TLDR: A research paper examines if synthetic data, used for AI in healthcare, falls under EU’s GDPR. Through an empirical study and legal analysis, it concludes that well-generated synthetic data is likely anonymous due to high identification effort and low attack certainty. The paper advocates for clearer regulations to balance privacy with the societal benefits of AI development in healthcare.
Artificial intelligence (AI) holds immense promise for transforming healthcare, but its development heavily relies on access to vast amounts of health data. However, the strict interpretations of the EU’s General Data Protection Regulation (GDPR) have created significant hurdles for sharing this sensitive information, slowing down crucial AI-driven research and treatment advancements. This challenge has led to a growing interest in synthetic data, which is artificially generated data that mimics the statistical properties of real data without containing actual personal information.
A recent research paper, “Processing of synthetic data in AI development for healthcare and the definition of personal data in EU law,” explores whether synthetic data should be classified as personal data under the GDPR. The authors, Vibeke Binz Vallevik, Anne Kjersti C. Befring, Severin Elvatun, and Jan Franz Nygård, conducted a systematic analysis of relevant legal sources and an empirical study to shed light on this complex issue. You can read the full paper here.
The Challenge of Data Sharing in Healthcare AI
The core problem lies in the tension between the need for data to train powerful AI models and the imperative to protect patient privacy. While synthetic data is proposed as a solution to enable data sharing with reduced privacy risks, there’s ongoing uncertainty about its legal status under GDPR. If synthetic data is considered “personal data,” it falls under the regulation’s stringent rules, adding administrative burdens and limiting its benefits. If it’s deemed “anonymous,” it falls outside GDPR’s scope, allowing for freer use.
Empirical Study: Testing Identifiability
To understand the practical risks, the researchers conducted an empirical case study using a synthetic dataset derived from real colorectal cancer patient data from the Florence project. This dataset included 16 attributes describing diagnosis and treatment for over 30,000 patients. Direct identifiers were removed, and other variables were modified to reduce granularity. The study focused on “real data-inspired” synthetic data, which replicates statistical distributions without preserving exact individual records.
The team performed a “Membership Inference Attack” (MIA) simulation. An MIA attempts to determine if a specific individual’s data was part of the original dataset used to train the synthetic data generator. The simulation aimed to assess the “residual identification risk.” The attack assumed a realistic scenario where an attacker would have access to the synthetic dataset and publicly available additional information, such as details patients might share on social media (e.g., approximate diagnosis dates, age, gender). It did not assume access to the original, sensitive patient records.
The results showed a relatively low identification success rate for the MIA, with an average ROC AUC score of 0.75 in the “Upper Bounds” scenario (where the attacker has access to all possible parameters without the original training data). In the “Likely Scenario” (with limited, realistic auxiliary information), no records had a score higher than 0.75. An AUC score of 0.5 indicates random guessing, while 1.0 is perfect identification. The study highlighted that the attack’s effectiveness was significantly influenced by the target record’s uniqueness and that the attacker would not be able to verify their guesses without access to the original data, which significantly lowers the motivation for such a resource-intensive attack.
The empirical study concluded that the effort required for a successful attack was substantial (weeks of work, high data science competence, high-performance computing), and the resulting information was highly uncertain. This suggests that, in this specific case, the synthetic data should most likely be considered anonymous, as identification is “reasonably unlikely.”
Legal Analysis: Interpreting GDPR and Case Law
The paper delves into the GDPR’s definition of “personal data” (Article 4(1)) and the concept of “identifiability” (Recital 26). The GDPR states that data is personal if it relates to an “identified or identifiable natural person,” considering “all the means reasonably likely to be used” to identify them. This phrasing emphasizes a risk-based assessment, not just a theoretical possibility.
Key EU court cases were analyzed:
- Breyer case: Established that identification is not “reasonably likely” if it’s prohibited by law or “practically impossible” due to disproportionate effort (time, cost, effort). It also clarified that not all identifying information needs to be held by one person.
- Banco Popular case: Suggested that pseudonymized data might not be personal data for a recipient who lacks the means to re-identify individuals, even if the original data controller retains an identification key. This supports the idea of a dynamic definition of personal data based on the recipient’s context.
- OC v Commission: Showed that if additional information can be easily found (e.g., via an internet search) and there’s sufficient motivation (e.g., public interest in fraud), identification can be considered “reasonably likely.”
The paper argues that the mere existence of the original dataset should not automatically preclude synthetic data from being considered anonymous, especially if the recipient of the synthetic data lacks the means to access that original data. The “risk-based approach” from case law contrasts with some earlier, more conservative guidelines from bodies like WP29, which suggested “irreversible” anonymization was required.
Human Rights and Proportionality
The research emphasizes that GDPR’s primary goal is to safeguard fundamental rights, but data protection is not an absolute right. It must be balanced against other fundamental rights and societal needs, such as the right to health, the right to benefit from scientific progress, and the principle of non-discrimination. The proportionality principle, central to EU law, allows for weighing the risks of privacy infringement against the benefits of data sharing, especially when it leads to advancements in healthcare, such as developing tools for rare diseases or addressing biases in datasets.
Also Read:
- Crafting Synthetic Medical Data: How Smart Prompts Guide AI for Privacy and Quality
- Advancing Causal Machine Learning Through Better Synthetic Experiments
Policy Implications and the Way Forward
The paper concludes that well-generated synthetic datasets, like the one in the empirical study, should rarely be considered personal data under GDPR, given the low motivation and high effort required for identification, coupled with the uncertainty of attack outcomes. This perspective is supported by the dynamic interpretation of personal data in EU case law and the broader human rights framework.
The current legal ambiguity, however, leads to overly cautious interpretations and hinders the potential of synthetic data in healthcare AI development. The authors call for clearer regulations or guidance that define what constitutes “reasonably likely” identification, potentially including penalties for unauthorized identification of synthetic datasets. Such clarifications are crucial to unlock the full potential of synthetic data, foster innovation within the EU, and ensure that healthcare benefits from advanced AI technologies without compromising fundamental rights.
