Medical AI glossary

Short, plain-language definitions of the terms used across the course. Each term has its own link, so lessons can point straight to it.

AI agent: A system in which a language model decides which steps to take and which tools to call (for example, searching records or running a calculation) to complete a task, rather than producing a single answer.See also: Tool calling, Human-in-the-loop
Ambient AI scribe: A tool that records a clinical consultation, with consent, transcribes it and drafts a clinical note for the clinician to review, edit and sign.See also: Hallucination, Human-in-the-loop
AUROC (area under the ROC curve): The probability that a model gives a randomly chosen patient with the outcome a higher score than a randomly chosen patient without it. 0.5 is chance and 1.0 is perfect ranking. Also called the C-statistic. It measures ranking, not whether predicted probabilities are correct.See also: Calibration, Sensitivity (recall), Specificity
Benchmark contamination: When test questions, or close versions of them, were included in a model's training data, so its benchmark score partly reflects memorisation rather than ability.See also: LLM-as-a-judge, Large language model (LLM)
Brier score: The mean squared difference between predicted probabilities and actual outcomes (coded 0 or 1). Lower is better. It reflects both discrimination and calibration.See also: Calibration, AUROC (area under the ROC curve)
C-index (concordance index): For survival models, the probability that of two patients, the one who has the event first was given the higher predicted risk. It is the survival equivalent of AUROC.See also: AUROC (area under the ROC curve), Censoring
Calibration: How closely a model's predicted probabilities match observed outcome rates. In a well-calibrated model, about 20% of patients given a 20% risk go on to have the outcome.See also: AUROC (area under the ROC curve), Brier score, Dataset shift
Censoring: In time-to-event data, when a patient's follow-up ends before the event has happened, so the event time is known only to be later than the last follow-up.See also: Hazard ratio, Competing risks
Class imbalance: When one outcome is much rarer than the other, as with most clinical events. For risk prediction it usually needs no correction; oversampling or reweighting distorts predicted probabilities.See also: Calibration
Competing risks: Events that prevent the outcome of interest from ever occurring, such as death before kidney failure. Standard Kaplan-Meier estimates overestimate risk when competing events are common.See also: Censoring
Confounding: A distortion of the association between an exposure (such as a treatment) and an outcome, caused by a third factor that affects both. Confounding by indication occurs when treatments are given to patients who are sicker, making the treatment look harmful.See also: Target trial emulation
Confusion matrix: A table counting true positives, false positives, false negatives and true negatives at a chosen threshold. Sensitivity, specificity and predictive values are all calculated from it.See also: Sensitivity (recall), Specificity, Positive predictive value (PPV)
Context window: The maximum amount of text, measured in tokens, that a language model can consider at once, including the prompt, any supplied documents and its own response.See also: Token, Large language model (LLM)
Cross-validation: Estimating a model's performance by repeatedly training it on part of the data and testing it on the rest, for example in five 'folds'. It estimates performance on similar data, not in new settings.See also: Data leakage, External validation
Data leakage: When information that would not be available at the moment of prediction gets into the training data, for example a test result recorded after the diagnosis. It makes a model look far better in development than it will be in practice.See also: External validation
Dataset shift: A difference between the data a model was trained on and the data it meets in use, such as a new patient population, a new lab analyser or a change in clinical practice. It is a common reason models lose accuracy after deployment.See also: External validation, Calibration
De-identification: Removing or transforming information that could identify a person (names, dates, record numbers, addresses and so on) so data can be used for research or development with lower privacy risk. It reduces, but rarely removes, re-identification risk.See also: PHI (protected health information)
Decision-curve analysis: A method that estimates the net benefit of using a model to make decisions across a range of risk thresholds, compared with treating everyone or no one. It asks whether a model would actually improve decisions.See also: Calibration
DICOM: Digital Imaging and Communications in Medicine: the international standard for storing and exchanging medical images, such as X-ray, CT and MRI, together with metadata about the patient, study and equipment.See also: De-identification
Differential privacy: A mathematical guarantee that the output of an analysis or model changes very little whether or not any one person's data is included, usually achieved by adding calibrated random noise.See also: De-identification, Federated learning
EHR (electronic health record): The digital record of a patient's care held by a health system, including diagnoses, medicines, test results, vital signs and clinical notes. EHR data are usually stored as time-stamped events.See also: FHIR (Fast Healthcare Interoperability Resources), OMOP Common Data Model
Embedding: A list of numbers that represents a piece of text, an image or another item so that similar items have similar numbers. Embeddings power semantic search and retrieval in RAG systems.See also: Vector store, RAG (retrieval-augmented generation)
External validation: Testing a model on data from a different setting, time period or population than the one it was developed in. It is the main evidence that a model will work outside the place it was built.See also: Dataset shift, Data leakage
Federated learning: Training a model across several institutions without pooling their data: each site trains locally and shares only model updates, which are combined centrally.See also: Differential privacy
Few-shot prompting: Including one or more worked examples of input and ideal output in a prompt, so the model can follow the format and judgement shown.See also: Large language model (LLM), Prompt injection
FHIR (Fast Healthcare Interoperability Resources): An HL7 standard for exchanging healthcare data through web APIs. Data is organised into resources such as Patient, Observation, Condition and MedicationRequest.See also: HL7, OMOP Common Data Model
Fine-tuning: Further training a pre-trained model on a smaller, task-specific dataset so it performs better on that task. Parameter-efficient methods such as LoRA update only a small part of the model.See also: Foundation model, LoRA (low-rank adaptation)
Foundation model: A large model trained on broad data that can be adapted to many tasks, for example a large language model or an imaging model pre-trained on millions of scans.See also: Large language model (LLM), Fine-tuning
Gradient boosting: A machine-learning method that builds many small decision trees in sequence, each fitted to the errors of the trees before it. It captures non-linear effects and interactions in tabular data.See also: Logistic regression, SHAP
Grounding: Tying a model's output to specific, verifiable sources, so each claim can be traced to a document the user can check.See also: RAG (retrieval-augmented generation), Hallucination
Hallucination: When a generative model produces content that is fluent and confident but false or unsupported by its sources, such as an invented reference or a wrong dose. A central safety risk in clinical use.See also: Grounding, RAG (retrieval-augmented generation)
Hazard ratio: The factor by which a characteristic multiplies the rate of an event at any time during follow-up. In a Cox model it is assumed constant over time.See also: Censoring, C-index (concordance index)
HL7: Health Level Seven, a standards organisation and family of standards (including HL7 v2 messaging and FHIR) for exchanging clinical and administrative data between systems.See also: FHIR (Fast Healthcare Interoperability Resources)
Human-in-the-loop: A design in which a qualified person reviews, approves or can override an AI system's output before it affects a patient or record.See also: AI agent
ICD (International Classification of Diseases): The World Health Organization's classification of diseases and health problems, used for mortality and morbidity statistics and, in many countries, billing. ICD-11 came into effect on 1 January 2022; many systems still use ICD-10 or national versions.See also: SNOMED CT
k-anonymity: A property of a dataset in which every combination of quasi-identifiers (such as age band, sex and area) is shared by at least k people, so no one can be singled out by those details alone.See also: De-identification, Pseudonymisation
k-means clustering: An unsupervised method that splits data into k groups by repeatedly assigning each item to the nearest group centre and moving the centres. It always returns k clusters, whether or not real groups exist.See also: Unsupervised learning
Large language model (LLM): A neural network trained on large amounts of text to predict the next token, which lets it generate and transform language. Examples include general-purpose chat models and medically adapted models.See also: Foundation model, Hallucination, Token
LLM-as-a-judge: Using a language model to grade other models' answers, for example against a rubric. It scales evaluation but must be checked against human expert ratings.See also: Benchmark contamination
Logistic regression: A statistical model that estimates the probability of a binary outcome from a weighted sum of predictors on the log-odds scale. It is the standard baseline for clinical risk prediction.See also: Odds ratio, Calibration
LOINC: Logical Observation Identifiers Names and Codes: a free, universal set of codes for laboratory tests and clinical observations, maintained by the Regenstrief Institute. For example, 4548-4 identifies HbA1c.See also: UCUM, FHIR (Fast Healthcare Interoperability Resources)
LoRA (low-rank adaptation): A parameter-efficient fine-tuning method that trains small add-on matrices instead of all of a model's weights, making fine-tuning cheaper and faster.See also: Fine-tuning
MCP (Model Context Protocol): An open protocol that lets AI applications connect to external tools and data sources through a standard interface of 'servers' that expose tools and resources.See also: Tool calling, AI agent
Missing data mechanisms (MCAR, MAR, MNAR): Rubin's classification of why data are missing: completely at random, at random given recorded data, or not at random. In clinical data, values are often missing for reasons linked to the patient's condition.See also: Data leakage
Odds ratio: The ratio of the odds of an outcome in one group to the odds in another. In logistic regression, the exponent of a coefficient is the odds ratio per unit of that predictor, holding the others constant.See also: Logistic regression, Confounding
OMOP Common Data Model: A standard database structure and vocabulary set, maintained by the OHDSI community, that lets observational health data from different sources be analysed in the same way.See also: FHIR (Fast Healthcare Interoperability Resources), SNOMED CT
PHI (protected health information): Health information that can identify an individual. The term comes from US HIPAA; other laws, such as the EU GDPR and India's DPDP Act, use their own definitions of personal and health data.See also: De-identification
Positive predictive value (PPV): Of the people a test or model flags as positive, the proportion who truly have the condition. Called 'precision' in machine learning. It falls as the condition becomes less common.See also: Sensitivity (recall), Specificity, Confusion matrix
Prompt injection: An attack in which text inside a document, web page or tool result contains instructions that try to make a language model ignore its original task. A key risk for agents and RAG systems.See also: AI agent, RAG (retrieval-augmented generation)
Pseudonymisation: Replacing direct identifiers with a code while someone keeps a separate key that can link the code back to the person. Under the GDPR, pseudonymised data is still personal data for anyone able to re-identify it.See also: De-identification, k-anonymity
RAG (retrieval-augmented generation): A method in which relevant passages are first retrieved from a trusted collection (such as clinical guidelines) and then given to a language model to answer from, so answers can cite sources.See also: Embedding, Vector store, Grounding
SaMD (Software as a Medical Device): Software intended for a medical purpose, such as diagnosis or treatment decisions, that performs that purpose without being part of a hardware device. Many clinical AI tools are regulated as SaMD.
Sensitivity (recall): Of the people who have the condition, the proportion the test or model correctly identifies. TP / (TP + FN).See also: Specificity, Positive predictive value (PPV)
SHAP: SHapley Additive exPlanations: a method that attributes a model's prediction to each input feature, based on Shapley values from game theory. It explains the model's behaviour, not the patient's biology.
Shortcut learning: When a model relies on a feature that happens to correlate with the label in its training data, such as a hospital marker on an X-ray, rather than the clinical signal. It fails when that correlation changes.See also: Dataset shift, External validation
SNOMED CT: A large clinical terminology with coded concepts for findings, diagnoses, procedures and more, used to record clinical information consistently in health records.See also: OMOP Common Data Model, FHIR (Fast Healthcare Interoperability Resources)
Specificity: Of the people who do not have the condition, the proportion the test or model correctly clears. TN / (TN + FP).See also: Sensitivity (recall), Positive predictive value (PPV)
Synthetic data: Artificial records generated by a simulator or model to resemble real data. Simulated data such as Synthea's contains no real patients; data generated from real records must be tested to make sure it does not leak them.See also: De-identification, Differential privacy
Target trial emulation: Designing an observational study to mimic a specific randomised trial, with explicit eligibility criteria, time zero, treatment strategies, outcomes and follow-up, to reduce common biases when estimating treatment effects.See also: Confounding
Temperature: A setting that controls how a language model chooses each next token: low values almost always pick the most likely token; higher values make less likely tokens more probable, increasing variety and the chance of errors.See also: Token, Hallucination
Token: The unit of text a language model reads and writes, usually a word or part of a word. Model context limits and costs are measured in tokens.See also: Large language model (LLM)
Tool calling: A model capability in which the model returns a structured request to run a specific function (for example, 'look up this patient's latest HbA1c'), which the application executes and returns to the model.See also: AI agent, MCP (Model Context Protocol)
UCUM: The Unified Code for Units of Measure: a system of unambiguous codes for units such as mg/dL, mmol/L and %, used in FHIR and other standards so software can convert and compare measurements reliably.See also: LOINC
Unsupervised learning: Machine learning without an outcome label, used to find structure such as groups of similar patients (clustering) or a few underlying dimensions (dimensionality reduction).See also: k-means clustering
Vector store: A database that stores embeddings and finds the items most similar to a query, used for retrieval in RAG systems.See also: Embedding, RAG (retrieval-augmented generation)
Vision-language model: A multimodal model that processes images and text together, either by aligning them in a shared embedding space or by connecting an image encoder to a language model that can describe images.See also: Embedding, Zero-shot learning
Zero-shot learning: Performing a task, such as classifying an image, without examples specific to that task, for instance by comparing an image with text descriptions of each possible category.See also: Vision-language model, Few-shot prompting

Educational content only. This material teaches methods; it is not clinical advice. Code, models and examples here are not validated medical devices and must not be used to make decisions about real patients without proper clinical validation, governance and any required regulatory approval. Examples use synthetic or openly licensed de-identified data only. Never paste identifiable patient information into public AI tools. Mentions of models, tools or vendors are for identification, not endorsement.