Glossary

AI basics

Artificial intelligence (AI)

AI

A broad term for computational systems that perform tasks such as recognising patterns, making predictions, generating content or supporting decisions.Most modern AI — including generative AI — is built using machine learning. It learns patterns from data rather than being explicitly programmed for every situation. In healthcare, the important questions are the task, data, evidence and use of the system — not the “AI” label.

Machine learning

ML

A method in which a model learns patterns from data or examples rather than having every rule explicitly programmed.A model can reflect the populations, coding practices and care pathways represented in its training data, so its performance may differ in another hospital.

Deep learning

Machine learning using multi-layer neural networks that can learn useful representations from data such as images, signals or text.

Training

The process by which a model learns patterns or has its parameters fitted using training data. Training may happen initially and again during later updates; running the trained model on new input is inference.

Inference

Running a trained model on new input to produce an output. This is what happens each time the tool is used in practice.

Generative AI

Fine-tuning

Further training an existing model on additional examples to adapt its behaviour to a specific task or style.Fine-tuning can adapt behaviour and sometimes encoded knowledge, but it does not provide dependable access to current, local or patient-specific facts. Those should come from governed context or retrieval.

Generative AI

GenAI

AI models designed to generate new content, such as text, images, audio or code. The same models can also be used for tasks such as classification, extraction or scoring.Generated output can be fluent yet unsupported or wrong, so review and traceability remain important; non-generative scores also require appropriate evaluation and oversight.

Large language model (LLM)

LLM

A language model trained at large scale to model sequences of tokens, commonly through next-token prediction. This enables uses such as answering, summarising, drafting, extracting and classifying text.An LLM does not by itself provide a reliable source index, freshness guarantee or provenance trail. These have to be designed into the surrounding system.

Foundation model

A large, general-purpose model trained on broad data and then adapted to many downstream tasks rather than built for one.A foundation model is one component. The clinical product is the system around it: context, tools, guardrails, evaluation and workflow.

Prompt

Instructions, context or examples given to a model to guide an output or action — the specification it is asked to follow.

Context engineering

Deciding what information a model is given for each request — instructions, patient context, retrieved passages, tool results — and in what form.

Hallucination

Generated content or claims that are not adequately supported by the supplied context, evidence or reliable facts — for example an invented medication, date, guideline or citation.Grounding, source checking and appropriate human review can reduce or catch hallucinations; simply asking the model to be careful is not a reliable control.

Retrieval-augmented generation (RAG)

RAG

Retrieving material from a source collection and supplying it to a generative model so its output can use — and, where designed, cite — those sources.Retrieval can improve currency, local relevance and citability. It does not guarantee correct retrieval, faithful use of sources or adequate oversight.

AI agent

agent

agentic AI

A system that pursues a goal across multiple steps and itself chooses or sequences at least some of the next actions or tool calls based on intermediate results, within defined boundaries. Tool use alone does not make a system an agent: a fixed, pre-scripted sequence of tool calls is a workflow.When an agentic system can take actions through tools, permissions, approval gates, audit logging and stop rules are part of the surrounding system design.

Model Context Protocol (MCP)

MCP

An open protocol for connecting a model-based application to tools and data sources through a standard interface.MCP is a connector standard inside the tools layer. It is not the agent, and it does not by itself grant, restrict or audit access.

Grounding

Supplying a model with relevant source material — such as retrieved passages, records or guidelines — so its output can use information beyond its learned parameters.Grounding is an input-side design decision, and source quality still matters. Whether the output stays faithful to the supplied material is groundedness — a separate, measurable property.

Groundedness

The degree to which claims in a generated output are supported by the sources or context supplied to the model.Groundedness is an output-side quality you evaluate, for example by checking citations. A system can be grounded by design and still produce ungrounded sentences.

Embeddings

Numeric vector representations of data such as text or images, designed so related items can be compared by proximity and used for semantic search or retrieval.Embeddings power retrieval over guidelines and protocols, but similarity is not clinical equivalence — retrieval results still need source review.

Vector database

A data store designed to index embeddings and retrieve nearby items according to a chosen similarity measure.In a governed RAG system the vector database indexes only the approved source collection; it is infrastructure, not a source of truth itself.

Context window

The maximum amount of information, measured in tokens, that a model can process in its active context — such as instructions, text, images, retrieved material and conversation history.A large window is not a memory strategy: important instructions and patient context can still be crowded out, which is why context engineering matters.

Tokens

Units a model processes — often parts of words, but also whole words, punctuation, code or pieces of other modalities, depending on the model. Context limits and often usage are measured in tokens.

Temperature

A sampling setting exposed by some models and APIs: lower values generally make output more predictable, while higher values permit more variation. Implementations differ in whether and how they expose it.A lower temperature may improve consistency where the setting is supported, but it does not guarantee repeatability or factual correctness.

Structured outputs

Producing a model response in a defined schema, such as specified JSON fields, so a surrounding system can validate and parse its structure.Where schema enforcement is supported, it can guarantee structural conformance — not truth, grounding, completeness or clinical correctness.

Tool calling

function calling

A model emits a structured request to use a defined tool; the surrounding application decides whether and how to execute it and may return the result to the model.Tool calling can give a system access to current, patient-specific or external data and actions. Permissions, validation and execution controls belong to the surrounding application, not the model.

Agent2Agent (A2A)

A2A

An open protocol for letting independent AI agents discover, delegate to and exchange tasks with one another.A2A connects agents across system boundaries; governance of what each agent may do still has to be designed and audited separately.

Agent Skills

Skills

Packaged, reusable instruction sets (often with scripts or templates) that an agent can load to perform a specialised task consistently.Skills encode institutional know-how — a referral format, a discharge checklist — as versioned, reviewable artefacts rather than prompt folklore.

Agent harness

harness

The orchestrating code around a model that manages the loop: prompts, tool calls, retries, limits and stopping conditions.Two products on the same model can behave very differently because their harnesses differ; evaluate the system, not just the model.

Multi-agent system

A design where several specialised agents cooperate — for example a planner, a retriever and a checker — rather than one agent doing everything.More agents means more hand-offs to test, more failure modes and more audit surface; added complexity must earn its place.

Data

Electronic health record (EHR)

EPR

EMR

The digital record of a patient's care. Also called an EPR (electronic patient record) in several European systems.EHR data is primarily recorded during care and administration, not collected as a purpose-built research dataset: timing, coding and completeness reflect how the service works.

FHIR

An interoperability standard defining how health data is structured and exchanged between systems.FHIR defines exchange structures and includes terminology resources and bindings. Implementations also commonly use external code systems such as SNOMED CT and LOINC to represent clinical meaning.

SNOMED CT

SNOMED

Clinical terminology

A clinical terminology that gives granular coded meaning to concepts recorded in the clinical record — findings, disorders, procedures and body structures — as opposed to a classification such as ICD, which groups cases for counting.

Provenance

The record of where information came from: which source, which version, when, and through which process.Two senses are used in the curriculum — data provenance (where a dataset came from) and source provenance (which passage supports this sentence). Both must be inspectable.

Dataset shift

distribution shift

A difference in the distribution of data between development and use — across sites, populations or periods — that can affect a model's behaviour or performance. Drift commonly refers to distribution change over time and is one form of dataset shift.

Data leakage

leakage

target leakage

temporal leakage

When model development or evaluation uses information that would not legitimately be available at prediction time, or allows related records to cross data partitions, producing an unrealistically favourable result.Leakage can inflate measured performance and undermine deployment. Common forms include target or temporal leakage and patient-, site- or preprocessing-related leakage between development and evaluation data.

LOINC

Lab codes

Observation identifiers

A standard set of identifiers for laboratory tests, clinical measurements, panels and documents — it names what was measured or observed, not what the patient has.LOINC codes the observation itself; the result may still need units (UCUM) and reference ranges to be interpretable by a model.

ICD

International Classification of Diseases

ICD-10

ICD 10

ICD10

ICD-11

ICD 11

ICD11

Medical coding

Clinical coding

Diagnosis coding

The WHO International Classification of Diseases: a classification that sorts diagnoses into a bounded set of categories for statistics, registers and reimbursement — distinct from a clinical terminology such as SNOMED CT.ICD-11 came into effect on 1 January 2022, but countries may continue reporting in ICD-10 during transition, so both are met in practice. WHO's ICD-10↔ICD-11 mapping tables support that transition; many relationships are one-to-many or context-dependent, so a mapped series is an interpretation rather than a lossless conversion. ICD codes also reflect coding practice and reimbursement as much as clinical reality; models trained on them inherit those incentives.

UCUM

A standard for encoding units of measure (mg, mmol/L, mmHg) so quantities can be exchanged and compared unambiguously.Unit errors are a classic silent data failure: using a value without a trustworthy unit can make a model input unsafe or misleading.

OMOP Common Data Model

OMOP

A research-oriented common data model that maps data from many sources into one standard structure and vocabulary for analysis.OMOP supports standardised multi-site analyses and external validation, but mapping quality and local provenance still matter. It complements rather than replaces FHIR's exchange role.

SMART on FHIR

A set of OAuth-based patterns for apps to launch inside or outside an EHR and obtain authorised access to FHIR data with defined scopes and context.A common integration route for clinical applications: an app may run with EHR launch context and a defined scope of authorised access.

CDS Hooks

A standard for triggering clinical decision support at defined points in the EHR workflow, such as opening a chart or signing an order.CDS Hooks can call a decision-support service at a defined workflow moment and return cards or permitted actions. It is an integration pattern, not evidence that the advice is clinically sound.

Evidence & evaluation

External validation

Evaluating a model on data from a different site, population or period than the one it was developed on.It gives evidence about transportability beyond development data, but relevance to a particular setting still depends on its population, workflow and time period.

Calibration

Whether predicted probabilities match observed frequencies — if the model says 20%, roughly 20 in 100 such patients should have the outcome.Calibration is important when interpreting predicted probabilities. A decision threshold also depends on discrimination, prevalence and case mix, harms and benefits, workflow and what action follows.

Positive predictive value (PPV)

Of the cases a tool flags, the proportion that meet the reference standard for the condition or outcome. It depends strongly on prevalence as well as the tool's sensitivity and specificity in that setting.

Sensitivity

recall

true positive rate

Of everyone who had the event, the proportion the tool flagged: TP / (TP + FN). It is counted among people with the outcome, so it does not depend directly on how common the outcome is.For a fixed score, lowering the threshold commonly increases sensitivity and the number flagged, with consequences for specificity, alert volume and workload. Sensitivity can also vary across populations, case mix and verification processes.

Specificity

true negative rate

Of everyone who did not have the event, the proportion the tool left unflagged: TN / (TN + FP).Unlike PPV, sensitivity and specificity do not depend mathematically on prevalence alone. They can still change across settings because of case mix, disease spectrum, measurement and verification processes.

AUROC

AUC

ROC AUC

C-statistic

A single ranking measure: how well a model separates people who had the outcome from those who did not, across all thresholds, in the cohort evaluated.AUROC is not a proportion of correct decisions and is not tied to any threshold. What a score means, what a chosen threshold produces and whether anyone can act on it all lie outside it.

Shadow mode

Running an AI system on live cases while withholding its outputs from clinicians so they do not intentionally influence care, often to evaluate performance before activation.Shadow mode can reveal workflow-fit and data problems missed offline while reducing risk from output-driven decisions. Privacy, security, data and operational risks remain, and the results do not by themselves prove clinical benefit.

Silent mode

silent pilot

Running a system on live data while withholding its outputs so they do not intentionally influence care; often used for alert-style systems to study alert rates and timing before activation.Silent-mode results can inform threshold and workload decisions before outputs are shown, while privacy, security, data and operational risks still require controls.

Total cost of ownership (TCO)

TCO

The full cost of a system over its life: licences, integration, infrastructure, validation, training, monitoring, change management and decommissioning.Integration, validation, training and ongoing monitoring can materially add to — or exceed — licence costs, depending on the deployment. A business case based only on licence price is incomplete.

NASSS framework

A framework for assessing complexity in health-technology adoption across the condition, technology, value proposition, adopters, organisation, wider system, and adaptation over time.NASSS helps teams examine and explain risks of non-adoption, abandonment, and difficulty with scale-up, spread and sustainability; it is not a universal predictive rule.

DECIDE-AI

A reporting guideline for early-stage, live clinical evaluation of AI-driven decision-support systems, covering performance, safety and human factors in their clinical context.

TRIPOD+AI

TRIPOD-AI

A reporting guideline for prediction-model studies using regression or machine learning, covering development and validation reporting.Guideline-compliant reporting enables appraisal; it is not itself proof a model is safe or effective.

Safety & oversight

Hazard

A potential source of harm. A hazardous situation is a circumstance in which people, property or the environment are exposed to one or more hazards; harm is the resulting injury or damage.A hazard need not arise from a failure, and it is distinct from both the hazardous situation and any resulting harm. Safety review therefore considers hazards, hazardous situations, near misses and harms.

Harm

Injury or damage to the health of people, or damage to property or the environment. It is the consequence, not the hazard or hazardous situation itself.Keeping hazard and harm separate matters because harm is also unevenly distributed, so subgroup performance and workload belong in the same analysis.

Stop rule

stop criteria

stopping rule

A pre-agreed condition under which use of a system is paused or withdrawn, with named signals, thresholds, an owner and a review rhythm.Without those four elements a stop rule is a sentence in a document rather than a control — and it has to be executable out of hours, not only in working hours.

Human oversight

human-in-the-loop

meaningful oversight

A defined role and process in which people have the competence, information, time and authority to understand, intervene in, override or stop a system as appropriate to its use and risks.Oversight need not mean checking every output before use, but it must be meaningful for the workflow and risk. A review click is not enough if the reviewer lacks context, competence, time or authority to intervene.

Automation bias

The tendency to accept a system's output because it came from the system, and to under-weigh contradicting evidence.

Audit logging

A record of system events and actions, which may include the inputs used, model and version, output produced, review and subsequent action.Appropriately designed, protected and retained logs help teams reconstruct incidents and support traceability without recording unnecessary personal data.

Monitoring

Ongoing review of a live system's inputs, outputs, performance, overrides, workload and safety signals, linked to named owners and actions. Observability is related infrastructure that makes internal system states and events inspectable; it is not the same as monitoring.

Idempotency

A property of an operation where performing it twice has the same effect as performing it once — so retries do not create duplicates.Agent actions that write to clinical or administrative systems should be idempotent or guarded where retries are possible; otherwise a retried request can book or order twice.

Prompt injection

An attack where hostile instructions hidden in retrieved or user-supplied content try to override the system's own instructions.Any content the model reads — letters, web pages, patient messages — can carry injected instructions, so tool permissions must not depend on model judgement alone.

Regulation (EU)

Educational summaries of EU law, not legal advice. Checked against the primary texts between 25 August and 10 September 2026; obligations and dates change, so verify the article that governs your own case: AI Act, MDR, EHDS.

Intended purpose

intended use

What the manufacturer states the software is for: the clinical task, users, population and setting. It is central to qualification, classification and the evidence required.Intended purpose is not one sentence in a regulatory file. It is shaped by every manufacturer-controlled claim — instructions, labelling, website, sales material and procurement answers — and a mismatch between them matters.

Medical device software (MDSW)

MDSW

Software intended, alone or in combination, for a purpose covered by the EU definition of a medical device or an in vitro diagnostic medical device.Qualification follows the stated intended purpose; classification is a separate step. MDR Rule 11 commonly applies to software providing information used for diagnostic or therapeutic decisions, but that function does not define all MDSW.

Software as a medical device (SaMD)

SaMD

IMDRF terminology for software intended for one or more medical purposes that performs those purposes without being part of a hardware medical device.SaMD is an international IMDRF term; EU legislation and guidance commonly use MDSW. In the EU, qualification under the MDR or IVDR depends on the software's intended purpose and function.

MDR / IVDR

Two separate EU laws: Regulation (EU) 2017/745 governs medical devices (MDR), while Regulation (EU) 2017/746 governs in vitro diagnostic medical devices (IVDR). Software is covered when it qualifies under the relevant regulation.

EU AI Act

Regulation (EU) 2024/1689 sets risk-, use- and role-based rules for AI systems, and separate rules for general-purpose AI models, across the EU.Healthcare AI is not automatically high-risk. Classification depends on the system's intended purpose and whether Article 6 with Annex I or Annex III applies; the Act can apply alongside medical-device and data-protection law.

GDPR

The EU General Data Protection Regulation, governing how personal data — including health data — may be processed.

European Health Data Space (EHDS)

EHDS

Regulation (EU) 2025/327 on the European Health Data Space, covering access to and exchange of electronic health data for care, EHR-system rules, and governed secondary use. It is in force, with phased application beginning on 26 March 2027.

DPIA

A Data Protection Impact Assessment: the GDPR assessment required before processing that is likely to result in a high risk to people's rights and freedoms. Use of health data is an important factor, but does not make every AI project identical in scope or risk.

FRIA

A Fundamental Rights Impact Assessment required by Article 27 of the EU AI Act for specified deployers before first use of certain high-risk AI systems. It is not required for every high-risk system or every deployer.A body governed by public law, or a private entity providing a public service, may be in scope when deploying a covered high-risk system. A FRIA complements rather than automatically replacing a DPIA or clinical safety assessment.

Notified body

An independent conformity-assessment body designated by an EU member state to carry out the assessment that MDR/IVDR require for certain device classes before CE marking.Not every device needs one. Simple Class I devices are generally self-declared by the manufacturer, while Class I sterile, measuring or reusable-surgical devices and the higher classes need notified-body involvement to the extent the applicable conformity-assessment route requires it. Where a notified body is involved, its capacity and timelines are a real project constraint.

CE marking

The mark by which a manufacturer indicates that a product conforms to applicable EU requirements. For qualifying medical software, it follows the relevant MDR or IVDR conformity-assessment route, with notified-body involvement where required.CE marking relates to the assessed device and intended purpose. A change must be evaluated to determine whether it affects conformity or requires further assessment.