Comparative evaluation of feature-based, deep learning, and foundation models for ECG-based detection of left heart dilatation and hypertrophy

Objectives Left Atrial Enlargement (LAE), Left Ventricular Dilatation (LVD), and Left Ventricular Hypertrophy (LVH) represent early manifestations of adverse remodeling. Electrocardiography (ECG) is widely used in clinical practice, but diagnostic performance for detecting structural abnormalities remains limited. This study compared four Deep Learning (DL) strategies for ECG-based prediction of left heart dilatation and hypertrophy. Methods We retrospectively analyzed 10,840 adults with paired ECG and transthoracic echocardiography within one week. Echocardiographic criteria were used as reference. Patients with non-sinus rhythm, prior thoracic or cardiac interventions were excluded. Four strategies were evaluated: (1) ECG amplitude and interval feature-based model by gradient boosting; (2) Transformer-based DL model integrating temporal and frequency; (3) Hybrid late-fusion model incorporating handcrafted ECG features into DL architecture; and (4) Pre-trained foundation model. Results For LAE prediction, the ECG foundation model achieved AUC of 0.8965 and F1 score of 0.7468, exceeding the performance of feature-based, deep learning, and hybrid models. For LVD, it achieved AUC of 0.9489 with F1 score of 0.8806, accompanied by higher sensitivity while maintaining high specificity. For LVH, the foundation model achieved AUC of 0.9574 and F1 score of 0.7330. Feature-based models tended to show higher sensitivity but lower specificity, whereas ML and hybrid models demonstrated balanced performance. Across all three targets, the foundation model consistently yielded higher discriminative performance as reflected by AUC and F1 metrics. Conclusion Fine-tuned ECG foundation models provide improved discrimination for ECG-based detection of left heart dilatation and hypertrophy and may support automated ECG screening for early structural heart disease.