Privacy-Preserving Synthetic Data Generation in Healthcare

Abstract

The advancement of Artificial Intelligence (AI) is having a substantial influence on the healthcare sector. Nonetheless, a principal obstacle to implementing this technology is acquiring high-quality data. At the same time, legal limits arising from privacy and safety concerns have created challenges, positioning synthetic data (SD) as a viable option. The notion of SD production offers an alternative that enables data sharing and utilisation in ways unattainable with real-world data. SD, by maintaining the integrity of genuine data distribution, safeguards sensitive information and facilitates the training of machine learning (ML) classifiers without compromising privacy. This research assesses the efficacy of SD generated via a hard-coded SMOTE technique, focusing on creating SD samples for three healthcare-related datasets (breast cancer, heart disease, and prostate cancer). We specifically assessed the initial ML outcomes on the original datasets. Thereafter, we analysed them alongside those generated by the proposed technique to determine similarities and variations. Experimental results indicate that the proposed technique achieves performance comparable to the original dataset across most ML classifiers, with some classifiers performing better. This proposed approach guarantees data integrity and privacy by maintaining the confidentiality of the original datasets. In contrast, the SMOTE-generated data acts as an effective proxy, keeping the representational correctness of the original data. The Decision Tree classifier demonstrated the most significant performance improvement on the prostate cancer dataset, achieving +5.00, +8.34, +16.66, and +8.18 improvements in Acc., Pre., Rec., and F1 respectively, compared to its original performance.