Sparse Temporal AutoEncoder for ECG Anomaly Detection

Electrocardiogram (ECG) analysis is a fundamental tool for diagnosing various cardiac conditions; however, accurately distinguishing between normal and abnormal ECG signals remains challenging due to high inter-individual variability and the inherent complexity of ECG waveforms. In this study, We propose a novel Sparse Temporal Autoencoder (STAE) for unsupervised ECG anomaly detection that leverages Temporal Convolutional Networks (TCNs) to extract hierarchical features from both time-domain and frequency-domain representations of ECG signals. Unlike traditional approaches requiring annotated abnormal samples, the proposed model is trained exclusively on normal ECG data, making it well-suited for real-world deployment. A STAE integrates a masked signal reconstruction strategy and a hybrid sparse attention mechanism combining sparse block and sparse strided attention to capture critical temporal and spectral patterns efficiently. The proposed method is evaluated on the PTB-XL dataset, where it achieves the highest ROC-AUC of 0.872 among compared unsupervised methods while maintaining a low inference time of 0.009 s, demonstrating that STAE achieves state-of-the-art performance in ECG anomaly detection, highlighting its potential as a powerful tool for automated and intelligent ECG analysis.