Automated cardiovascular disease classification is crucial to enabling real-time and continuous monitoring using wearable electrocardiogram (ECG) devices. However, due to the limited number of pathological class samples and difficulty in separating some hard-to-discriminate ECG classes (such as normal sinus rhythm (N) and premature supraventricular contraction (S)) due to their morphological similarity, existing deep learning models frequently fail to ensure sufficient inter-class separation in the latent space, limiting their discriminative power. The proposed deep metric learning framework with dynamic margin triplet loss (DynaECG-Net) extracts feature embeddings that maximize latent space separability between N, S, and premature ventricular contraction (V) heartbeats. Disease-specific experiments conducted on MIT-BIH ECG Arrhythmia dataset for 3-class classification with 50% data used for training, yield overall: accuracy 98.97%, sensitivity (Sen) 97.76%, and, F1-score (F1) 96.82% and classwise: N (Sen=99.47%, F1=99.35%), S (Sen=95.23%, F1=93.36%) and V (Sen=99.44%, F1=98.19%) achieving better performance than the state-of-the-art in low inter-class separability and data-scarce conditions. Additionally, a t-SNE visualization demonstrated well-separated embedding clusters for each class.