Improving ECG Diagnosis of Left Ventricular Hypertrophy with Contrastive ECG-echocardiogram Learning

Because existing ECG criteria for left ventricular hypertrophy (LVH) have poor sensitivity at high specificity, we investigated whether deep learning (DL) could improve ECG-based LVH screening.We trained a vision transformer via supervised contrastive ECG-Echocardiogram learning (CEEL), which uses, in addition to the cross-entropy (CE) loss, a loss function similar to the language-image CLIP loss to align ECG and Echo embeddings.The model, denoted ViT-CEEL, was trained on a MIMIC-IV subset of 2,028 patients.The same ViT architecture was also trained with the CE loss alone (ViT-CE).Both models were compared to traditional ECG-LVH methods (Cornell, Sokolow-Lyon, Romhilt-Estes) on a local study (AFICIONADO), which consisted of 57 asymptomatic patients but at high risk of heart failure.On AFICIONADO, ViT-CE performs better (F 2 = 0.641, AUPRC = 0.38) than ViT-CEEL (F 2 = 0.556, AUPRC = 0.246), suggesting that CEEL does not boost performance as initially hypothesized.At similar specificity (0.941), the Cornell criterion has a better sensitivity (0.5) than ViT-CE (0.167).These results confirm previous findings that DL models based on the ECG alone achieve comparable performance to traditional ECG-LVH methods.