Abstract / Summary
Abstract Risk of sudden cardiac arrest may be expressed differently by anatomy, mechanical function and electrical activity, while studies that contain all three modalities for the same patient are uncommon. We developed CardiacGuard-Net, a modular deep-learning framework that learns modality-specific representations from cardiac magnetic resonance imaging (CMR), echocardiography and twelve-lead electrocardiography (ECG), then combines them for cardiac-arrest classification. The modality-specific models achieved mean Dice 0.953 for CMR segmentation, 3.8% mean absolute error for echocardiographic ejection-fraction estimation and AUC 0.938 for 24-h arrest prediction from ECG. For multimodal evaluation, 1,204 records were assembled into a matched benchmark using age band, sex, ejection-fraction band and diagnostic category because no dataset available to the study contained all three modalities for the same patients. On this benchmark, CardiacFusion-Net achieved AUC 0.947, compared with 0.920 for late fusion. Removing the ECG, CMR or echocardiographic branch reduced AUC by 0.035, 0.026 and 0.016, respectively. A single trained model also retained discrimination when modalities were withheld, with AUCs of 0.890 for ECG alone, 0.905 for MRI+echo, 0.921 for MRI + ECG and 0.931 for echo + ECG. The multimodal results are therefore best interpreted as an internal proof-of-concept benchmark of complementary information, not as patient-level clinical validation.