Lead-Level ECG Signal Quality Assessment Based on Weak Supervision and Multi-Branch Fusion

12-lead Electrocardiogram (ECG) is a common test for diagnosis of cardiovascular disease and each ECG lead reflects the heart activities from different aspects. Due to the environmental noise interference in daily life, the assessment of ECG signal quality is of great importance for accurate diagnosis. However, due to the lack of lead-level quality labels, existing methods only perform coarser record-level quality assessment, which may reject lead of good quality or accept poor-quality lead. To overcome this limitation, the paper proposes a multi-branch fusion model for lead-level quality assessment under weak supervision. In this framework, each lead corresponds to a multi-scale convolutional network which produces a lead-level score, and a dense gating network assigns a weight value to each individual lead. The record-level quality score of the multi-lead ECG signals is obtained by aggregating the lead-level quality scores. The proposed method was evaluated on the PhysioNet/CinC2011 dataset and achieved an accuracy of 95.5%, a sensitivity of 98.07%, and an F1 score of 97.13%. The results demonstrate that the proposed method can effectively assess the quality of 12-lead ECG signal at both record and lead levels.