Click here to open the filter panel
#
510(k) premarket notification
A
Accountability
A
Accuracy
A
Adequacy decision
A
Ad hoc services
A
Adverse event
A
AI literacy (Art. 4 EU AI Act)
A
AI Officer
A
Analysis Data Model (ADaM)
A
Anonymisation
A
Appropriate safeguards
A
Article 29 Working Party (WP29)
A
Artificial intelligence (AI)
A
Authorised representative (EU AI Act)
A
Automated individual decision-making
A
Auxiliary medicinal product
B
Best-in-class drugs
B
Binding corporate rules (BCR)
B
Biobank
B
Biometric data
B
Biotechnology
B
Broad consent
B
Business associate agreement (BAA)
C
Central imaging
C
Central laboratory
C
Chemistry, Manufacturing and Controls (CMC)
C
Clinical outcome assessment (COA)
C
Clinical research
C
Clinical study
C
Clinical trial
C
Clinical trial agreement (CTA)
C
Clinical trial phases
C
Clinical Trials Information System (CTIS)
C
Clinical trial site
C
Clinical Trials Regulation (CTR) 536/2014
C
Clinician-reported outcome (ClinRO)
C
CNIL (Commission nationale de l'informatique et des libertés)
C
Codes of conduct
C
Consent
C
Contract research organisation (CRO)
C
Controller
C
Cookies
C
Cross-border processing
D
Data Act
D
Data concerning health
D
Data Governance Act (DGA)
D
Data management plan (DMP)
D
Data minimisation
D
Data monitoring committee (DMC)
D
Data processing agreement (DPA)
D
Data protection authority (DPA)
D
Data protection by default
D
Data protection by design
D
Data Protection Impact Assessment (DPIA)
D
Data Protection Officer (DPO)
D
Data Protection Representative (DPR)
D
Data quality
D
Data retention
D
Data subject
D
Data subject access request (DSAR)
D
Decentralised clinical trial (DCT)
D
De-identification (HIPAA)
D
Differential privacy
E
Early termination of a clinical trial
E
Electronic case report form (eCRF)
E
Electronic clinical outcome assessment (eCOA)
E
Electronic data capture (EDC)
E
Electronic health record (EHR)
E
Electronic informed consent (eConsent)
E
Encryption
E
End of a clinical trial
E
Enterprise
E
Ethics committee
E
EU Artificial Intelligence Act (AI Act)
E
European Data Protection Board (EDPB)
E
European Data Protection Supervisor (EDPS)
E
European Health Data Space (EHDS)
E
European Medicines Agency (EMA)
E
EU-US Data Privacy Framework (DPF)
E
Explicit consent
F
Fairness and transparency
F
Federated learning
F
Filing system
F
First-in-class drugs
F
Food and Drug Administration (FDA)
F
Fundamental rights impact assessment (FRIA)
G
General Data Protection Regulation (GDPR)
G
General-purpose AI model (GPAI)
G
Genetically modified organism (GMO)
G
Genetic data
G
Global public health security
G
Good clinical practice (GCP)
H
Health data warehouse (HDW)
H
High-risk AI system
H
HIPAA (Health Insurance Portability and Accountability Act)
H
Human oversight (AI Act)
H
Human subjects protection review board
C
Imported item 10
S
Imported item 11
C
Imported item 12
Click a letter to view all its definitions
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.

Term of the Day

Natural history study

A natural history study is a preplanned observational study intended to track the course of a disease over time, identifying demographic, genetic, environmental and other variables that correlate with its development and outcomes in the absence of intervention, or under standard of care. Designs may be retrospective (chart review of existing records) or prospective (longitudinal follow-up of a cohort or registry).

Natural history data is particularly important in rare and paediatric diseases, where randomised placebo-controlled trials may be infeasible or unethical. The FDA (guidance on rare disease natural history studies, 2019) and the EMA accept well-designed natural history studies to define endpoints and biomarkers, identify patient subgroups, estimate sample sizes and, in some cases, serve as external or historical control arms for single-arm trials supporting orphan products.

Because they are non-interventional, natural history studies fall outside the CTR and are governed by national law (for example France's MR-003 or MR-004 reference methodologies) and by the GDPR. They typically involve secondary use of medical records, long-term follow-up, genetic data and small populations in which anonymisation is rarely achievable, so pseudonymisation, a DPIA and a robust research legal basis under Art. 9(2)(j) are essential. Registries maintained by patient organisations or academic consortia raise additional questions of joint controllership and data access governance.

M

Imported item 50

A membership inference attack attempts to determine whether a particular record was included in a machine-learning model’s training dataset. It does not necessarily recover the complete record. In a health context, learning that someone belonged to a particular training cohort may itself disclose sensitive information about that person or their care.

Key takeaways

  • Dataset membership can itself be sensitive.
  • An attack need not reconstruct a full record.
  • Test privacy risk under realistic access conditions.

What an Attacker Tries to Learn

The NIST definition focuses on identifying training-set membership. The ICO’s AI security guidance explains why model access, overfitting and exposed outputs can affect this risk.

QuestionMembership inference
TargetWas this record used in training?
Possible signalDifferences in model responses to known records.
Privacy consequenceRevealing association with a sensitive dataset.

Why It Matters for Data Protection

Our recommendation is to include membership inference in model threat assessment where training-set association would be sensitive. Consider the information an attacker could already possess and the access they could obtain. Evaluate measures such as access restrictions, limiting unnecessary output detail, reducing overfitting and appropriately implemented differential privacy. These measures need testing and trade-off analysis; none should be described as a universal guarantee.

Hypothetical example: A research team trains a model on records from a specialist oncology cohort. It assesses whether an external user with some prior information could infer that a person contributed to training. The team tests realistic queries before release and records which controls reduce the observed risk and which uncertainties remain.

Frequently Asked Questions

Is this the same as training-data extraction?

No. Membership inference asks whether a record was present, rather than necessarily recovering its contents.

Does removing names eliminate the risk?

Do not assume so. An attacker may combine other information with model responses.

Does differential privacy solve every exposure?

No. Its protection depends on design and implementation; access controls and broader security review still require attention.

Related reading: large language models, differential privacy, training-data extraction. For practical support, see AI compliance services.

Updated: 27 October 2026