Echo-SMADS: A hierarchical planning model for predicting ejection fraction using echocardiography

Current deep learning approaches for predicting ejection fraction primarily rely on end-to-end regression. While effective in certain cases, these methods often lack intermediate structural foundations and limit the interpretability of the clinical decision-making process. This limitation reduces transparency and may hinder clinical adoption. The objective of this study is to design a clinically aligned, modular system that emulates the diagnostic workflow of physicians, improving interpretability, stability, and applicability in real-world ejection fraction assessment.

We propose Echo-SMADS, a clinically aligned system that adopts the concept of hierarchical planning from artificial intelligence. The overall prediction task is decomposed into three clinically relevant subtasks: structure identification, phase selection, and volume estimation. These subtasks are implemented as decoupled functional modules, each optimized independently while producing interpretable intermediate outputs. The modular design allows reasoning evidence to propagate across stages, enabling a transparent diagnostic reasoning path.

Experiments were conducted on the EchoNet-Dynamic dataset. Echo-SMADS achieved a mean absolute error of 5.48 ± 0.17 and a root mean square error of 7.64 ± 0.20 for ejection fraction prediction. The results demonstrate improved performance stability compared with traditional end-to-end models, while providing meaningful intermediate outputs that enhance interpretability and trustworthiness.

Echo-SMADS integrates modular, clinically aligned components into a structured diagnostic process that closely reflects real-world clinical workflows. By combining interpretability, physical grounding, and performance stability, the proposed system offers a promising approach for reliable ejection fraction prediction, with potential for future clinical application.