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

Zhuoyi Huang, Nutan Sahoo, Anamika Kumari, Girish Kumar, Kexuan Cai, Shixing Cao, Yue Kang, Tian Xia, Somya Chatterjee, Nicholas Hausman, Aidan Jay, Eric S. Rosenthal, Soundararajan Srinivasan, Sadid A. Hasan, Alex Fedorov, Sulaiman Vesal

Keywords: Structure State Space Model, ECGs, Mel-Spectrogram, Clinical Time Series, Cardiovascular Medicine, Waveforms, Low data regime, Diffusion Models, Privacy Preserving, Cardiac care, Generative AI in healthcare, Synthetic Clinical Data, Demographic Condition Generation

TL;DR: MIST-ECG uses time–frequency supervised diffusion to generate high fidelity personalized synthetic ECGs reducing interlead correlation error by 74%, boosting metrics by 4–8% with real-level performance in low-data regimes while preserving privacy.

Abstract: Generative AI holds promise for overcoming data access restrictions that hinder cardiovascular research, but the real-world use of synthetic ECG models is limited by persistent gaps intrustworthiness and clinical utility. In this work, we address two key shortcomings of current methods: insufficient morphological fidelity and the inability to generate personalized, patient-specific signals. To close these gaps, we build on a state-of-the-art diffusion model with two principled innovations: (1) MIDT-ECG (Mel-Spectrogram Informed Diffusion Training), a frequency-domain supervision framework that enforces physiological realism, and (2) multimodal demographic conditioning to enable patient-specific synthesis. We comprehensively evaluate our approach on the PTB-XL dataset, measuring signal fidelity, clinical coherence, privacy risk, and downstream utility. Our results show that MIDT-ECG improves morphological coherence, reducing inter-lead correlation error by 70%, while demographic conditioning enhances signal-to-noise ratio and personalization. Critically, classifiers trained solely on synthetic ECGs achieve performance comparable to those trained on real data in low-data regimes. Together, these contributions deliver a scalable and privacy-preserving approach for generating trustworthy, personalized medical time series, advancing the responsible use of generative AI in healthcare.

Submission Number: 168