Predicting Cardiac Arrest Using Ultrasound and MRI Images: A Multimodal Machine Learning Approach on Left Ventricular Myocardium Parameters
P. Katheeja Khanam, V. Anantha NatarajanPurpose
Sudden Cardiac Arrest (SCA) is a life-threatening cardiovascular emergency, with survival rates below 10% without immediate intervention. Accurate prediction requires comprehensive characterization of left ventricular (LV) myocardial structure and function, which cannot be fully captured by a single imaging modality. This study proposes MuMo-CardNet, a multimodal deep learning framework for SCA risk prediction using complementary echocardiography and cardiac magnetic resonance imaging (MRI).
Design/methodology/approach
The proposed five-stage pipeline integrates 2D echocardiography from the CAMUS dataset and short-axis cine MRI from the ACDC dataset. The framework performs image preprocessing, deep feature extraction using transfer-learned ResNet-50 encoders, dimensionality reduction through Recursive Feature Elimination (RFE), and cross-modal fusion using a Variational Autoencoder with Cross-Attention Transformer (VAE-CAT). XGBoost is subsequently employed for classification. A harmonization module projects heterogeneous imaging features into a common representational space.
Findings
MuMo-CardNet achieved a 93.7% reduction in inter-modality Kullback–Leibler divergence, indicating effective harmonization of heterogeneous imaging characteristics. On a merged test dataset of 271 subjects, the model achieved an AUC-ROC of 0.967, accuracy of 96.5%, and F1-score of 0.945, outperforming single-modality and simple feature-concatenation approaches. Based on the findings, it is clear that MuMo-CardNet is capable of delivering reliable and explainable risk stratification for cardiac diseases using echocardiography and cardiac MRI features. This is because of the discriminatory and sensitive nature of MuMo-CardNet that makes it suitable for use in early identification of cardiac risks.
Originality/value
SHAP-based interpretation identified Ejection Fraction, End Diastolic Volume, and Wall Motion Score as the most influential predictors, providing clinically consistent feature-level explanations. The framework demonstrates the potential of multimodal imaging fusion for accurate and interpretable SCA risk prediction.