MedNet: A Medical Overview-Focus Network With Uncertainty-Regularized Loss for Unsupervised ECG Anomaly Detection.

Abstract

Electrocardiogram (ECG) is a diagnostic tool used to determine whether abnormalities exist in cardiac electrical activity. Supervised ECG analysis methods are limited in identifying rare abnormalities because training data cannot encompass all abnormal patterns. In contrast, Unsupervised Anomaly Detection (UAD) methods can identify anomalies solely by learning the characteristics of normal ECG signals. However, existing UAD approaches still exhibit notable deficiencies in recognizing local anomalies and addressing reconstruction instability caused by noise interference and uncertainty fluctuations. Accordingly, we propose an unsupervised ECG anomaly detection network-MedNet (A Medical Overview-Focus Network). MedNet simulates the diagnostic workflow of ECG specialists and adopts a single-branch "global overview-local focus" architecture. This architecture first extracts global rhythm features through OverviewNet and uses them as prior information to guide FocusNet in concentrating on potential abnormal regions, thereby enhancing the model's capability to detect subtle anomalies. Meanwhile, MedNet incorporates a Context-aware Feature Reweighting (CFR) mechanism, which dynamically adjusts feature weights to highlight discriminative information and further improve the model's sensitivity to anomalous signals. In addition, we propose an Uncertainty-Weighted Regularization Loss (UWR Loss), which uses model predicted uncertainty as a regulatory factor to jointly constrain reconstruction errors and non-physiological discontinuities resulting from noise and uncertainty fluctuations, thus enhancing reconstruction stability. Extensive experiments on two benchmark ECG datasets demonstrate that MedNet achieves performance comparable to existing methods across major metrics, showcasing its effectiveness and generalization ability in unsupervised ECG anomaly detection tasks.