Existing ElectroCardioGram (ECG) representation learning approaches treat the signal as an arbitrary time series, ignoring that ECG morphology is the surface projection of continuous cardiac electrophysiological dynamics. We propose training separate Modulated Neural Ordinary Differential Equation Variational Autoencoders (MoNODE-VAE) on segmented P-waves and QRST complexes, imposing a continuous-time inductive bias that mirrors the generative structure of cardiac electrophysiology. We argue that the Neural Ordinary Differential Equation vector field constraints the latent space during training, pushing clinically distinct classes toward well-separated regions. Evaluated on MedalCare-XL via linear probing, K-means clustering, and latent trajectory visualisation, MoNODE-VAE consistently outperforms a Gated Recurrent Unit-VAE baseline, yielding more structured and clinically coherent latent representations.