MorphDiff-ECG: A Diffusion Model for Multimodal Feature Fusion of ECG Signals.

OBJECTIVE: Electrocardiogram (ECG) generation is crucial for computer-aided diagnosis and medical data augmentation, yet existing generative models struggle with waveform fidelity, morphological accuracy, and rhythm consistency. METHODS: We propose a novel diffusion model for multimodal feature fusion of ECG signals, namely MorphDiff-ECG. Inspired by the hierarchical "morphology-rhythm-trend" logic that clinicians follow when interpreting ECGs, we designed a two-stage hierarchical decoding architecture comprising two core stages. The first stage, termed Feature Decoupling Enhancement (FDE), comprises three parallel branches dedicated to modeling the trend, periodic, and morphological components of physiological signals. This explicit decoupling design not only aligns the learned features with clinically interpretable terms, but also enhances the model's sensitivity to abnormal morphological patterns. Specifically, the trend branch captures low-frequency baseline drift and global amplitude variations, ensuring stable signal horizons; the periodic branch extracts rhythmic cardiac patterns through multi-scale time-frequency analysis, preserving both regular and irregular heart rate dynamics; the morphological branch isolates fine-grained local waveform geometry, enabling precise reconstruction of diagnostically critical features such as P-waves, QRS complexes, and T-waves. This tripartite decomposition embeds the hierarchical diagnostic logic of cardiology directly into the network architecture, providing both physiological plausibility and mechanistic interpretability. The second stage, termed Cross-Feature Attention Fusion (CFAF), integrates the decoupled features through an efficient attention-based mechanism to reconstruct complete physiological signals. To further preserve critical high-frequency transients, exemplified by QRS complexes, we introduce a hybrid training objective that combines time-domain reconstruction with wavelet-based time-frequency loss. This approach takes advantage of the inherent time-frequency coupling of the ECG signals, ensuring structural integrity and spectral fidelity in the generated outputs. RESULTS: Experiments on three classic public datasets show that MorphDiff-ECG outperforms state-of-the-art methods in generation fidelity on the MIT-BIH, CPSC2020, and extended nine-class CPSC2018 benchmarks. Compared to the best-performing baseline, it reduces PRD by 1.97%, FD by 2.38%, RMSE by 10.39% and MAE drops by 0.61%. CONCLUSION: This work provides an effective solution for high-quality and controllable ECG synthesis. SIGNIFICANCE: MorphDiff-ECG advances cardiac diagnosis and personalized medical data generation through improved quality of ECG synthesis.