Uncertainty-Guided Boundary Localization Network for Weakly Supervised ECG Waveform Delineation

Jiaran Chen
University of Auckland


Abstract

Cardiac arrhythmias manifest as alterations in electrocardiogram (ECG) recordings, where diagnosis relies on both inter-beat rhythm and intra-beat electrophysiological features. Existing ECG analysis in clinical devices predominantly depends on rule-based algorithms with manually tuned parameters, which lack robustness under severe waveform distortions. Meanwhile, the development of automated artificial intelligence (AI) methods is constrained by the high cost of large-scale manual annotation. Here, we propose a weakly supervised framework for ECG waveform boundary localization without human annotations. Multiple rule-based algorithms are leveraged to generate complementary weak labels, which are integrated through joint modeling of data distribution, prior knowledge, and prediction subsets to construct a boundary distribution. Boundary uncertainty is further quantified and incorporated into a distributional clustering scheme to identify high-confidence boundary points. This enables reliable supervision for stable model training and facilitates downstream electrophysiological feature extraction. Training and evaluation on a PhysioNet dataset demonstrate that the proposed method achieves precise waveform boundary localization without manual annotations, providing a practical solution for AI-driven boundary detection on large-scale unlabeled ECG data.