Prediction of Left Ventricular Systolic Dysfunction from ICU Electrocardiograms Using Vision Transformer Embeddings...

In critically ill patients, left ventricular ejection fraction (LVEF) is a key prognostic indicator, yet its assessment via echocardiography or cardiac magnetic resonance (CMR) is frequently delayed or unavailable in the intensive care unit (ICU). This study proposes a novel multimodal approach that combines deep embeddings from a Vision Transformer (ViT) pretrained on CMR data with conventional quantitative ECG features to predict reduced LVEF (≤40%) directly from 12-lead ICU electrocardiograms. Using a retrospective cohort of approximately 900 ICU admissions from a single tertiary centre, three classifiers-Random Forest, XGBoost, and a shallow multilayer perceptron (MLP) trained with focal loss-were compared under stratified 5-fold cross-validation. The MLP achieved the best performance (AUC = 0.87, recall = 0.68, F1 = 0.65), demonstrating that cross-modal ViT embeddings transfer meaningful cardiac-structural information to ECG-based prediction tasks. These results support the feasibility of using routine ECGs as an early screening tool for ventricular dysfunction in the ICU.