Cardiac echocardiographic analysis with multi-scale effective fusion module: a novel stroke prediction approach

Stroke remains a leading cause of death and disability, and early risk stratification is critical for prevention. Existing models based only on clinical factors may miss subtle cardiac structural and hemodynamic information visible on echocardiography. We aimed to develop a multimodal stroke prediction model integrating multi-view echocardiographic images and clinical indicators.

In this retrospective study, 712 hypertensive patients (10,992 echocardiographic images; 27 clinical variables) were included. Long-axis, short-axis, and apical four-chamber views were analyzed. We developed a Multi-Scale Effective Fusion (MSEF) module combining Global Feature Fusion, Multi-Feature Reconstruction, Channel Attention, and Positional Attention to improve multi-scale feature representation. Imaging features were integrated with clinical variables to build multimodal models. Model performance was evaluated on validation and test sets using accuracy, precision, recall, and F1 score.

The MSEF-based imaging model outperformed comparator fusion variants and achieved an accuracy of 76.8% and an F1 score of 64.7% on the test set. After integrating clinical indicators, performance further improved, with a test accuracy of 80.2% and an F1 score of 72.1%.

The proposed MSEF-based multimodal framework improves stroke risk prediction by effectively combining echocardiographic and clinical information, and may support earlier risk identification and clinical decision-making.