Deep neural networks interpreting 12-lead ECGs achieve cardiologist-level diagnostic sensitivity. However, microscopic imperceptible adversarial perturbations injected into raw biopotential waveforms can flip classifications from normal sinus rhythm to life-threatening ventricular tachycardia. This study details adversarial robustness testing frameworks under FDA cybersecurity guidance.
Applying bounded L-infinity perturbations mimicking electrical sensor noise demonstrated that unhardened models could be misled by noise amplitudes under 12 microvolts. Adversarial training restored robust classification boundaries.
Passing raw biopotential inputs through multi-resolution wavelet thresholding removes high-frequency adversarial perturbations before tensor inference, raising model robustness by over 38%.