REAN: Reconstruction-aware ECG Anonymization Based on Privacy–Utility Orthogonality

Electrocardiograms (ECGs) are essential for diagnosing cardiac conditions, yet they function as biometric fingerprints that can reveal a patient's identity, age, and gender. Existing anonymization methods often force a trade-off: they either degrade the diagnostic quality of the signal or fail to sufficiently hide biometric identifiers. This paper investigates whether biometric information can be removed from raw ECG waveforms without damaging the morphology required for clinical diagnosis.

The authors observe that the directions in signal space that correspond to diagnostic utility (e.g., arrhythmia detection) and privacy (e.g., identity, age, gender) are nearly orthogonal—meeting at an angle of approximately 90 degrees. They propose REAN (REconstruction-aware ECG ANonymizer), a 1-D U-Net that exploits this geometric property. During training, REAN uses frozen diagnostic and biometric classifiers to optimize a single objective function: it minimizes diagnostic loss and distortion while maximizing privacy loss. This allows the model to learn a residual perturbation that effectively shifts the signal along the privacy-suppression direction while leaving the diagnostic features intact.

REAN significantly outperforms existing anonymization baselines across four public PhysioNet databases. It reduces re-identification accuracy to chance (0.00) while maintaining arrhythmia macro-AUROC scores statistically indistinguishable from clean, non-anonymized data (0.9991 vs. 0.9982). Furthermore, the model is robust against unseen attacker architectures and standard signal-denoising purification techniques. Because REAN is an amortized model—meaning it learns a transformation that can be applied in a single forward pass—it is over 100 times faster than per-sample optimization methods like REACT.

This work provides a practical, high-performance solution for sharing sensitive medical data. By identifying the geometric orthogonality between clinical utility and biometric leakage, the authors provide a principled way to escape the traditional privacy-utility trade-off. This enables researchers to share large-scale ECG datasets for diagnostic research without compromising patient privacy, while the single-pass inference makes it suitable for real-time clinical applications.