Original title: DiSR-ECG: Residual Shifting Conditional Diffusion for Robust ECG Super-Resolution
Fan-Yi Hsu, I Chiu, Kuan-Chen Wang, Sheng-Yu Peng +1 · September 2026
Wearable devices track heartbeats over long periods, but saving battery often forces them to record lower-quality heart signals, known as ECGs. Doctors need sharp, high-resolution signals to spot irregular heart rhythms. Existing artificial intelligence models can sharpen these signals, but they often struggle when real-world data looks different from their training data. To solve this, researchers built DiSR-ECG. It uses a diffusion process, which gradually builds clean data, combined with a technique called residual shiftingA method that focuses on predicting the difference between the low-resolution and high-resolution signals, rather than the full signal. and a fast sequence-tracking tool named MambaA type of neural network designed to efficiently process sequences of data, like time series, by capturing long-range patterns.. These parts guide the model to reconstruct complex heart signals recorded from multiple angles. When tested, DiSR-ECG reached top performance both on standard data and on new, unfamiliar data, proving it can reliably clean up heart recordings.
Imagine trying to sketch a detailed portrait from a blurry photo. You can fill in some details, but if the photo was taken under different lighting than what you are used to, your guesses might be wrong. Here, the blurry photo is the low-rate ECG signal, and the portrait is the high-resolution ECG. The model learns to fill in the missing details, and it does so even when the blurry photo comes from a different camera or lighting condition than what it trained on.
Wearable ECG devices save battery by recording at low rates, but the resulting signals can miss crucial details for diagnosing heart arrhythmias. Existing super-resolution methods work well on the data they were trained on but often fail when the data comes from a different device or patient population. This model maintains accuracy across such shifts, making it practical for real-world monitoring where conditions vary.
The abstract does not report inference speed or computational cost, which is important for real-time monitoring. It also only tests on two datasets, so generalization to other devices or patient populations is not fully established. The model's performance on rare arrhythmias or noisy recordings is not discussed. Clinical validation in a real setting is still needed.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
DiSR-ECG: Residual Shifting Conditional Diffusion for Robust ECG Super-Resolution — Fan-Yi Hsu, I Chiu, Kuan-Chen Wang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.