Self-Supervised ECG Anomaly Detection Based on Time-Frequency Specific Waveform Mask Feature Fusion

The imbalance of ECG signal data and the complexity of labeling pose significant challenges for deep learning-based anomaly detection. Traditional contrastive learning approaches for ECG anomaly detection often rely on reconstruction or generation; however, normal signals that resemble abnormal ECG samples may be incorrectly clustered, leading to suboptimal performance. To address this issue, we propose an anomaly detection framework TFMAD that integrates ECG signal mask reconstruction with time-frequency contrastive learning, leveraging the correlation between time- and frequency-domain features for anomaly detection. Specifically, the proposed method incorporates an auto-encoder module, a time-frequency mask module, and a contrastive learning module to extract masked time-frequency domain features of ECG signals. The model then reconstructs the signal using time-frequency feature fusion and employs contrastive learning to structure the feature space, ensuring abnormal distributions are effectively learned. We evaluated this method on six datasets, and the results demonstrate that TFMAD outperforms nine state-of-the-art methods.