An adaptive threshold-based semi-supervised learning method for cardiovascular disease detection.

Deep cardiovascular disease (CVD) detection usually achieves good performance with large-scale labeled electrocardiograms (ECGs), but manual labeling of ECGs is tedious. Semi-supervised learning (SSL) aims to improve model performance through unlabeled data. In this study, an adaptive threshold-based semi-supervised learning model (ATSS-LGP) is proposed. It introduces the multibranch network (MBN) to generate local and global predictions for 12-lead ECG. The labeled ECGs are used to train the model in supervised manner and produce adaptive thresholds through the proposed prediction voting decision mechanism. The unlabeled ECGs are divided into high-confidence and low-confidence parts by the adaptive thresholds. The pseudo-labeling technique and consistency regularization are used to jointly guide the unsupervised learning, which can fully utilize all unlabeled ECGs. To the best of our knowledge, this is the first SSL algorithm that explores adaptive threshold in ECG. ATSS-LGP shows impressive performance in CVD detection. Using the same number of labeled ECGs, it achieves at least a 5.15% increase in accuracy over pure supervised learning. Moreover, ATSS-LGP achieves comparable performance to fully supervised method using only 10% of labeled ECGs. In summary, ATSS-LGP is a suitable SSL algorithm for 12-lead ECGs, which can greatly reduce the burden of ECG labeling in deep learning. [ABSTRACT FROM AUTHOR]