Privacy Protection in Data Synthesis: Effects on Survival Analysis Performance

Area of Science:

• Medical Informatics
• Data Science
• Biostatistics

Background:

• Survival analysis is crucial for predicting patient outcomes.
• Synthesizing data with differential privacy offers enhanced patient confidentiality.
• Balancing privacy and accuracy in synthetic data for survival analysis remains a challenge.

Purpose of the Study:

• To investigate the impact of differential privacy on synthetic data accuracy for survival analysis.
• To evaluate the trade-off between privacy protection and predictive performance.
• To identify optimal methods for integrating differential privacy in survival data synthesis.

Main Methods:

• Synthesized a German lung cancer patient dataset using CTAB-GAN+.
• Applied CoxPH and DeepSurv models for survival analysis.
• Utilized Miss Forest for imputation and various encoding techniques for categorical variables.

Main Results:

• Differential privacy budgets significantly influenced model accuracy.
• Model choice and data preprocessing improved accuracy by up to 4.5%.
• The CoxPH model with Miss Forest imputation and one-hot encoding achieved a concordance index exceeding 0.68 under differential privacy.

Conclusions:

• Differential privacy integration is feasible for survival analysis data synthesis.
• Methodological choices in imputation and encoding are critical for maintaining accuracy.
• Achieving a balance between robust privacy and reliable predictive power is attainable.