High-Fidelity Synthetic ECG Generation via Mel-Spectrogram Informed Diffusion Training

TL;DR

This work tackles privacy-driven barriers to ECG data by proposing MIDT-ECG, a Mel-Spectrogram Informed Diffusion Training method built on $SSSD-ECG$ that enhances morphological fidelity while enabling patient-specific synthesis through disentangled multimodal conditioning. By supervising diffusion in the time-frequency domain and prioritizing diagnostically relevant spectral content, MIDT-ECG achieves substantial gains in inter-lead coherence and reduces privacy leakage compared with baselines. The authors also establish a comprehensive, multi-faceted benchmarking framework that includes fidelity, trustworthiness, clinical alignment, and privacy metrics, as well as faithfulness-based evaluation to guide dataset curation. In data-scarce settings, synthetic data generated by MIDT-ECG can serve as a viable privacy-preserving surrogate for real data, accelerating model development and collaborative healthcare research while highlighting areas for future formal privacy guarantees and cross-dataset validation.

Abstract

The development of machine learning for cardiac care is severely hampered by privacy restrictions on sharing real patient electrocardiogram (ECG) data. Although generative AI offers a promising solution, the real-world use of existing model-synthesized ECGs is limited by persistent gaps in trustworthiness and clinical utility. In this work, we address two major shortcomings of current generative ECG methods: insufficient morphological fidelity and the inability to generate personalized, patient-specific physiological signals. To address these gaps, we build on a conditional diffusion-based Structured State Space Model (SSSD-ECG) with two principled innovations: (1) MIDT-ECG (Mel-Spectrogram Informed Diffusion Training), a novel training paradigm with time-frequency domain supervision to enforce physiological structural realism, and (2) multi-modal demographic conditioning to enable patient-specific synthesis. We comprehensively evaluate our approach on the PTB-XL dataset, assessing the synthesized ECG signals on fidelity, clinical coherence, privacy preservation, and downstream task utility. MIDT-ECG achieves substantial gains: it improves morphological coherence, preserves strong privacy guarantees with all metrics evaluated exceeding the baseline by 4-8%, and notably reduces the interlead correlation error by an average of 74%, while demographic conditioning enhances signal-to-noise ratio and personalization. In critical low-data regimes, a classifier trained on datasets supplemented with our synthetic ECGs achieves performance comparable to a classifier trained solely on real data. Together, we demonstrate that ECG synthesizers, trained with the proposed time-frequency structural regularization scheme, can serve as personalized, high-fidelity, privacy-preserving surrogates when real data are scarce, advancing the responsible use of generative AI in healthcare.

Figures (10)

• Figure 1: Mel-Spectrogram Informed Diffusion Training Overview. This diagram illustrate how MIDT works in on single MIDT training Step. Signal illustrations of $X_0, \epsilon, X_t, \epsilon_\theta, \hat{X_0}$ on the right grey box are generated by our trained model.
• Figure 2: Comparison of inter-lead correlation matrices for real and synthetic ECG data.
• Figure 3: Comparison of correlation difference matrices for the baseline and proposed models. Darker colors indicate larger errors, 1 is positive relevant, -1 is negative relevant. The significantly paler colors demonstrate the superior performance of our MIDT-ECG method.
• Figure 4: Statistical and Morphological metric distribution across baseline SSSD-ECG and 4 variants: mel - mel-spectrogram loss variant, multi - Disease + All demographic conditioned variant, bmi - Disease + BMI conditioned variant, da - Disease + Age conditioned variant. Boxplots show that conditioning models achieve higher SNRs compared to baseline, but exhibit a larger number of extreme outliers in error metrics (MSE, RMSE, Hausdorff distance, Fourier Transform distance), indicating greater variability and consistent occasional failure cases.
• Figure 5: Lead distribution of outliers across conditioning variants in different metrics. Dashed red line represents the uniform distribution (evenly distributed across 12 leads). Lead 4 and 5 are observed to have consistent high frequency in outliers.