Evaluation of Avatarization for Privacy-Preserving Synthetic Health Data Generation: A Case Study in Cancer...

Area of Science:

• Health informatics
• Data privacy
• Synthetic data generation

Background:

• Generating synthetic health data is crucial for research while protecting patient privacy.
• The Avatar method offers a potential solution for creating realistic yet private datasets.
• Evaluating the utility-privacy trade-off of synthetic data generation methods is an ongoing challenge.

Purpose of the Study:

• To evaluate the Avatar method for synthetic health data generation.
• To assess the balance between data utility and privacy preservation using the Avatar method.
• To identify critical parameters influencing the utility-privacy trade-off in Avatarization.

Main Methods:

• Utilized a cancer prediction dataset comprising 1,500 patients.
• Analyzed the Avatar method across various parameter settings.
• Quantified data utility using metrics like Hellinger distance and privacy using re-identification and hidden rates.

Main Results:

• Avatar method achieved high data utility (97.96%) and approximated the original data's statistical structure.
• Demonstrated significant privacy protection with a privacy rate of 91.7% and hidden rate of 92.4%.
• Identified neighborhood size (k) as a critical parameter for optimizing the utility-privacy balance.

Conclusions:

• The Avatar method shows promise for generating private synthetic health data.
• Achieving an optimal utility-privacy balance requires careful parameter selection and tuning.
• Further validation is needed before widespread application in real clinical contexts.