Cross-Task Joint Network for Detecting Heart Failure with Reduced Ejection Fraction and Estimating Ejection Fraction from Electrocardiograms

Heart failure with reduced ejection fraction (HFrEF) is an advanced-stage manifestation of various heart diseases and seriously threatens human health. Left ventricular ejection fraction (LVEF) is a key clinical indicator for diagnosing HFrEF and measuring its severity. Existing deep learning methods can quickly detect HFrEF from electrocardiograms (ECGs), but cannot simultaneously provide clinicians or patients with specific LVEF values to assess severity. In this paper, we developed a novel cross-task joint network (CTJNet) using multi-channel ECGs, which could not only detect HFrEF and estimate specific LVEF values to measure its severity, but also effectively utilize the correlations between the above two tasks to improve performance. Specifically, we first designed a two-stage architecture for the CTJNet consisting of a channel-specific stage and a task-specific stage. In the channel-specific stage, we proposed a cross-task dualconstraint mechanism to enable the feature representations extracted from each channel ECG to have good generality across different tasks. For the task-specific stage, we developed a crosstask interaction mechanism to better share information across tasks by learning inter-task correlations. Experimental results showed that compared with the existing state-of-the-art methods, the CTJNet improved the average accuracy of HFrEF detection by at least 4.05% and reduced the mean absolute error of LVEF estimation by at least 9.85%. Meanwhile, the CTJNet also outperformed single-task learning methods, indicating that it could effectively utilize inter-task correlations. These results demonstrated the good potential of our method for clinical applications in HFrEF detection and LVEF estimation.