On the Effectiveness of Data Anonymization and Synthetic Data Generation for Privacy-Preserving Machine Learning in Healthcare

Machine learning on tabular medical data is used for clinical analysis and data-driven decision making. However, these datasets contain sensitive information, which raises privacy concerns. This calls for privacy-preserving procedures while keeping analytical utility. Anonymization and synthetic data generation have been explored extensively, but few direct comparisons across several machine learning tasks have been made. In this paper asses the effect of anonymized and synthetic data on classification, regression and clustering using six medical tabular datasets from Kaggle and UCI Machine Learning Repository ranging from 569 to 110,527 records. Anonymization was performed by a combination of pseudonymization, generalization, and attribute-based perturbation, while synthetic data were created by a Conditional Tabular Generative Adversarial Network (CTGAN). Results show that anonymization often preserves the machine learning value with respect to the original data, although the performance depends on the characteristics of the dataset and the analytical tasks. When classifying, anonymised data performed as well as original data (97.37% accuracy for the Breast Cancer dataset), whereas synthetic data showed a significant performance drop due of lower discriminative feature boundaries. Regression demonstrated the best predictive accuracy of anonymised data (R² up to 0.926) and synthetic data (largest reduction to 0.431) suggesting little preservation of complicated numerical relationships and extreme values. Anonymized data preserved data structure better, but synthetic data changed cluster distributions and reduced inherent variability.In general, anonymization provides a better balance between privacy and utility when it comes to predictive analytics in healthcare, while synthetic data is more ideal for exploratory and secondary analysis.