The proposed morphology-keyed secure representation learning framework achieved 98.6% accuracy and an F1-score of 97.6 for five-class ECG arrhythmia classification.
A novel morphology-keyed secure representation learning framework provides highly accurate (98.6%) and privacy-preserving ECG arrhythmia classification.
The Electrocardiogram signal is a sensitive physiological signal and it is commonly used in arrhythmia classification systems. Despite the fact that recent learning-based methods have yielded high diagnostic reliability, most methods either do not take privacy considerations into account or use generic protection methods which are not correlated with the cardiac morphology and diminish interpretability. In this research work a morphology-keyed secure representation learning framework is introduced for the classification of the ECG arrhythmia. The proposed model is designed to guarantee privacy preservation, diagnostic accuracy, and signal recoverability. The variability of the rhythm, depolarization slope energy, and segment-level deviation features are used to obtain reproducible signal-dependent keys. This provides a lightweight protection at the latent representation rather than encrypting the raw waveforms. A self-recoverable decoding scheme is then incorporated to reconstruct diagnostic quality ECG signals to test fidelity. The proposed work is evaluated on the MIT-BIH Arrhythmia Database based on five heartbeat categories. The proposed model attained 98.6% accuracy, 98.2% as precision, 97.1% sensitivity, 98.9% as specificity and an F1-score of 97.6 along with low reconstruction error in the secured latent space. The findings reveal that morphology-aware security integrated smoothly into the representation learning of representation privacy-sensitive and interpretable ECG analysis in the biomedical signal processing applications.
Sumathi et al. (Sun,) conducted a other in ECG Arrhythmia. MorphoKey-SRNet vs. Conventional Wavelet-SVM, autoencoder, CNN, GRU, and LSTM-based models was evaluated on Classification accuracy. The proposed morphology-keyed secure representation learning framework achieved 98.6% accuracy and an F1-score of 97.6 for five-class ECG arrhythmia classification.