Appendix D — Glossary of Medical AI Terms
Medical AI terms often carry different meanings across statistics, computer science, regulation, law, and clinical practice. This glossary uses evidence-bounded definitions and points readers to the chapter where the operational or legal context is developed.
After using this glossary, readers should be able to:
- Distinguish technical performance terms from clinical utility, regulatory authorization, and legal standards.
- Interpret common metrics without relying on universal thresholds detached from prevalence, consequences, and use case.
- Recognize when a popular label, such as hallucination or AI-associated psychosis, is informal or causally unsettled.
- Follow cross-references into the handbook’s evidence, safety, privacy, liability, and specialty chapters.
Quick reference for 100+ medical AI terms. Key concepts:
- AI vs. ML vs. Deep Learning: AI is the broad field; ML is AI that learns from data; deep learning uses neural networks
- FDA pathways: 510(k) clearance generally addresses substantial equivalence; De Novo establishes a class I or II device type when no legally marketed predicate exists; PMA evaluates reasonable assurance of safety and effectiveness for a class III device’s intended use
- Model metrics: Sensitivity (catches disease), specificity (avoids false alarms), AUC (overall discrimination)
- LLMs: Large language models like GPT-4, trained on text, generate human-like responses
- Validation types: Internal (same data source), external (different population), prospective (real-time)
Introduction
This glossary defines key terms used throughout the Physician AI Handbook. Terms are organized alphabetically and explained in practical, clinically relevant language.
A
AI (Artificial Intelligence): Computer systems designed to perform tasks that typically require human intelligence, such as visual perception, speech recognition, decision-making, and pattern detection. In medicine, AI encompasses diagnostic algorithms, clinical decision support systems, and predictive models.
AI-Associated Psychosis, sometimes called AI-Induced Psychosis or Chatbot Psychosis: An informal label for reports in which psychotic-spectrum symptoms arise or worsen in temporal association with conversational-AI use. It is not a formal diagnosis, and current evidence does not establish that chatbots cause psychosis. A published case report describes new-onset psychosis following intensive chatbot interaction, but a case report establishes temporal association rather than incidence or causation (Pierre et al., 2025). A 2026 prompt-response study found inappropriate or partially appropriate responses to psychotic prompts across tested ChatGPT versions; it evaluated single exchanges and response appropriateness, not whether chatbot use caused illness (Shen et al., 2026). See also Chatbot Psychosis and Sycophancy, and Psychiatry and Behavioral Health.
Algorithm: A step-by-step procedure or formula for solving a problem. In medical AI, algorithms process patient data (images, labs, clinical notes) to generate predictions, diagnoses, or treatment recommendations.
Algorithmic Bias: Systematic model or system behavior that creates differential error, allocation, access, or burden across groups. Sources can include the target chosen, historical inequity, measurement, sampling, missingness, labels, model design, thresholds, workflow, access, and institutional response, not only underrepresentation in training data.
Anchoring Bias (AI Context): The tendency for an initial AI output to influence subsequent clinical judgment disproportionately. Interface sequencing, required independent review, uncertainty displays, and training may affect anchoring, but mitigation should be tested in the actual workflow. See also Automation Bias.
Automation Bias: The tendency to favor an automated output in a way that contributes to omission or commission errors. The effect varies with task, user, interface, workload, trust, and error pattern; high average accuracy does not eliminate the need to study what happens when the system is wrong.
AUROC (Area Under the Receiver Operating Characteristic Curve): A threshold-independent summary of how a model ranks observations from two classes. Under common conditions, 0.5 corresponds to random ranking and 1.0 to perfect ranking. No universal AUROC threshold establishes clinical usefulness. Calibration, clinically relevant thresholds, sensitivity, specificity, predictive values, decision consequences, uncertainty, and comparator performance must also be assessed.
Autonomous AI: A context-dependent label for a system that completes a defined function without a specialist interpreting the input before the output is issued. Autonomy is bounded by the indication, user, equipment, quality checks, referral pathway, and fallback process. IDx-DR, now marketed as LumineticsCore, was authorized for an autonomous diabetic-retinopathy detection function within its labeling; it does not operate without any human, organizational, or clinical oversight. Contrast with adjunctive decision support.
B
Batch Learning: Training or updating a model on a defined dataset before deployment. A deployed version may remain fixed until a controlled update. Many authorized devices are version-controlled, but the exact update mechanism must be verified in the product’s labeling and authorization record. Contrast with online or continuous learning.
Bedside to Bench: A discovery route that begins with patient-level observations or clinical data and generates hypotheses for mechanistic or laboratory investigation. Pierson and colleagues used knee radiographs to predict pain and examined whether image information could help explain racial disparities not captured by conventional severity grading; the study did not itself identify a biological mechanism (Pierson et al., 2021).
Bias (Statistical): Systematic deviation of AI predictions from true values. Sources include non-representative training data, flawed data collection, or algorithmic design choices. See also Algorithmic Bias.
Black Box: AI models whose internal logic is opaque: input goes in, prediction comes out, but how the model reached its decision is unclear. Deep neural networks often criticized as black boxes. Contrast with interpretable/explainable models.
C
CAD (Computer-Aided Detection): AI systems that flag suspicious findings for physician review, e.g., highlighting lung nodules on chest X-rays or calcifications on mammograms. CAD augments but does not replace human interpretation.
Calibration: The degree to which an AI model’s predicted probabilities match observed frequencies. A well-calibrated model that predicts “70% probability of disease” should be correct 70% of the time. Poorly calibrated models may be overconfident or underconfident.
Chatbot Psychosis: Colloquial, non-diagnostic term used for reports in which chatbot interactions and psychotic symptoms are temporally associated. A Danish preprint described records in which chatbot use was judged potentially harmful, including delusion-related concerns, but observational clinical records cannot establish incidence or causation (Olsen et al., 2025, preprint). See also AI-Associated Psychosis and Sycophancy.
CDSS (Clinical Decision Support System): Software providing clinicians with patient-specific assessments or recommendations to aid decision-making. Includes simple rule-based systems (alert if drug interaction detected) and complex AI-powered predictions (sepsis risk scores).
Cognitive De-skilling: Potential degradation of unaided performance after sustained reliance on automated assistance. In a multicenter retrospective observational colonoscopy study, adenoma detection during non-AI-assisted examinations was lower after centers introduced AI exposure, an association the authors interpreted as possible deskilling rather than proof of an irreversible causal effect (Budzyń et al., 2025). The risk, magnitude, reversibility, and mitigation strategy must be studied for each task and workforce.
Confounding: Distortion of an exposure-outcome association by another variable related to both. In prediction modeling, shortcut learning is a related but distinct problem: a model may exploit site, equipment, acquisition, or care-process signals that correlate with the label without representing the target pathology.
Continuous Learning: A model-update process in which new data contribute to subsequent model changes. The term should not be used for a fixed model that is merely monitored or periodically replaced. In regulated devices, the permitted change process depends on the authorization, quality system, labeling, and any authorized predetermined change control plan.
D
Data Augmentation: Techniques to artificially increase training dataset size by creating modified versions of existing data (e.g., rotating medical images, adding noise). Helps prevent overfitting and improves generalization.
Deep Learning: A subset of machine learning using artificial neural networks with multiple layers. It is widely used for representation learning and pattern recognition in medical images, text, waveforms, and other complex data. Performance depends on data, task, validation, comparison, and deployment rather than the method name alone.
Deployment Drift: Change over time in input distributions, clinical practice, labels, prevalence, equipment, or the relationship between predictors and outcomes. Drift does not automatically mean performance has degraded; monitoring must connect distributional change to prespecified performance and clinical consequences.
Deferral (AI Context): The deliberate design of a system to abstain or route a case for human review when an input is outside scope, quality is inadequate, or uncertainty exceeds a defined action threshold. Deferral changes the evaluated population and workload, so performance must be reported for deferred and non-deferred cases, along with unsafe non-deferral, under-triage, omission, and capacity consequences. See Clinical AI Safety.
Differential Privacy: A mathematical framework that bounds how much an output distribution can change when one individual’s data are added or removed, under stated parameters and assumptions. It can be implemented with calibrated randomness and may create a privacy-utility tradeoff. Differential privacy provides a formal guarantee for a defined mechanism; it does not by itself secure the full data pipeline, prevent every attack, or establish HIPAA compliance. See Privacy, HIPAA, and Patient Data Security.
Dual-Sided Liability Argument: A potential litigation theory in which a claimant alleges unreasonable reliance on an AI system or, in a different case, unreasonable failure to use an established tool. Neither allegation creates a universal legal rule. Liability remains fact-, duty-, jurisdiction-, product-, workflow-, and evidence-specific. See Physician AI Liability and Regulatory Compliance.
E
Edge AI: AI models running locally on devices (smartphones, medical equipment, wearables) rather than cloud servers. Benefits include faster processing, offline capability, and enhanced privacy. Examples: smartphone-based ECG interpretation, on-device ultrasound AI.
Ensemble Model: Combining multiple AI models to improve prediction accuracy. For example, averaging predictions from 5 different lung nodule detection algorithms may outperform any single algorithm.
Explainable AI (XAI): Methods intended to provide human-interpretable information about model behavior or a particular output, such as feature attributions, example-based explanations, or highlighted image regions. An explanation may be unstable, incomplete, or unfaithful to the model’s actual reasoning and should be evaluated for the intended user and decision.
External Validation: Evaluation on data meaningfully separated from model development, such as another institution, population, geography, or time period. It provides evidence about transportability within the tested conditions; one external dataset does not establish universal generalizability.
F
False Negative: When AI incorrectly predicts “no disease” for a patient who actually has the disease. Clinically dangerous: missed diagnoses, delayed treatment. Measured by sensitivity (1 - false negative rate).
False Positive: When AI incorrectly predicts “disease present” for a healthy patient. Leads to unnecessary anxiety, testing, procedures, costs. Measured by specificity (1 - false positive rate).
FDA (Food and Drug Administration): U.S. agency that regulates medical products, including medical devices. Some AI-enabled software functions are medical devices subject to premarket review, while others may be non-device functions, exempt, enforcement-discretion functions, or outside FDA’s device jurisdiction. Device pathways include 510(k) clearance, De Novo authorization, and PMA approval; the exact function and intended use determine the analysis.
Feature: An individual measurable property or characteristic used as input for an AI model. In medical AI, features may include patient age, lab values, pixel intensities in images, or words in clinical notes.
Federated Learning: A distributed training approach in which participating sites compute local updates without pooling raw records in one central dataset. Federated learning can reduce central data movement, but updates may still leak information and the system still requires privacy, security, governance, and legal controls.
Fine-Tuning: Adapting a pre-trained AI model to a specific task or dataset. For example, taking a general image recognition model and fine-tuning it to detect lung nodules using chest X-rays.
Foundation Model: A model trained on broad data and adaptable to multiple downstream tasks through prompting, fine-tuning, adapters, retrieval, or other methods. Scale alone does not establish that a model is a foundation model, and downstream clinical claims require task-specific evidence.
G
Generalization: An AI model’s ability to perform well on new, previously unseen data. Poor generalization (overfitting) occurs when a model memorizes training data but fails on real-world cases.
Generative AI: AI that creates new content (text, images, audio, video). In medicine, applications include generating synthetic medical images for training, drafting clinical notes, or creating patient education materials. Examples: ChatGPT, DALL-E.
GMLP (Good Machine Learning Practice): Guidelines developed by FDA, Health Canada, and UK MHRA outlining best practices for developing and maintaining medical AI. Covers data quality, model validation, monitoring, and transparency.
Ground Truth: Common but potentially misleading shorthand for the reference standard used to evaluate a model. Expert labels, pathology, laboratory results, follow-up, and consensus adjudication can all contain error or represent different constructs. The reference standard and its uncertainty should be reported rather than assumed to be truth.
H
Hallucination: When AI generates plausible-sounding but incorrect or fabricated information. Common problem with large language models, such as citing non-existent research papers or recommending unproven treatments. Dangerous in clinical contexts.
Health AI Developer Foundations (HAI-DEF): Google’s program name for a set of health-AI model and developer resources, including MedGemma-family releases. Model licenses, versions, supported modalities, access routes, and acceptable uses must be checked in the current official model documentation. Availability does not establish clinical validation or authorization.
Hyperparameter: Settings that control how an AI model learns, such as learning rate, number of layers in a neural network, or regularization strength. Tuned during model development to optimize performance.
I
Imbalanced Data: When one class vastly outnumbers others in training data. Example: 95% normal chest X-rays, 5% with pneumonia. AI may achieve high accuracy by predicting “normal” for everything, but fail to detect rare disease. Requires special techniques to address.
Informed Consent and AI: Consent and disclosure duties depend on the intervention, jurisdiction, risk, materiality, data use, institutional policy, and whether the AI changes the care decision or merely supports an operational task. A universal patient-consent rule does not apply to every AI function.
Interpretability: The degree to which a person can understand a model’s structure, behavior, or output for a defined purpose. Interpretability is user- and task-dependent; a technically interpretable model may still be misunderstood, and a post hoc explanation may not faithfully describe the computation.
L
Labeled Data: Training data with known outcomes or expert annotations. Example: chest X-rays labeled “pneumonia present” or “normal” by radiologists. Most supervised learning requires large amounts of labeled data, which is expensive and time-consuming to obtain.
Large Language Model (LLM): AI trained on vast amounts of text to generate human-like language. Examples: GPT-4, Claude, Med-PaLM. In medicine, LLMs answer clinical questions, summarize literature, generate documentation, but risk hallucinations.
Latent Space: Abstract, high-dimensional representation of data learned by neural networks. In medical imaging, latent space captures patterns (textures, shapes) that distinguish diseased from healthy tissue, even if not explicitly programmed.
M
Machine Learning (ML): Subset of AI where systems learn patterns from data without explicit programming. Includes supervised learning (learning from labeled data), unsupervised learning (finding patterns in unlabeled data), and reinforcement learning (learning through trial-and-error).
Model: The mathematical representation learned by an AI algorithm from training data. Once trained, the model makes predictions on new data. In medical AI, models range from simple logistic regression to complex deep neural networks.
MedGemma: A Google medical-model family distributed through Health AI Developer Foundations. Version, license, modalities, benchmark results, and access routes change over time. A model release is a research and development resource, not evidence of clinical safety, effectiveness, or FDA authorization. See Emerging AI Technologies for the current version-specific review.
Multimodal AI: A system that processes or relates more than one data modality, such as images, text, genomics, laboratory results, or wearable signals. Multimodal input can add information, but it can also add missingness, shortcut learning, alignment errors, privacy exposure, and workflow complexity.
N
Natural Language Processing (NLP): AI techniques for analyzing human language. In medicine, NLP extracts structured information from unstructured clinical notes, generates documentation, or powers chatbots answering patient questions.
Neural Network: AI model inspired by brain structure, consisting of interconnected nodes (neurons) organized in layers. Information flows from input layer through hidden layers to output layer. Deep neural networks have many hidden layers.
NPV (Negative Predictive Value): Probability that a patient with a negative AI prediction truly does not have the disease. Depends on disease prevalence: high when disease is rare, lower when common.
O
Oculomics: Inferring systemic disease risk from ocular imaging (fundus photographs, OCT layer thicknesses, or retinovascular indices). An association with a systemic endpoint is not a stand-alone diagnostic test for that endpoint. See Cardiology.
Opportunistic screening: Using an already-acquired test or image to look for a condition other than the original indication. Echo-derived CKD screening is one example: discrimination in echo-referred cohorts is not a replacement for eGFR or proof that alerts improve outcomes. See Cardiology.
Overfitting: When a fitted model captures development-data idiosyncrasies that do not transport to new observations. Regularization, resampling, simpler representations, larger or more representative samples, and strict separation of tuning from final evaluation can reduce optimism, but no method guarantees prevention.
P
PCCP (Predetermined Change Control Plan): A device-specific plan included in a marketing submission that describes anticipated modifications, the methods used to develop, validate, and implement them, and an impact assessment. When FDA authorizes a PCCP, modifications implemented consistently with that authorized plan may not require a new marketing submission. A PCCP is not blanket permission for continuous or unbounded model change.
Precision (Positive Predictive Value): Proportion of positive AI predictions that are correct. High precision = few false positives. Critical when false positives are costly (unnecessary biopsies, anxiety, procedures).
Pre-training: Initial training of an AI model on a large, general dataset before fine-tuning on a specific medical task. Example: pre-train on millions of general images, then fine-tune on chest X-rays. Improves performance, especially with limited medical data.
Prospective Validation: Evaluation using data accrued after the evaluation plan or model is fixed. Prospective data collection can reduce some forms of leakage and better represent intended workflow, but design still determines the claim: a prospective single-arm accuracy study is not a randomized outcome trial.
R
Radiomics: Extracting quantitative features from medical images (texture, shape, intensity patterns) and using them for AI prediction. Example: tumor heterogeneity on CT predicting treatment response.
Recall (Sensitivity): Proportion of actual cases correctly identified by AI. High recall = few false negatives. Critical when missing disease is dangerous (e.g., cancer screening).
Reinforcement Learning: AI learns through trial-and-error, receiving rewards for correct actions and penalties for incorrect ones. Rare in clinical medicine (ethical concerns about “trial-and-error” on patients) but used in drug discovery, treatment optimization simulations.
Reproducibility and Legal Risk: Nondeterministic outputs, undocumented version changes, or inconsistent records may complicate reconstruction of an AI-assisted decision. Whether that creates legal exposure depends on the facts, applicable duties, jurisdiction, record-retention rules, product, workflow, and harm. Risk controls can include version and output logging where appropriate, change management, independent review, and documentation of the clinical basis for a decision. See Physician AI Liability and Regulatory Compliance.
Retrospective Validation: Evaluation using previously collected data rather than data accrued prospectively for the evaluation. It can provide important evidence, especially with external data, but is vulnerable to selection, missingness, temporal, workflow, and reference-standard biases. Its limitations depend on the actual design.
ROC Curve (Receiver Operating Characteristic): Graph plotting true positive rate (sensitivity) vs. false positive rate (1-specificity) at various decision thresholds. Used to evaluate AI diagnostic performance. See also AUROC.
S
Sensitivity (Recall): Proportion of reference-positive cases classified positive. Formula: True Positives / (True Positives + False Negatives). The required sensitivity depends on the use, downstream confirmation, consequences of misses, specificity, capacity, and uncertainty.
Specificity: Proportion of healthy patients correctly identified as healthy by AI. Formula: True Negatives / (True Negatives + False Positives). High specificity critical when false positives costly.
Standard of Care and AI: The standard of care is a jurisdiction- and fact-specific legal concept, not a numerical adoption threshold. Relevant evidence may include professional guidance, customary practice, training, available alternatives, patient circumstances, institutional resources, product labeling, and expert testimony. Coverage or widespread use alone does not determine negligence. See Physician AI Liability and Regulatory Compliance.
Supervised Learning: Model development using examples paired with target labels or outcomes. It is common in medical prediction and image-analysis research, but label quality, leakage, sampling, and reference-standard error can limit the resulting model.
Sycophancy (AI Context): A model behavior in which responses become unduly agreeable with a user’s stated view, preference, or premise at the expense of accuracy or appropriate challenge. It can arise from training objectives, preference data, prompting, or conversational context. In mental-health interactions, reinforcing a false or delusional premise is a safety concern, but a response pattern does not establish that the model caused a psychiatric disorder. See also AI-Associated Psychosis and Chatbot Psychosis.
Synthetic Data: Artificially generated records, signals, images, or text intended to approximate aspects of real data without directly representing the original individuals. Synthetic data can support simulation, testing, augmentation, or privacy strategies, but it may reproduce bias, leak information, omit rare structure, or create unrealistic patterns.
T
Test Set: Data reserved from model fitting and tuning for a final evaluation. A test estimate is valid only to the extent that the set is independent, representative of the target use, adequately sized, correctly labeled, and not repeatedly used for model selection. It does not by itself predict real-world deployment performance.
Training Data: Data used to teach an AI model. Model learns patterns from this data. Quality and representativeness of training data critically determine model performance and bias.
Transfer Learning: Applying knowledge learned from one task to a different but related task. Example: AI trained to recognize cats and dogs can be fine-tuned to detect tumors in medical images. Reduces data requirements for medical AI.
Triage: A function that prioritizes cases for review or response, such as flagging a head CT suspected of intracranial hemorrhage. Triage is not necessarily diagnosis. Its effect on review time, treatment, workload, false-alert burden, and outcomes must be measured in the actual workflow.
U
Underfitting: When an AI model is too simple to capture patterns in data, performing poorly on both training and test data. Opposite of overfitting. Fixed by using more complex models or better features.
Unsupervised Learning: AI learning patterns from unlabeled data without explicit outcomes. Applications include clustering similar patients or discovering disease subtypes. Less common in clinical medicine than supervised learning.
V
Validation: Evaluation of a model, device, or workflow against prespecified data, reference standards, comparators, and endpoints. Internal validation estimates optimism within the development context. External validation evaluates transportability in another context. Prospective, reader, workflow, and outcome studies answer additional questions that external validation alone cannot.
Vanishing Gradient Problem: Technical challenge in training very deep neural networks where learning signals become too weak to update early layers. Addressed through architectural innovations (e.g., residual networks).
W
Weakly Supervised Learning: Training AI with imprecise or incomplete labels. Example: using billing codes (imperfect) instead of chart review (expensive) to label diagnoses. Reduces labeling burden but may introduce errors.
X
XAI: See Explainable AI.
Z
Zero-shot Learning: Performance on a task or label without task-specific training examples being supplied in the evaluated adaptation or prompt. For large pretrained models, it is usually impossible to prove that the underlying subject matter never appeared in pretraining data. Zero-shot evaluation therefore requires careful contamination and task-definition analysis.
Conclusion
A precise definition should narrow a claim, not make it sound more certain than the evidence allows. For deeper exploration of a technical, clinical, regulatory, privacy, or legal concept, follow the glossary’s cross-reference into the relevant handbook chapter and verify any time-sensitive product or policy detail at its primary source.