Glossary

A curated, clinically grounded reference index of AI and digital health terminology as it applies to healthcare contexts. Covers AI/ML technical terms (model drift, hallucination, AUROC, sensitivity/specificity, foundation models), regulatory terminology (SaMD, 510(k), De Novo, PMA, post-market surveillance), clinical workflow terms (ambient clinical intelligence, clinical decision support, interoperability, HL7 FHIR), and methodological concepts (algorithmic bias, prospective validation, training/test set). Each entry provides a plain-language definition, clinical relevance explanation, and where applicable, links to authoritative sources (FDA glossary, CHAI standards, primary literature). This group serves all audience segments — from clinicians new to AI to researchers needing precise regulatory terminology. Excludes marketing definitions and vendor-specific terminology presented as neutral.

AI/ML fundamentals

Federated Learning in Healthcare AI: Definition, Privacy Mechanisms, and Clinical EvidenceA clinically grounded reference entry on federated learning as it applies in healthcare AI — covering how distributed mo…Clinicians and health IT leaders evaluating multi-site AI tools need to understand that FL enables cross-institutional model training without raw data transfer — addressing HIPAA and GDPR barriers — but that baseline FL provides no formal privacy guarantee without additional layers such as differential privacy or secure aggregation. Procurement teams should require disclosure of which privacy-enhancement mechanisms are deployed and at what epsilon level.
Hallucination in Clinical LLMs: Definition, Causes, Detection, and Deployment ImplicationsA clinically grounded reference entry defining LLM hallucination as it applies to healthcare AI — covering its taxonomy,…Clinical LLM hallucination is a patient safety risk, not merely a usability issue. Clinicians evaluating or deploying AI tools must understand hallucination taxonomy, detection limits, and mitigation requirements — particularly that no current technique eliminates hallucination entirely and that task-specific validation, ongoing monitoring, and human-in-the-loop workflows are minimum deployment safeguards.
Model Drift in Clinical AI: Correction Strategies, Retraining Governance, and Regulatory Lifecycle ManagementA structured reference for clinical informaticists, AI governance leads, and health system engineers on how to respond w…Clinicians and governance leads overseeing deployed AI must distinguish calibration drift (recalibrate only) from concept drift (full retraining required) to avoid both uncorrected degradation and unnecessary retraining that introduces new failure modes. Contractual clarity with vendors on correction responsibilities is essential, as only 16% of health systems have a systemwide AI governance policy addressing model updates.
Transfer Learning and Fine-Tuning in Clinical AI: Definitions, Strategies, and Governance ImplicationsA clinically grounded glossary reference defining transfer learning, fine-tuning, and the full taxonomy of adaptation st…Clinicians and health IT teams evaluating AI tools need to understand how a vendor's model was adapted—whether via full fine-tuning, PEFT, or domain adaptation—because the adaptation strategy directly affects failure risks (catastrophic forgetting, negative transfer, overfitting), regulatory status (locked vs. adaptive classification, PCCP requirements), and the validity of performance claims made on fine-tuning datasets versus real-world deployment populations.

AI/ML methodology

Foundation Models in Healthcare: Definition, Architecture, and Clinical Scope — with Med-PaLM as the Reference CaseA structured reference entry defining foundation models as a distinct class of healthcare AI — large systems pre-trained…
Model Drift in Deployed Clinical AI: Definition, Types, Causes, Detection, and MonitoringA structured reference entry defining model drift in the context of deployed clinical AI systems — covering its clinical…

Regulatory terminology

Predetermined Change Control Plan (PCCP): The FDA Mechanism for Iterative AI/ML Medical Device UpdatesA Predetermined Change Control Plan (PCCP) is the FDA regulatory mechanism that allows AI-enabled medical device manufac…

clinical workflow terms

Ambient Intelligence in Healthcare: Definition, Technologies, Clinical Applications, and EvidenceA comprehensive glossary entry on ambient intelligence (AmI) in healthcare for clinicians, researchers, and health IT pr…Reduces documentation burden and burnout by passively capturing clinician-patient conversation, enabling more eye contact and patient-centered care.
Conversational AI in Healthcare: What the Evidence Shows About Accuracy, Safety, and OutcomesAn evidence-grounded reference for clinicians, researchers, and clinical informaticists evaluating conversational AI too…Evaluates conversational AI for screening, documentation, triage, and mental health; strongest evidence for screening and documentation, significant gaps for diagnostic use.
NLP in Clinical Documentation: AI Scribes, Coding, and Clinical Documentation ImprovementA clinically grounded appraisal of NLP-based documentation tools across three operationally distinct domains — ambient A…Clinicians and procurement teams evaluating NLP documentation tools must distinguish three operationally distinct domains — ambient AI scribing, computer-assisted ICD/CPT coding, and CDI query flagging — each with different evidence quality, failure modes, and regulatory status. Peer-reviewed evidence is consistently limited by small samples, short follow-up, and absent patient outcome data; mandatory clinician review is a clinical necessity, not optional oversight.
What Is AI in Revenue Cycle Management? A Glossary-Style Reference on Technologies, Applications, and EvidenceA structured, evidence-grounded glossary entry for healthcare professionals, HIM specialists, and revenue cycle managers…AI in RCM directly impacts provider revenue, denial rates, and administrative burden; understanding these technologies is essential for HIM specialists and revenue cycle managers evaluating tools and responding to payer AI strategies.

ethics and bias

Algorithmic Bias in Clinical AI: Audit Frameworks and Mitigation MethodsA structured reference for healthcare professionals, researchers, and compliance officers covering the lifecycle of algo…Bias in clinical AI can cause systematic underdiagnosis or misdiagnosis of patient subgroups, widening health disparities; audit frameworks and stage-specific mitigation methods are essential for equitable AI deployment.
Algorithmic Bias in Healthcare AI: Definition, Taxonomy, and Mitigation FrameworksA structured reference entry defining algorithmic bias in healthcare AI across its four origin categories — human, data,…Clinicians evaluating AI tools for adoption should request demographic subgroup performance data stratified by race, ethnicity, sex, age, and site — aggregate AUC or sensitivity figures do not reliably indicate equitable performance across the patient populations a health system serves. Procurement teams should assess vendor documentation for bias testing across the full AI lifecycle, including post-deployment surveillance plans.

performance metrics

Clinical AI Model Evaluation Metrics: A Five-Domain Reference for AUROC, Calibration, Sensitivity, Specificity, and Net BenefitA structured reference for clinicians, researchers, and procurement specialists explaining the five distinct performance…Clinicians and procurement specialists evaluating AI tools need all five performance domains — discrimination, calibration, overall performance, classification, and clinical utility — to judge deployment readiness. A model reporting only AUROC or F1 is incompletely characterized; missing calibration data can mask systematic overtreatment or undertreatment, and absent net benefit analysis leaves the core clinical question unanswered.

regulatory terms

CONSORT-AI Reporting Standard for Artificial Intelligence in Clinical TrialsA structured reference entry explaining the CONSORT-AI and SPIRIT-AI reporting standards — the 2020 international guidel…Ethics committees, trial methodologists, and peer reviewers working with AI-based interventional trials must apply SPIRIT-AI (15 items, protocol stage) and CONSORT-AI (14 items, report stage) together to ensure AI trials are reproducible, transparently registered, and safe. Algorithm version, input data eligibility, and performance error planning — the most under-reported items — directly affect patient safety monitoring and trial reproducibility. As of Q2 2026, neither SPIRIT 2025 nor CONSORT 2025 incorporates AI-specific guidance, so these 2020 extensions remain the operative standards.
Substantial Equivalence and Predicate Device: FDA 510(k) Concepts for AI-Enabled SaMDA clinically grounded glossary entry defining substantial equivalence and predicate device as FDA regulatory concepts, e…Clinicians and procurement teams evaluating AI-enabled SaMD should understand that 510(k) clearance via substantial equivalence does not guarantee functional similarity to a predicate — predicate creep and non-AI-to-AI predicate transitions mean a cleared AI device may differ materially from its regulatory benchmark. Procurement due diligence should include reviewing a device's predicate chain and whether demographic-stratified performance data was required under 2025 FDA guidance.

technical AI/ML

AI in Medical Research: A Glossary Reference on Technologies, Applications, and LimitationsA structured glossary entry defining AI in medical research — covering machine learning, deep learning, LLMs, generative…Defines AI technologies used across the medical research lifecycle, enabling clinicians and researchers to evaluate applications in drug discovery, trials, imaging, and publishing.
Building an Institutional Monitoring Program for Clinical AI Model DriftA practical implementation guide for health system CIOs, clinical informaticists, and AI governance committees on how to…Clinicians and health IT leaders procuring or overseeing deployed AI tools need to know whether their institution has formal drift monitoring in place — including assigned roles, detection thresholds, and escalation workflows — because undetected model degradation can affect patient safety without triggering any visible alert.
Foundation Models in Healthcare: Clinical Applications and LimitationsA structured glossary reference covering the four primary families of clinical foundation models — CLaMs, FEMRs, vision …Clinicians and procurement teams evaluating foundation model tools need to distinguish which FM family applies to their clinical data type (text, structured EHR, imaging, or multimodal), verify the evidence tier of supporting studies before deployment, and treat benchmark performance as distinct from clinical readiness — particularly given the 33% LLM outperformance rate and absence of prospective multi-site validation across all imaging FM families as of mid-2026.