Estimation of future occurrence of hemoglobin-A1c elevation with and without differential privacy.

backgroundPrivacy protection is crucial for legal compliance and ethical practices, particularly in medical data sharing. Maintaining trust is essential for effective healthcare delivery, preventing discrimination, and protecting patient information. Glycated hemoglobin is an essential parameter for diagnosing diabetes, and its analysis is sometimes necessary for decision-making. Preventing the leakage of this valuable information is a crucial consideration.

methodsLocal differential privacy (LDP) is one of the best methods to address this problem. In this study, we applied four LDP algorithms (Logistic regression, tree-based methods, and Gaussian naïve Bayes) to determine an appropriate degree of privacy budget for data perturbation demands. We also applied two dimensionality reduction techniques, principal component analysis and discriminant component analysis (DCA), with Gaussian naïve Bayes to observe their effects on the results. Theoretically, a higher value of e enhances usefulness but decreases privacy protection.

resultsThe logistic regression model achieved an accuracy of > 70% for the smallest e values, outperforming the other algorithms. In the dimensionality reduction techniques, the area under the receiver operating characteristic curve of DCA was > 80% with LDP when the value of 𝜀 was < 0.1.

conclusionsThis study provides important insights into data privacy for researchers, practitioners, and policy makers, particularly those who handle sensitive medical data.