AI-Enhanced CNN-LSTM and MRI-CFD-Based Hemodynamic Modeling for Non-Invasive Diagnosis of Hemodynamically Significant Coronary Artery Disease in High-Risk Referred Patients

Abstract

The majority of cardiovascular deaths and disabilities globally are caused by coronary artery disease (CAD). The aim of the present study was to determine whether a deep learning (DL) fused model, consisting of a hybrid of cardiac MRI-derived morphological and perfusion features, and computational fluid dynamics (CFD)-derived hemodynamic indices, was able to non-invasively predict hemodynamically significant CAD in high-risk adults undergoing cardiac evaluation. A total of 1,180 patients participated in the study from January 2022 until December 2024 and 694 (58.8%) were confirmed to have hemodynamically significant CAD by invasive coronary angiography (ICA) with pressure wire fractional flow ratio (FFR). This CNN-LSTM model showed a significantly higher value of accuracy (AUC 0.934; 95% CI 0.914–0.954), sensitivity (88.7%), and specificity (93.4%) compared with stress echocardiography (AUC 0.821), CT coronary angiography (AUC 0.876), and SPECT myocardial perfusion imaging (AUC 0.848). The integrated multi-modal approach proved superior to imaging features and hemodynamic indices, using ablation analysis. Overall, this excellent result shows the power of the CNN-LSTM model when combined with MRI morphology, perfusion, and tissue characterization with CFD-based functional assessment to enhance the non-invasive diagnostic accuracy of hemodynamically significant CAD in high-risk referred patients, which can boost appropriately diagnostic efficiency and clinical outcomes.