sources and draft the graph. The draft passes review, SHACL validation and approval, then becomes a stable snapshot.

Stage	What the agent layer does	What the platform does next
Knowledge prep	Collects sources and drafts the knowledge graph	Moves the draft into review
Review	Not used	Assists review, SHACL validation and approval
Runtime	Not used	Compiles and packages the module from the approved snapshot
Improvement	Reviews application materials, drafts the next graph	Decides what goes into the next snapshot

The agent layer does not replace the knowledge author or approving expert, does not change the snapshot and does not issue equipment commands. Its usefulness and trustworthiness are measured on the “Why Not AI” page →

Boundaries of responsibility

The safety kernel guarantees isolation of rule semantics from executable code:

• The application service — handles transport, data decoding, secrets and version checks;
• The embedded library — executes compiled rules;
• The knowledge graph — is refined through approved snapshots with discrete knowledge changes.

Key facts

Fact	Value
Interoperability	HL7 v2/MLLP · IHE PCD · ISO/IEEE 11073 SDC · FHIR R4 · DICOMweb
Medical-equipment adapter levels	Level 1: HL7 v2/MLLP, read-only · 2: IHE PCD, structured data · 3: SDC, read and bidirectional observation · 4: SDC, governed parameter write
Command validation phases	Level 1: shadow only · 2: internal review · 3: recommendations only · 4: supervised write · 5: restricted write · 6: full control

Technical composition — platforms, compilers, core languages — in the full documentation →