Objective To construct and compare a video prediction model (end-to-end model) based on deep regression and an automated calculation procedure for left ventricular ejection fraction (LVEF) based on multi-model stepwise, and compare the standard Simpson's method under ultrasound to evaluate their overall accuracy and consistency in cardiac function classification and LVEF error distribution. Methods Two representative LVEF calculation processes were constructed. One was end-to-end model using convolutional neural networks and spatiotemporal convolutional structures for feature extraction and regression prediction from echocardiographic videos. Another was multi-model step-by-step process method adopting a three-step process of " frame recognition + segmentation mask + geometric operation" . This included self-supervised learning for left ventricular segmentation and LVEF calculation using both Simpson's method and the area-length method. This study used the EchoNet-Dynamic dataset (for training) and the CAMUS database. External validation was performed using clinical imaging data from our hospital. Model performance was assessed using bias, mean absolute error, correlation and agreement metrics. Results On the CAMUS database, the LVEF calculated based on the area-length method of segmented models in a multi-model step-by-step process was closer to human measurement than the Simpson method. Its accuracy was slightly better (r = 0. 74 vs. r = 0. 72, MAE = 6. 54 vs. 8. 39). In external validation, the end-to-end model showed similar predictive performance in the fourchamber and two-chamber views (r = 0. 76 vs. 0. 77, MAE = 5. 46 vs. 5. 30). Compared to the Simpson method in the multi-model stepby-step process, the end-to-end model combined with the biplane means performance was more accurate (r = 0. 82 vs. 0. 59, MAE = 4. 65 vs. 9. 69) . Compared with Simpson ' s method in the multi-model stepwise pipeline, the end-to-end model combined with biplane averaging provided more accurate results ( r = 0. 82 vs. 0. 59, MAE = 4. 65 vs. 9. 69). Within the multi-model stepwise pipeline, Simpson's method underestimated LVEF more than the area-length method (bias: -6. 64 vs. -3. 85) . However, it showed better correlation ( r: 0. 59 vs. 0. 56) . Conclusions The end-to-end model constructed in this study has comparable accuracy in LVEF prediction to manual measurement using the standard Simpson method under ultrasound. Compared to excellent models with multi-step calculations, it has advantages in terms of accuracy and simplified procedures. However, clinical interpretability remains an issue that needs to be addressed.