Harmonizing Fidelity, Utility, and Privacy in Medical Data Synthesis: A Dual Approach via Survival-Aware CVAE and XAI-Guided DDPM.
Authors
Affiliations (4)
Affiliations (4)
- Computer Engineering, Fırat University, 23119, Elazıǧ, Turkey.
- Computer Engineering, Manisa Celal Bayar University, 45140, Manisa, Turkey.
- Information Systems Engineering, Kocaeli University, 41001, Kocaeli, Turkey.
- Information Systems Engineering, Kocaeli University, 41001, Kocaeli, Turkey. [email protected].
Abstract
Strict data privacy regulations severely constrain the sharing and secondary use of clinical datasets in biomedical research. Generative artificial intelligence offers a viable solution by synthesizing realistic medical data devoid of original patient identifiers and augmenting existing cohorts. This study introduces two distinct modality-specific generative pipelines tailored for medical tabular synthesis and class-aware medical image generation that share a common design principle-the clinically decisive variable (event status or diagnostic class) is an explicit, controlled condition of the generator-and are assessed with a common fidelity-utility-privacy protocol. For tabular synthesis, a Conditional Variational Autoencoder (CVAE) was trained and benchmarked on the METABRIC, GBSG2, and WHAS500 datasets. Experimental results demonstrate that the conditional architecture effectively mitigates the event-collapse phenomenon prevalent in conventional VAE frameworks. For class-conditional medical image synthesis, an explainable AI (XAI)-guided Denoising Diffusion Probabilistic Model (DDPM) was developed and trained on chest X-ray and CT modalities using a stepwise strategy incorporating low-frequency, edge, and density-based anatomical regularizers. The stepwise XAI guidance coupled with anatomy-preserving constraints increased independent class consistency, whereas its effect on downstream classification was modality-dependent. Across five random seeds and six dataset-regime conditions, the CVAE achieved a mean SDV-style quality score of 0.931 ± 0.001 and a Train on Synthetic, Test on Real (TSTR) C-index of 0.739 ± 0.004, with no exact record matches and a mean distance-based membership-inference AUC of 0.555 ± 0.009, i.e., modestly above the 0.5 chance level. Across three seeds, the XAI-guided DDPM reached 98.1 ± 0.1% independent class consistency on chest X-ray and 86.1 ± 2.9% on COVID-CT; however, on COVID-CT its TSTR F1-score (0.639 ± 0.019) was lower than that of the unguided DDPM baseline (0.776 ± 0.017), indicating that the benefit of XAI guidance is contingent upon imaging modality and sample size. Rather than claiming universal superiority, the study contributes explicit event control with a characterized fidelity-privacy trade-off for survival tables and a systematic component analysis of when XAI guidance helps and when it harms image synthesis.