Dual-branch AI network cleans wearable ECG signals while protecting diagnostic heart waves

Wearable electrocardiogram monitors promise continuous cardiac surveillance outside the hospital, but the motion of everyday life constantly fights back. Walking, coughing, laughing, and even talking inject artifacts into the faint electrical traces that clinicians rely on, and the delicate P and T waves that carry diagnostic information are often the first casualties. A research team from the Hefei Institutes of Physical Science at the Chinese Academy of Sciences and the University of Science and Technology of China now reports a denoising framework built specifically to protect those fragile waveforms rather than merely to maximize a global signal-quality score. Their model, called FG-DPUNet, is described in Frontiers in Signal Processing.

The core problem the authors identify is architectural. Single-branch U-Net networks, a popular family of encoder-decoder models for signal enhancement, face an inherent receptive-field trade-off. Small convolution kernels excel at retaining local P-QRS-T waveforms but lack the capacity to suppress large-scale baseline wander and muscle interference. Large kernels remove low-frequency artifacts effectively but tend to blur and attenuate faint diagnostic components. Existing multi-branch designs attempt to combine both scales, yet most rely on fixed averaging or static convolutional fusion in which every time sample shares the same fusion weights. Because motion artifact intensity fluctuates dynamically in wearable recordings, static fusion can only reach a suboptimal compromise, unable to adjust each branch’s contribution to the noise arriving at any given instant.

FG-DPUNet answers this with a compact dual-branch design. A shared shallow convolution layer first extracts low-level features from a noisy single-lead input segment of 1250 samples, corresponding to five seconds at 250 Hz. Those features feed two parallel U-Net branches with different kernel sizes: one with a kernel of 7, whose receptive field covers roughly 28 milliseconds and captures fine temporal detail such as sharp QRS complexes and subtle P and T wave features, and one with a kernel of 21, spanning about 84 milliseconds, which targets slow drifts like baseline wander and electrode motion. Skip connections within each branch help recover fine-grained details lost during downsampling, and the authors note that this makes the difference in QRS preservation between the two branches relatively modest in practice.

The distinctive element is a lightweight feature-aware gating module. Rather than duplicating gating sub-modules at nearly every encoder-decoder and skip-connection layer, as a prior dynamic-gating denoising autoencoder did, FG-DPUNet generates its fusion weights from the shared shallow features using only average pooling and a sigmoid activation. The module aggregates channel-wise information and produces a temporal gate vector bounded between zero and one, which then modulates the element-wise weighted combination of the two branch outputs. In low signal-to-noise segments dominated by baseline wander and electrode motion, a larger gate value boosts the long-receptive-field branch to suppress broad-scale interference. In clean, high-SNR segments, a smaller gate value lets the short-receptive-field branch dominate reconstruction and preserve fine morphological detail. The authors emphasize that the gating adds negligible computational overhead and supports fully end-to-end joint optimization.

Equally important to the team is how denoising quality is measured. Conventional metrics such as signal-to-noise ratio improvement, percentage root-mean-square difference, and root mean square error reflect overall amplitude errors but cannot capture morphological distortion within diagnostically vital waveforms. Clinical interpretation depends heavily on P-wave morphology, PR interval, T-wave morphology, and QT interval, and even minor distortions can impair diagnosis of arrhythmias and myocardial ischemia. To close this gap, the researchers introduce three morphological-temporal fidelity metrics: a P-QRS-T weighted correlation degradation index called Dw, which assigns a weight of five to P-wave and T-wave regions and one elsewhere; and two peak-deviation measures, ΔPR and ΔRT, which quantify absolute errors in the PR and RT intervals in milliseconds. Peak localization is refined by upsampling the 250 Hz signals to 1000 Hz with cubic interpolation before searching for P and T peaks in windows relative to each detected R peak.

Benchmarking took place on two datasets with distinct roles. The MIT-BIH Arrhythmia Database provided controlled additive-noise conditions: six recordings were used, with 30 minutes of clean ECG and 30 minutes of mixed noise from the MIT-BIH Noise Stress Test Database, including baseline wander, electrode motion, and muscle artifact, synthesized at six input SNR levels from −6 dB to 24 dB in five-second segments. The BUT Smartphone PPG Database supplied real-world wearable recordings from 50 subjects aged 19 to 76, captured during walking, laughing, talking, coughing, and resting. For the cross-subject evaluation, 35 eligible subjects were divided into 27 for training, three of whom served for validation, and 8 held out entirely for testing, so no subject overlapped between training and test sets.

On MIT-BIH under the random-split protocol, FG-DPUNet outperformed the 2D-CEDNet comparison model across all six noise levels in SNR improvement and PRD. At the most challenging −6 dB condition, it reached an SNR improvement of 21.61 dB against 19.54 dB for 2D-CEDNet and cut PRD from 89.59% to 18.91%. At 24 dB, where the input is already fairly clean, FG-DPUNet still delivered a positive SNR improvement of 1.85 dB while 2D-CEDNet went negative at −1.66 dB, and PRD fell to 5.77% versus 77.20%. Averaged across all levels and compared against four state-of-the-art methods, FG-DPUNet achieved the highest mean SNR improvement of 12.69 dB and the lowest mean PRD of 10.07%, with an RMSE of 0.04 mV matching the best competitor. The abstract summarizes this as an average SNR improvement gain of 5.94 dB and a PRD reduction from 63.87% to 10.07% relative to prior methods.

Ablation experiments reinforced the design logic. Single-branch U-Nets with either kernel size performed worst overall: the small-kernel variant could not capture broad-scale motion interference, while the large-kernel variant suppressed noise more weakly. All dual-branch models beat the single-branch baselines, and fixed averaging fusion underperformed learnable fusion at every SNR level. On the BUT data, FG-DPUNet recorded the minimum Dw across all motion categories, and the cross-subject experiments showed roughly a 16% improvement in Dw over static dual-branch structures. The trade-offs were candidly reported: DP-UNet, a static concatenation-fusion variant, achieved slightly better RMSE and SNR improvement for walking samples, and FG-DPUNet showed marginally larger ΔPR during walking and slightly higher ΔRT during coughing. Subject-aggregated paired comparisons found no statistically significant inter-model differences for Dw, ΔPR, or ΔRT, which the authors attribute to moderate inter-subject variability and the small test cohort of eight subjects.

The gating module’s adaptivity was tested statistically. Across all BUT test samples, mean gate values correlated negatively with input SNR, with Pearson and Spearman coefficients both strongly significant. Within activity categories, high-motion scenarios showed strong significant negative correlations for walking, coughing, and laughing, with laughing yielding the largest magnitude, while talking showed a weak non-significant negative correlation and resting a trivial non-significant positive one, indicating stable gate outputs under nearly noise-free conditions. Visualized recordings told the same story: walking and laughing segments produced higher mean gate values of 0.57 and 0.55 with substantial real-time fluctuations, whereas resting signals held a stable mean near 0.54. Gate values during QRS complexes stayed close to one, leaning on the long-receptive-field branch, whose skip connections preserved adequate QRS fidelity, though the authors note that P and T wave morphology remains more susceptible to the smoothing effect of the wide receptive field.

The correlation analysis of the new metrics added a further argument for waveform-aware evaluation. Dw correlated strongly with PRD and with ΔRT but only weakly with SNR improvement and ΔPR, suggesting it chiefly characterizes T-wave and RT-segment distortion, information that global amplitude metrics miss. ΔPR and ΔRT were themselves weakly correlated, indicating that PR-interval and RT-interval distortions are relatively independent patterns. In resting scenarios, conventional metrics placed FG-DPUNet and DP-UNet nearly on par, yet the fidelity metrics revealed a 28.0% reduction in Dw and a 2.77 ms shorter ΔRT for the gated model. The team is careful about scope: the metrics’ clinical validity has not yet been established, the model has not been tested on extreme noise such as electrode detachment, and cardiologist validation is planned. The model itself remains light, containing 2.6 million parameters in 9.93 MB, with an average CPU inference latency of 16.58 ms per five-second segment, which the authors present as feasible for real-time use on wearable devices.

Subject of Research: A dual-branch gated deep learning network for removing motion artifacts from wearable ECG signals while preserving diagnostic waveform morphology

Article Title: FG-DPUNet: a feature-aware gating dual-path U-Net for ECG motion artifact removal with waveform integrity

Article References: Lv, B., Wang, S., Yu, C., Xu, Y., Sun, Y., & Yang, X. (2026). FG-DPUNet: a feature-aware gating dual-path U-Net for ECG motion artifact removal with waveform integrity. Frontiers in Signal Processing, 6, Article 1945432. https://doi.org/10.3389/frsip.2026.1945432

Image Credits: AI Generated

DOI: 10.3389/frsip.2026.1945432

Keywords: wearable ECG, motion artifacts, ECG denoising, dual-path U-Net, feature-aware gating, waveform integrity, P-wave, T-wave, deep learning, signal processing, ambulatory monitoring, morphology metrics