Memory-Augmented Autoencoding with Self-Supervised Learning for Unsupervised Detection of Abnormal Physiological...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Accurate detection of abnormal events in physiological signals (EEG, ECG) is crucial for diagnosing neurological and cardiovascular disorders.
• Current unsupervised anomaly detection methods struggle with limited representation capacity and poor generalization across signal types.

Purpose of the Study:

• To develop a novel unsupervised anomaly detection framework, MAGE, to overcome limitations of existing methods.
• To improve the detection accuracy and robustness of abnormal event identification in physiological signals.

Main Methods:

• MAGE utilizes a unified convolutional autoencoder architecture integrating multi-head memory gating, self-supervised learning, and adversarial training.
• A memory-augmented gating mechanism adaptively preserves and integrates salient features for enhanced discriminability.
• Self-supervised tasks and transformation-aware adversarial perturbations improve representation learning and robustness against distribution shifts.

Main Results:

• MAGE achieved over 98% detection accuracy and superior F1-scores on benchmark EEG and ECG datasets.
• The framework consistently outperformed state-of-the-art baselines in within-dataset evaluation settings.
• Demonstrated significant improvements in feature discriminability and robustness compared to prior memory-based approaches.

Conclusions:

• MAGE offers an effective and robust solution for unsupervised anomaly detection in physiological signals.
• The framework shows significant clinical potential for early anomaly detection and continuous health monitoring.
• MAGE's approach enhances representation learning and generalization capabilities for diverse signal domains.