Aller au contenu
Nous contacter

Clinical context for AI

What a model needs to reason about one patient: data bound to patient and stay, clinical time, coded values with units, record status and complete results.

6 min de lecture


Models, in-house or third-party, can be integrated and exchanged. What decides whether an AI system can say anything useful about a specific patient is the material it works from: clinical information bound to the right patient and stay, ordered in clinical time, coded consistently, marked with its status, and complete for the question asked. monaOS provides that material from its FHIR-native clinical record and delivers it to AI through one governed interface. This article explains what each property means in FHIR terms, how the record supplies it, and what happens to an AI system when it is missing.

Five properties of usable context#

An AI consumer of clinical data needs five things to be true of every value it receives. Each maps to concrete elements of the FHIR record.

Property What it means How the record carries it
Patient and stay binding The value belongs to this patient and to this hospital stay. subject references the Patient; encounter references the Encounter of the stay.
Clinical time When the value was true, as opposed to when it was stored. effectiveDateTime or effectivePeriod, separate from the record time of each version.
Coded meaning and units What was measured, and in which unit. LOINC codes for vital signs, SNOMED CT and ICD-10 for coded clinical content, UCUM units on quantities.
Status Whether the value is valid, corrected or invalidated. status: final, amended and corrected are current; entered-in-error is not.
Completeness Whether the result set holds everything that matched. FHIR search paging, and on the monaOS AI server an explicit completeness statement on every search.

Patient and stay binding#

A value reasoned over for the wrong patient produces output that looks coherent and is clinically wrong. Stay binding matters as much: a creatinine from last year's admission is not part of the current picture, although the patient identifier matches. In monaOS, a hospital stay is one Encounter carrying the hospital case identifier and the location history, with linked phases for the OR and the ICU. Admissions, transfers and discharges from the hospital information system keep that stay current, and documentation that has to start before the admission reaches monaOS is linked to the verified identity later.

Device data needs one more step, because a monitor does not know which patient is in front of it. monaOS associates device data with the patient by bed location or by manual assignment, for example for a transport monitor, so that a heart rate from bed 4 becomes an Observation of the patient in bed 4 (see medical device data in clinical workflows).

Clinical time#

Clinical reasoning is about trajectories: whether lactate is rising, whether a blood gas followed or preceded a ventilator change. The FHIR record keeps clinical time (effectiveDateTime) separate from record time (when a version was stored), so a laboratory value that arrives two hours after sampling still sits at the time it was drawn. The distinctions between clinical time, source time and arrival time are covered in clinical time and data arrival. An AI consumer reviewing a past decision also needs to know what was available at that moment, which the version history of the record answers.

Coded meaning and units#

The same analyte arrives from different systems under different local codes and in different units. As a general concept, a number without a trustworthy code and unit is not a measurement: a creatinine of 1.2 means one thing in mg/dL and another in µmol/L, where the factor between the two is roughly 88. A model has no reliable way to notice that a value was read in the wrong unit; it reasons from a number that is off by nearly two orders of magnitude.

monaOS maps local codes and units to standard terminologies in the HL7 Transformation Engine. Vital signs are coded in LOINC with UCUM units, and the expected unit per code is held in configuration, so values reach the record with a coded meaning and a unit the platform knows. The units and terminology article goes deeper into the pitfalls.

Status#

A corrected value is a new version of the same resource; an invalidated one carries the status entered-in-error and stays in the record with its history. Charts, fluid balances and score calculations in monaOS load only current values. An AI consumer has to apply the same rule, otherwise a value a clinician has already struck out re-enters the reasoning.

Completeness#

"No penicillin allergy found" means something only if the search returned everything that matched. FHIR search results are paged, and a client that stops before the last page holds a partial list. The monaOS AI server states on every search whether the result is complete, and explains truncation when a limit applies, so the AI and the clinician reading its output know when an answer rests on partial data.

An example of bound, timed, coded data#

Illustrative: A synthetic laboratory observation with standard FHIR elements. Identifiers and values are demo data; this is not a monaOS profile.

{
  "resourceType": "Observation",
  "id": "demo-crea-0311",
  "meta": { "versionId": "1", "lastUpdated": "2026-03-14T10:41:05+01:00" },
  "status": "final",
  "code": { "coding": [{ "system": "http://loinc.org", "code": "2160-0", "display": "Creatinine [Mass/volume] in Serum or Plasma" }] },
  "subject": { "reference": "Patient/demo-patient-17" },
  "encounter": { "reference": "Encounter/demo-stay-17" },
  "effectiveDateTime": "2026-03-14T08:30:00+01:00",
  "valueQuantity": { "value": 1.9, "unit": "mg/dL", "system": "http://unitsofmeasure.org", "code": "mg/dL" }
}

Everything a model needs to place this value is explicit: whose it is (subject), which stay (encounter), when the blood was drawn (effectiveDateTime), when the record received it (meta.lastUpdated), what was measured (LOINC 2160-0), in which unit (UCUM mg/dL), and that it is a current, valid result (final).

Why missing context produces confident errors#

Each gap in the table above has the same signature: the output stays fluent while the content goes wrong.

  • Without stay binding, a summary mixes a previous admission into the current one.
  • Without clinical time, a trend is computed in order of arrival rather than order of events.
  • Without units, a value is misread by a constant factor, and every downstream statement inherits the error.
  • Without status, invalidated values return.
  • Without a completeness signal, an absence of findings is reported as a negative finding.

Human review, described in human review in assisted workflows, is the safeguard on the output side, and it works when the inputs are sound. Establishing that soundness is a job for the data layer, done once, rather than for each application separately.

One governed interface#

The alternative to a shared context layer is each application parsing raw HL7 feeds and re-deriving patient association, time ordering and units on its own, which multiplies the failure modes above by the number of applications. monaOS uses a unified hospital API instead: applications and AI read the same FHIR record through the same standard interface.

  1. Clinical applications and AI
  2. monaOS AI server and unified hospital API
  3. FHIR-native clinical record
  4. HL7 Transformation Engine
  5. Hospital information systems and devices
Every consumer reads the same associated, timed and coded record through a governed interface.

For AI, the monaOS AI server delivers this context; it needs only the FHIR endpoint and an access key. It adds rules that matter specifically for models:

  • Read-only in code. There is no write path; results reach the record only when a clinician documents them.
  • Query rules. Only allowlisted resource types and search parameters are accepted, high-volume data requires a filter, and chained queries, reverse includes and server operations are refused before any request is sent.
  • Record text is data. Free-text narrative is stripped from every resource, and every result is labelled as record data that is never to be followed as an instruction.
  • Fixed limits per access key, with ceilings compiled into the server.
  • Audit without patient values: each access is logged with parameter names, never values.

The same record serves mona AI, the hospital's own assistant, and the AI tools a hospital already uses. How models, tools and orchestration build on this context is covered in models, tools and orchestration.

Note: monaOS is an assistive tool and does not replace the professional judgement of healthcare providers. It is not a patient monitoring system, has no alarm or diagnostic function, and receives data from medical devices without controlling them. AI outputs built on its context are prepared for a clinician to review.

Des questions auxquelles la documentation ne répond pas ?

Interrogez ceux qui ont conçu la plateforme : quels systèmes se connectent, quel flux de travail est concerné, ce qui doit arriver dans le dossier.