Privacy and Artificial Patients: Are Synthetic Patient Data Anonymous or Anonymized ?

DESCRIPTION OF SESSION

Generative Artificial Intelligence (AI) is commonly associated with Large Language Models that output human-like text, but it can also generate structured output like tabular data, enabling the creation of artificial patients or, more broadly, synthetic data (SD). A generative model can be trained using real personal health information, and then produces artificial patients from that learned data distribution. The most often applied use case for SD in healthcare is privacy-preserving sharing of data, particularly in secondary research. SD therefore can play an important role as a research and innovation accelerator while protecting individual privacy. Some large data custodians have already made synthetic versions of their health data publicly available such as Israel’s National Registry of Live Births or the French public health system claims and hospital dataset. Furthermore, SD is being increasingly shared as part of open science efforts to make research datasets more broadly accessible. The privacy-preserving capability of generative AI comes from there being no direct link between the synthetic data to real records (and thereby to real identities). As a result, it is often viewed as privacy preserving by design and thereby as a higher standard than more traditional anonymization techniques. The identifiability of SD is an important consideration for research ethics boards (REBs). In the Tri-Council Policy Statement 2 (TCPS 2), anonymous data is data that “never had identifiers associated with it (e.g., anonymous surveys) and risk of identification of individuals is low or very low” while anonymized data “is irrevocably stripped of direct identifiers, a code is not kept to allow future re-linkage, and risk of re-identification of individuals from remaining indirect identifiers is low or very low”. Anonymous data is exempt from the REB review process. The question arises: Is SD anonymous or anonymized ? There are instances where REBs in Canada and the USA treat research as exempt from review if it relied exclusively on synthetic data. At the same time, legal opinions on SD under European jurisdiction are accumulating that treat generative AI as analogous to other privacy-enhancing technologies, so that SD can qualify as non-identifiable data if residual risks are very low. Privacy laws do not typically explicitly refer to SD, nor do they spell out the standards to be met to lower the risk. Court cases that could create precedents are lacking, and very few (non-Canadian) data protection authorities have published SD guidelines. Therefore, the treatment of SD remains an open question. In the proposed session, we will provide an analysis of the privacy (risks) of SD and analyze its identifiability status to provide an answer to the question of whether it can and should be treated as anonymized or anonymous information in the context of REB reviews. The session will consist of presentations from the legal, research, and privacy policy perspectives. After the session attendees will understand what synthetic data is, the privacy risks in synthetic data, how these risks can be mapped to contemporary definitions of concepts of anonymous and anonymized as defined in the TCPS 2, and gaps between these definitions and modern privacy technologies such as synthetic data generation.