AI in Medicine Glossary | AI Medicine Today

#

510(k) Clearance premarket notification, 510k: The most common FDA route to market, based on demonstrating substantial equivalence to a legally marketed predicate device rather than proving clinical benefit.

A

Adverse Event Reporting MDR, MedWatch: The mandatory pathway for reporting device-associated deaths, serious injuries, and malfunctions to the FDA.
AI (Medical Abbreviation) AI, A.I.: In a clinical chart, AI usually means aortic insufficiency. In technology contexts it means artificial intelligence. In nutrition it means adequate intake.
Alert Fatigue alarm fatigue: Desensitization to system warnings caused by high volume and low specificity, leading clinicians to dismiss alerts including the important ones.
Algorithmic Bias model bias, health equity bias: Systematic differences in a model's performance across patient groups, usually inherited from unrepresentative training data or biased labels.
Ambient Clinical Documentation ambient AI, AI scribe, ambient scribe: Software that listens to a clinical encounter and drafts the note, currently the highest-adoption category of AI in physician workflows.
Artificial Intelligence in Medicine medical AI, clinical AI: Software that performs tasks in healthcare which normally require human judgment, most often pattern recognition in images, text, or signals.
Assistive AI adjunctive AI, clinician-in-the-loop AI: A system whose output informs a clinician who remains responsible for the decision, which describes almost all authorized medical AI.
AUROC AUC, C-statistic, area under the ROC curve: A single number summarizing discrimination across all thresholds: the probability the model ranks a random positive case above a random negative one.
Automation Bias automation complacency: The documented tendency of people to over-trust automated output, accepting incorrect suggestions and failing to catch what the system missed.
Autonomous AI autonomous diagnostic AI: A system that produces a clinical result without a clinician interpreting the underlying data, currently authorized in a small number of narrow indications.

B

Black Box Model opaque model: A model whose internal decision process cannot be meaningfully inspected, which describes most high-performing deep learning systems.

C

Calibration probability calibration: Whether a model's stated probabilities match observed frequencies ... when it says 30 percent, does the event happen about 30 percent of the time.
Class Imbalance imbalanced data: When one outcome is far more common than another in a dataset, which distorts both training and the interpretation of accuracy.
Clinical Decision Support Software CDS, CDSS: Software providing clinicians with information or recommendations to inform a decision, some of which is regulated as a device and some of which is not.
Computer Vision image analysis, CV: The field concerned with extracting information from images, and the technical basis of nearly all authorized medical imaging AI.
Computer-Aided Detection CADe: Software that marks locations of possible findings for the clinician to evaluate, without characterizing what the finding is.
Computer-Aided Diagnosis CADx: Software that characterizes a finding, for example estimating whether a lesion is benign or malignant, rather than only locating it.
Computer-Aided Triage CADt: Software that reprioritizes a worklist by flagging studies with suspected time-critical findings, without marking images or offering a diagnosis.

D

Dataset Shift distribution shift, domain shift, covariate shift: When the data a model sees in deployment differs statistically from its training data, degrading performance without any visible error.
De Novo Authorization De Novo classification, De Novo request: The FDA route for novel low-to-moderate risk devices with no existing predicate, which creates a new classification future devices can then cite.
De-identification anonymization, Safe Harbor: Removing identifiers from health data so it is no longer protected health information under HIPAA and can be used more freely.
Deep Learning deep neural networks: Machine learning using neural networks with many layers, which learn their own features from raw data instead of using features a human specified.
DICOM Digital Imaging and Communications in Medicine: The standard governing medical image formats and transmission, which is why imaging became the first viable target for medical AI.
Digital Pathology: The practice of reading pathology on digitized slides rather than through a microscope, and the platform on which pathology AI runs.

E

Explainability interpretability, XAI: The degree to which a model's reasoning can be understood and evaluated by a person, and a live regulatory and clinical concern.
External Validation independent validation: Evaluating a model on data from a different institution, population, or equipment than it was trained on.

F

Federated Learning distributed learning: Training a shared model across institutions without moving patient data, by exchanging model updates instead of records.
FHIR Fast Healthcare Interoperability Resources: A modern standard for exchanging healthcare data through web APIs, and the main route for AI tools to read and write EHR data.
Foundation Model base model, pretrained model: A large model pretrained broadly on general data, then adapted to specific downstream tasks rather than trained for one task from scratch.

G

Generalizability transportability, external validity: Whether a model's performance holds in populations, settings, and equipment other than those it was developed on.
Generative AI genAI: Models that produce new content ... text, images, or structured output ... rather than only classifying or scoring an input.
Ground Truth reference standard, gold standard: The labels a model is trained and evaluated against, which define the ceiling on what the model can learn to do.

H

Hallucination confabulation, fabrication: When a generative model produces content that is fluent and plausible but factually wrong or entirely invented.
HL7 Health Level Seven: The standards organization and its long-established messaging formats for exchanging clinical data between systems.
Human in the Loop HITL, clinician oversight: A design where a person reviews or approves model output before it affects care, which is the basis for most regulatory exemptions in clinical software.

I

Intended Use indications for use: The specific clinical purpose, population, and setting a device is authorized for, which defines the boundary of its clearance.
Internal Validation held-out validation, split-sample validation: Evaluating a model on data held out from the same source as its training data.

L

Large Language Model LLM: A model trained on large volumes of text to predict likely continuations, producing fluent language across a wide range of tasks.

M

Machine Learning ML, statistical learning: A method of building software where the system derives its behavior from patterns in training data rather than from explicitly written rules.
Model Drift performance drift, concept drift: Gradual degradation of a deployed model's performance over time as clinical practice, populations, or equipment change around it.

N

Narrow AI weak AI, task-specific AI: A system competent at one defined task and nothing else, which describes every AI system currently used in medicine.
Natural Language Processing NLP: Computational handling of human language, used in medicine to extract structured information from unstructured clinical text.
Negative Predictive Value NPV: The proportion of negative results that are genuinely negative, also dependent on the prevalence of the condition.
Neural Network artificial neural network, ANN: A model made of layered interconnected units with weighted connections, adjusted during training until the network produces the desired outputs.

O

Overfitting memorization: When a model learns noise and idiosyncrasies specific to its training data rather than the underlying pattern, performing well in training and poorly elsewhere.

P

PACS picture archiving and communication system: The system that stores, retrieves, and displays medical images, and the integration point where most imaging AI is deployed.
Positive Predictive Value PPV, precision: The proportion of positive results that are genuinely positive, which depends heavily on how common the condition is in the tested population.
Post-Market Surveillance PMS: Monitoring a device's real-world performance after authorization, which for adaptive AI is where most of the meaningful evidence has to come from.
Predetermined Change Control Plan PCCP: An FDA mechanism letting a manufacturer specify in advance how a model may be updated after authorization without requiring a new submission each time.
Predicate Device predicate: A legally marketed device that a new 510(k) submission claims substantial equivalence to.
Premarket Approval PMA: The FDA's most demanding route, requiring clinical evidence of safety and effectiveness, used for the highest-risk devices and rare in medical AI.
Prospective Validation prospective study: Evaluating a model on patients enrolled after the study begins, in the real workflow, rather than on previously collected data.
Protected Health Information PHI: Individually identifiable health information covered by HIPAA, which governs how clinical data may be used for AI development and deployment.

R

Radiomics quantitative imaging features: Extracting large numbers of quantitative features from medical images to use as inputs to statistical or machine learning models.
Retrieval-Augmented Generation RAG, grounded generation: Retrieving relevant source documents first, then generating an answer constrained to those sources, so output can be traced back to a citation.
Retrospective Validation retrospective study: Evaluating a model on previously collected data, which is faster and cheaper than prospective study and systematically more optimistic.

S

Saliency Map heat map, attention map, Grad-CAM: A visual overlay indicating which image regions most influenced a model's output, widely used and widely over-interpreted.
Sensitivity recall, true positive rate: The proportion of patients who have the condition that the test correctly identifies.
Software as a Medical Device SaMD: Software intended for a medical purpose that performs that purpose without being part of a hardware medical device.
Specificity true negative rate: The proportion of patients who do not have the condition that the test correctly identifies as negative.
Substantial Equivalence SE: The 510(k) standard: a new device has the same intended use and raises no new questions of safety and effectiveness compared with its predicate.

W

Whole Slide Imaging WSI, digital slide: Scanning an entire glass pathology slide into a high-resolution digital image, which is the prerequisite for any pathology AI.