Hybrid Differentially Private Conditional Diffusion for Synthetic EHR Generation in Medical IoT Environments

Abstract

Privacy-preserving synthesis of electronic health records is difficult because clinical data combine continuous measurements, categorical diagnoses, class imbalance, missingness, and multiple trust boundaries. PrivMedSynth is presented as a conditional diffusion framework for mixed type tabular health records in Medical Internet of Things edge cloud environments. The framework applies a local randomizer before transmission and trains a conditional diffusion model with differentially private stochastic gradient descent in the cloud. Continuous features are perturbed with a calibrated Gaussian mechanism, categorical variables are protected with generalized randomized response, and clinical conditions are used as explicit generation controls. Rényi differential privacy accounts for repeated cloud updates and is converted to an approximate differential privacy guarantee. The revised formulation distinguishes the formal privacy guarantee from empirical membership inference performance and makes the privacy unit, adjacency relation, modality budget, and noise parameterization explicit. The experiments report a total variation distance of 0.041, a downstream AUROC of 0.884, and a membership inference AUROC of 0.503 on MIMIC III. These utility and attack results remain conditional on the stated preprocessing and require reproduction under the corrected accounting before being interpreted as a final end to end guarantee. The framework provides a principled basis for studying the privacy and utility tradeoff in distributed synthetic clinical data generation.