Sustainable and Responsible ECG-Based AI Diagnostics: Masked Frequency Reconstruction with Peak-Aware Transformers

Electrocardiogram (ECG)-based deep learning systems play an increasingly important role in scalable and accessible cardiac health diagnostics, yet their effectiveness is often constrained by limited labeled data, data imbalance, and deployment in resource-restricted or vulnerable populations. Self-supervised learning (SSL) offers a path toward more equitable and sustainable medical AI, but existing ECG SSL approaches often overlook crucial frequency-domain information and clinically important waveform peaks. In this work, we introduce Electrocardiogram Masked Frequency Reconstruction with Peak-Aware Transformer (EMFR-PAT), a responsible and energy-efficient SSL framework that learns high-quality ECG representations by reconstructing masked frequency components and explicitly modeling peak-related features such as R-peaks. To improve robustness across diverse real-world and web-based ECG collection settings, we further incorporate a phase alignment module to mitigate the impact of temporal shifts. Experiments on two public datasets demonstrate that EMFR-PAT significantly outperforms existing SSL baselines in fine-tuning and linear probing, highlighting its ability to generalize even in low-label or data-poor environments. By advancing the reliability, interpretability, and efficiency of ECG representation learning, this work contributes to the development of web-enabled health technologies that support global health equity, responsible AI adoption, and scalable clinical decision support.