Disentangled conditional generation of ECG signals with interpretable latent factors.

BACKGROUND AND

OBJECTIVE

Electrocardiogram modeling supports signal synthesis, analysis, and data augmentation. Existing generative approaches primarily focus on waveform fidelity and diversity, with limited interpretability of the underlying generation mechanisms, whereas disentangled representation learning improves interpretability through semantically structured latent factors but has been less explored from a generative perspective.

METHODS

We propose an electrocardiogram modeling framework that integrates disentangled representation learning with conditional generative modeling to support interpretable and controllable ECG synthesis. Disentangled latent representations are learned using a β-Total Correlation Variational Autoencoder, and a self-attention mechanism is incorporated to promote semantic alignment between latent dimensions and signal morphological characteristics. A conditional flow module is further introduced in the latent space to enable condition-aware prior reshaping, mitigating distributional collapse while preserving the learned disentangled structure.

RESULTS

Experimental results demonstrate that the proposed framework mitigates the generation collapse induced by strong disentanglement learning. While preserving a morphology-oriented latent structure, as indicated by a proxy-based mutual information gap score of 0.53, the conditional flow module improves generative performance. In an out-of-domain evaluation where only the flow is fine-tuned on the target dataset, the framework achieves a fidelity of 0.66 and diversity of 0.80, compared with 0.33 and 0.32 under standard prior sampling without flow adaptation.

CONCLUSIONS

In summary, the proposed framework supports interpretable and controllable electrocardiogram generation through disentangled latent representations and condition-aware distribution alignment. The results demonstrate partial semantic consistency under out-of-domain evaluation and show that latent-space flow adaptation can mitigate generation collapse to some extent under strong disentanglement constraints.