Benchmarking privacy and utility in synthetic tabular cohorts for Alzheimer's disease research

The scarcity of large, clinically relevant cohorts is becoming a bottleneck in Alzheimer's disease (AD) research, as their sensitive nature makes open data sharing difficult. Privacy-preserving synthetic datasets generated with machine learning may help address this challenge.

We compared five frameworks for generating synthetic tabular data from the Alzheimer's Disease Neuroimaging Initiative and Anti-Amyloid Treatment in Asymptomatic Alzheimer's Disease cohorts, with a set of empirical privacy and utility metrics. Two of the methods, DataSynthesizer and TableDiffusion, provide $\epsilon$ -differential privacy guarantees.

Methods with differential privacy achieved high privacy ratings but low levels of utility. Deep learning methods like Tabular Prior-data Fitted Network (TabPFN) and Conditional Generative Adversarial Network (CTGAN) also showed high privacy with limited utility. In contrast, non-private DataSynthesizer and Synthpop offered higher utility at a cost of lower privacy.

The evaluated methods demonstrated a clear trade-off between privacy and utility. High privacy was generally associated with insufficient utility, highlighting the need for further research into synthetic data generation for AD.