Left ventricular ejection fraction (LVEF) is a key clinical indicator of cardiac function, yet accurate prediction from echocardiography remains challenging due to speckle noise, anatomical ambiguity, and complex cardiac motion. We propose an optical Flow-enhanced Dual-Stream Network (FDS-Net) for LVEF prediction that jointly models global spatiotemporal dependencies and fine-grained local motion. FDS-Net consists of a UniFormer-based spatiotemporal stream to capture long-range cardiac-cycle context and a Mamba-based optical-flow stream to encode pixel-level left ventricular motion. An adaptive gate-driven fusion module dynamically integrates the two streams, enabling complementary use of structural-temporal semantics and motion cues. To improve the compatibility of optical flow with ultrasound, we further design a pretraining framework for the flow stream, including an Anatomy-Motion Semantic Fusion (AMSF) network that anchors motion learning to left ventricular anatomy via segmentation, and a Dynamically Regularized Enhanced Hybrid (DREH) loss that combines weakly supervised and unsupervised constraints. Experiments on EchoNet-Dynamic, Cardiac Acquisitions for Multi-structure Ultrasound Seg-mentation (CAMUS), and EchoHPPS show that FDS-Net achieves mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R2) values of 3.75/4.99/0.825, 5.58/7.56/0.680, and 4.13/5.92/0.711, respectively. Moreover, it outperforms 11 benchmark methods on all three datasets and provides a robust solution for predicting LVEF in noisy echocardiograms.

Event Extraction (EE) is a key task in information extraction, which requires high-quality annotated data that are often costly to obtain. Traditional classification-based methods suffer from low-resource scenarios due to the lack of label semantics and fine-grained annotations. While recent approaches have endeavored to address EE through a more data-efficient generative process, they often overlook event keywords, which are vital for EE. To tackle these challenges, we introduce KeyEE, a multi-prompt learning strategy that improves low-resource event extraction by Event Keywords Extraction(EKE). We suggest employing an auxiliary EKE sub-prompt and concurrently training both EE and EKE with a shared pre-trained language model. With the auxiliary sub-prompt, KeyEE learns event keywords knowledge implicitly, thereby reducing the dependence on annotated data. Furthermore, we investigate and analyze various EKE sub-prompt strategies to encourage further research in this area. Our experiments on benchmark datasets ACE2005 and ERE show that KeyEE achieves significant improvement in low-resource settings and sets new state-of-the-art results.