EFFNet: Efficient feature fusion network for left ventricular hypertrophy identification based on 12-lead electrocardiogram signals

BACKGROUND Left ventricular hypertrophy (LVH) is a common cardiovascular disorder, yet its detection from electrocardiogram (ECG) signals remains challenging because of the low sensitivity of conventional criteria. OBJECTIVE We aimed to explore a new deep learning method for effective LVH identification based on 12-lead ECG signals. METHODS We proposed a novel ECG-based efficient feature fusion network (EFFNet) for LVH classification, incorporating a novel feature fusion module that fuses morphological features extracted by a convolutional neural network (CNN) with algorithm-derived amplitude features and a mixture of expert (MoE) module. Experiments of 5-fold cross-validations were performed on the 12-lead ECG signals in the UK Biobank dataset (n = 38289), with cardiovascular magnetic resonance measurement as the reference standard. We also tested the model on an external cohort Qinghai database using echocardiography as the standard (n = 142777). In the UK Biobank and Qinghai cohort, we assessed associations between EFFNet predicted LVH and cardiovascular rhythm abnormalities. RESULTS Experimental results of cross-validations on the UK Biobank dataset showed that, on average, EFFNet achieved the area under the receiver operating characteristic curve (AUC) of 0.933, which outperformed conventional ECG-based diagnostic methods and compared deep learning methods. On the external Qinghai validation set, EFFNet achieved an AUC of 0.654. EFFNet predicted LVH was associated with atrial fibrillation, ventricular premature beats, and atrial premature beats. CONCLUSION EFFNet significantly improves ECG-based LVH detection and introduces a novel deep-learning framework for LVH risk prediction, advancing intelligent ECG-based screening and early identification of LVH.