Glossary
AI × Medicine Glossary
The minimum set of concepts needed to follow artificial intelligence in medicine.
A compact edition. Terms are deduplicated: spelling variants, abbreviations and synonyms are held inside a single entry.
52 · 5 · 2026-08-25
Showing 52 of 52 concepts.
- Artificial IntelligenceAIEstablished
Artificial Intelligence
Computational systems that perform tasks normally requiring human cognition, such as perception, inference and language use.
The umbrella term used in every regulatory text and clinical guideline; its breadth is why narrower terms below matter more in practice.
- Machine LearningMLEstablished
Machine Learning
Methods that derive a model from data rather than from explicitly written rules.
Nearly all clinical AI is machine learning; the distinction from rule engines decides how a system must be validated and monitored.
- Deep LearningDLEstablished
Deep Learning
Machine learning using multi-layer neural networks that learn their own feature representations.
The basis of medical image analysis and of every current language model in healthcare.
- Foundation ModelEstablished
Foundation Model
A large model pre-trained on broad data and adapted to many downstream tasks rather than built for one.
Shifts validation from one model per task to one model reused across tasks, which regulators treat as a distinct risk profile.
- Large Language ModelLLMEstablished
Large Language Model
A foundation model trained to predict text, producing and interpreting natural language.
Behind documentation assistants, summarisation and patient-facing chat; also the source of hallucination risk in clinical settings.
- Medical Large Language ModelEmerging
Medical Large Language Model
A language model adapted to clinical text, terminology and reasoning tasks through domain training or tuning.
Domain adaptation improves terminology handling but does not by itself establish clinical safety; evaluation remains task-specific.
- Multimodal AIEmerging
Multimodal AI
A model that takes in more than one kind of input — for example images, text and signals — within a single system.
Matches how clinical reasoning actually works, combining a report, an image and a measurement rather than one alone.
- Software as a Medical DeviceSaMDRegulatory
Software as a Medical Device
Software intended for a medical purpose that performs that purpose without being part of a hardware medical device.
The single most consequential classification in health software: it determines conformity assessment, clinical evidence and post-market duties.
- Intended PurposeRegulatory
Intended Purpose
The use for which a product is intended according to the data supplied by the manufacturer in labelling, instructions or promotional material.
Classification follows what the manufacturer claims, including on a website — a disclaimer does not undo a stated medical purpose.
- AI-enabled Medical DeviceRegulatory
AI-enabled Medical Device
A medical device whose function depends wholly or partly on a machine-learning model.
Adds obligations that static software does not carry: data governance, change control and monitoring of performance after release.
- Predetermined Change Control PlanPCCPRegulatory
Predetermined Change Control Plan
An agreed description of how a model may be modified after authorisation without a new submission, and how each change will be verified.
The mechanism that lets a learning system be updated lawfully instead of freezing it at the version that was approved.
- Post-Market SurveillancePMSRegulatory
Post-Market Surveillance
Systematic collection and review of a device's real-world performance after it is placed on the market.
For learning systems this is where degradation is caught; a model that passed validation can fail quietly in a new population.
- Clinical Decision Support SystemCDSSEstablished
Clinical Decision Support System
Software that provides clinicians with patient-specific information or recommendations to inform a decision.
Whether a system merely displays reference material or acts on an individual's data is the line that turns it into a regulated device.
- Ambient Clinical IntelligenceACIEmerging
Ambient Clinical Intelligence
Systems that listen to a clinical encounter and produce documentation from it without the clinician typing.
The fastest-adopted category of clinical AI, and the one where consent, recording and record accuracy questions concentrate.
- AI Medical ScribeEmerging
AI Medical Scribe
A tool that drafts a clinical note from the audio of a consultation for the clinician to review and sign.
The draft-and-confirm pattern is what keeps it outside device classification in most readings; removing the review step changes that.
- Medical Imaging AIEstablished
Medical Imaging AI
Models that classify, detect, segment or register findings in radiological and other medical images.
The largest authorised category of AI medical devices, and the origin of most published evidence on clinical AI performance.
- RadiomicsEstablished
Radiomics
Extraction of large numbers of quantitative features from medical images for use as model inputs.
Turns an image into a table, which makes it analysable — and makes reproducibility across scanners the central problem.
- Computational PathologyEstablished
Computational Pathology
Analysis of digitised whole-slide tissue images by computational models.
Slide digitisation is the precondition; without it the discipline has no input, which is why adoption tracks scanner deployment.
- Clinical Natural Language Processingclinical NLPEstablished
Clinical Natural Language Processing
Extraction of structured meaning from free-text clinical documents such as notes, reports and discharge summaries.
Most of a health record is prose; nearly every downstream analysis depends on turning that prose into codes and values first.
- Digital BiomarkerEmerging
Digital Biomarker
A physiological or behavioural measure collected by a digital device and used as an indicator of a health state.
Turns continuous consumer-device data into something a study can use, provided the measure itself has been validated.
- Remote Patient MonitoringRPMEstablished
Remote Patient Monitoring
Collection of health measurements outside a care setting and their review by a care team.
Generates the continuous data stream that most predictive models in chronic care assume exists.
- Precision MedicineEstablished
Precision Medicine
Tailoring prevention and treatment to individual variability in genes, environment and lifestyle.
The clinical goal that most predictive modelling in healthcare is justified by, and the reason genomic data enters model inputs.
- Risk StratificationEstablished
Risk Stratification
Sorting a population into groups by estimated probability of an outcome, to direct attention or resources.
Where algorithmic bias does its most direct harm, because the output allocates care rather than merely describing it.
- Clinical Prediction ModelEstablished
Clinical Prediction Model
A model estimating the probability of a diagnosis being present or an outcome occurring for an individual.
The dominant form of clinical AI; reporting standards exist specifically because most published models were not usable.
- Electronic Health RecordEHREstablished
Electronic Health Record
The longitudinal record of a patient's care held by a healthcare provider organisation.
The primary source of training and inference data for clinical AI, and the primary reason that data is institution-shaped rather than person-shaped.
- Fast Healthcare Interoperability ResourcesFHIREstablished
Fast Healthcare Interoperability Resources
An HL7 standard defining health data as modular resources exchanged over web APIs.
The practical interface between an AI system and real clinical data; naming it is what makes an interoperability claim checkable.
- International Patient SummaryIPSEstablished
International Patient Summary
A minimal, specialty-agnostic set of health data intended to support unplanned cross-border care.
The concrete target for portable patient records, and the profile most export claims should be measured against.
- SNOMED CTEstablished
SNOMED CT
A clinical terminology providing coded concepts and their relationships for recording clinical meaning.
Gives a model something stable to attach a finding to; free-text labels do not survive translation between systems.
- Medical OntologyEstablished
Medical Ontology
A formal structure of concepts and relations in a medical domain, allowing meaning to be reasoned over.
The difference between matching strings and matching meaning; without one, alias handling silently fails.
- Knowledge GraphKGEstablished
Knowledge Graph
A graph of entities and typed relationships used as a queryable representation of a domain.
Used to ground generation in stated facts rather than model memory, and to trace why an output was produced.
- Health Data InteroperabilityEstablished
Health Data Interoperability
The ability of systems to exchange health data and use it with its meaning preserved.
Exchange without shared meaning produces data that arrives but cannot be relied on; the second half is the hard half.
- European Health Data SpaceEHDSRegulatory
European Health Data Space
An EU framework governing access to and exchange of electronic health data for care and for secondary use.
Sets dated obligations on record systems, and defines a labelling regime that reaches wellness applications, not only devices.
- Special Category Health DataRegulatory
Special Category Health Data
Personal data concerning health, whose processing is prohibited unless a specific condition applies.
Determines the lawful basis for every training set, and is why de-identification and consent design precede model design.
- Federated LearningFLEstablished
Federated Learning
Training a shared model across institutions by exchanging model updates instead of moving the data.
The main answer to hospitals that cannot pool records, though updates themselves can leak and need their own protection.
- Privacy-Preserving Machine LearningPPMLEmerging
Privacy-Preserving Machine Learning
Techniques that limit what can be inferred about individuals from a model or its training process.
Distinguishes an architectural guarantee from a policy promise, which is the distinction procurement asks about.
- De-identificationRegulatory
De-identification
Removing or transforming identifiers so that a record can no longer be attributed to a person without additional information.
Pseudonymised data remains personal data in EU law; conflating the two is the most common compliance error in AI projects.
- Synthetic Health DataEmerging
Synthetic Health Data
Artificially generated records that reproduce the statistical structure of real health data without corresponding to real people.
Useful for development and testing; it does not by itself remove privacy risk, since generators can memorise their training set.
- Retrieval-Augmented GenerationRAGEstablished
Retrieval-Augmented Generation
Retrieving source passages at query time and conditioning generation on them rather than on model memory alone.
The standard mitigation for hallucination in medical question answering, and the mechanism that makes a citation checkable.
- GroundingEmerging
Grounding
Tying a generated statement to a specific retrievable source that supports it.
Without it a correct answer and a fabricated one are indistinguishable to the reader, which in clinical use is the whole problem.
- HallucinationEstablished
Hallucination
Output that is fluent and plausible but not supported by the input or by any real source.
In medicine the failure mode is asymmetric: a fabricated dose or citation reads exactly like a correct one.
- Clinical ValidationRegulatory
Clinical Validation
Demonstration that a system's output achieves the intended clinical effect in the intended population and setting.
Distinct from technical accuracy: a model can score well on a held-out set and still change nothing, or harm, in practice.
- Model CalibrationEstablished
Model Calibration
The agreement between predicted probabilities and observed frequencies of the outcome.
A discriminating but miscalibrated model gives the right ranking and the wrong number, which is unsafe when the number drives a decision.
- Model DriftEstablished
Model Drift
Degradation of a deployed model's performance over time as conditions move away from those it was trained on.
The reason a one-time approval is insufficient and monitoring is a regulatory obligation rather than good practice.
- Dataset ShiftEstablished
Dataset Shift
A change between the data distribution a model was trained on and the one it is applied to.
Explains why a model validated in one hospital fails in the next: different scanners, coding habits and case mix.
- Algorithmic BiasEstablished
Algorithmic Bias
Systematic difference in a model's performance or effect across groups, arising from data, design or deployment.
In healthcare the harm compounds: a group under-represented in the data receives less accurate output and then less care.
- FairnessEstablished
Fairness
A stated criterion for how a model's errors or benefits should be distributed across groups.
Fairness criteria conflict mathematically; choosing one is a clinical and ethical decision, not a technical default.
- Explainable AIXAIEstablished
Explainable AI
Methods that make a model's behaviour or a specific output understandable to a person.
An explanation that is plausible but unfaithful to the model can increase misplaced trust rather than reduce it.
- Uncertainty EstimationEmerging
Uncertainty Estimation
Quantifying how much confidence a model's individual prediction warrants.
Enables the safest available behaviour in clinical AI — declining to answer — which a system without it cannot do.
- Human OversightRegulatory
Human Oversight
A requirement that a person can understand, monitor and override an AI system's operation.
A legal obligation for high-risk systems, and the mechanism that keeps a draft-and-confirm workflow lawful.
- Automation BiasEstablished
Automation Bias
The tendency to accept an automated recommendation and to under-weigh contradicting evidence.
The reason human oversight has to be designed for rather than assumed: a review step people rubber-stamp is not oversight.
- TRIPOD+AIEstablished
TRIPOD+AI
A reporting guideline for studies developing or evaluating clinical prediction models, including those using machine learning.
Gives a concrete checklist for judging whether a published model claim can be assessed at all.
- DICOMEstablished
DICOM
The standard for storing and transmitting medical images together with their acquisition metadata.
The metadata is what makes an image analysable and also what makes it identifying; both matter for AI pipelines.