LifWavNet: Lifting Wavelet-based Network for Non-contact ECG Reconstruction from Radar
Soumitra Kundu, Gargi Panda, Saumik Bhattacharya, Aurobinda Routray, Rajlakshmi Guha · 2025
Read LifWavNet: Lifting Wavelet-based Network for Non-contact ECG Reconstruction from Radar by Soumitra Kundu, Gargi Panda, Saumik Bhattacharya, Aurobinda Routray, Rajlakshmi Guha (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-27692.txt.
Introduction
Soumitra Kundu, Gargi Panda, Saumik Bhattacharya, Aurobinda Routray, and Rajlakshmi Guha present LifWavNet in 2025, a network that rebuilds heart traces from radar without touch. An electrocardiogram, or ECG, records the heart's electric signal and is key to care. Radar can sense chest motion from a distance, which suits watch without wires or pads.
LifWavNet uses learnable lifting wavelets, which split a signal into layers at distinct scales and can adapt their shape. A joint time-and-frequency loss keeps the rebuilt trace true in both views. Tests on two public sets beat prior methods on trace quality and on heart rate and its variation. Views of inner features aid trust in what the net has learned.
Catalog record
| CREATOR | Soumitra Kundu, Gargi Panda, Saumik Bhattacharya, Aurobinda Routray, Rajlakshmi Guha |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-27692
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
LifWavNet: Lifting Wavelet-based Network for Non-contact ECG Reconstruction from Radar Soumitra Kundu, Gargi Panda, Saumik Bhattacharya, Aurobinda Routray, Rajlakshmi Guha Abstract Non-contact electrocardiogram (ECG) reconstruction from radar signals offers a promising approach for unobtrusive cardiac monitoring. We present LifWavNet, a lifting wavelet network based on a multi-resolution analysis and synthesis (MRAS) model for radar-to-ECG reconstruction. Unlike prior models that use fixed wavelet approaches, LifWavNet employs learnable lifting wavelets with lifting and inverse lifting units to adaptively capture radar signal features and synthesize physiologically meaningful ECG waveforms. To improve reconstruction fidelity, we introduce a multi-resolution short-time Fourier transform (STFT) loss, that enforces consistency with the ground-truth ECG in both temporal and spectral domains. Evaluations on two public datasets demonstrate that LifWavNet outperforms state-of-the-art methods in ECG reconstruction and downstream vital sign estimation (heart rate and heart rate variability). Furthermore, intermediate feature visualization highlights the interpretability of multi-resolution decomposition and synthesis in radar-to-ECG reconstruction. These results establish LifWavNet as a robust framework for radar-based non-contact ECG measurement. I Introduction Electrocardiogram (ECG) is the clinical gold standard for measuring cardiac electrical activity, and its continuous monitoring provides essential information for the diagnosis and management of various cardiac diseases [ 1 ] . However, conventional ECG systems that use skin-contact electrodes can cause discomfort, skin irritation, and reduced compliance during long-term use—particularly in infants, elderly individuals, and patients with burns or sensitive skin. Advancements in optical photoplethysmography (PPG) have introduced wearable devices that offer a less obtrusive option, but these still require skin contact. Remote PPG (rPPG) enables contact-free monitoring using visual cameras, yet its performance is constrained by ambient lighting conditions. Radar sensing has emerged as a promising non-contact alternative, capable of detecting sub-millimeter chest wall displacements caused by cardiac activity, without being affected by lighting conditions. Early radar-based methods followed a signal processing pipeline beginning with radar I/Q extraction, demodulation, and phase unwrapping to obtain chest displacement signal. Respiration suppression was then applied using techniques such as bandpass/adaptive filtering, subspace methods, or wavelet decomposition. The cardiac band was subsequently isolated using time–frequency analysis, and heartbeats were tracked through spectral peaks, harmonic summation, or template matching [ 2 , 3 , 4 , 5 , 6 ] . While these approaches could provide reasonable mean heart rate (HR) estimates, they faced significant limitations when extended to full ECG waveform reconstruction, which requires an accurate representation of all P, Q, R, S, and T waves. Accurate ECG reconstruction from radar remains challenging because the weak cardiac motion signals are often masked by stronger respiratory movements and body motion. Furthermore, the relationship between mechanical and electrical cardiac activity is inherently nonlinear and highly sensitive to noise. In recent years, deep learning (DL) methods [ 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 ] have shown considerable promise in addressing these challenges and enabling radar-based ECG reconstruction. Fig. 1 : Overview of radar signal-based non-contact ECG reconstruction. The radar measures the chest movement signal in a non-contact manner, from which a neural network reconstructs the ECG signal. I-A Motivation ECG records the heart’s electrical activity through its characteristic PQRST waveform, with each component corresponding to a distinct phase of the cardiac cycle. The P-wave represents atrial depolarization, initiating atrial contraction and blood transfer to the ventricles. The QRS complex, reflecting sequential ventricular depolarization, drives ventricular contraction and blood ejection, while the T-wave denotes ventricular repolarization, leading to relaxation and filling in preparation for the next cycle. This correspondence between the ECG’s electrical events and the heart’s mechanical actions, arising from the excitation–contraction coupling (ECC) mechanism [ 18 ] , motivates modeling the nonlinear, cross-modal transformation from radar signal to ECG waveform. Recent deep learning-based radar to ECG reconstruction methods mainly adopt CNN–LSTM [ 7 ] , U-Net [ 10 ] , and transformer-based [ 12 ] architectures, with time and frequency domain-based loss functions to perform the non-linear mapping. Generative models such as GANs [ 9 , 16 ] and diffusion networks [ 11 ] employ probabilistic generative perspective to learn the data-distributions. However, these architectures are pure DL-based, designed with trial-and-test strategy, and largely operate as black box. Some recent approaches [ 8 , 14 , 15 ] have proposed to estimate an intermediate 4D cardiac representations to better model the radar to ECG reconstruction, at the cost of heavy pre-processing. Also, the above-mentioned methods lack explicit inductive bias toward the known structure of cardiac signals. This is where wavelets become an important alternative. Unlike Fourier or short time Fourier transform (STFT), which assume fixed resolution across time and frequency, wavelets provide adaptive time–frequency localization. In radar‑to‑ECG, this property is crucial because the radar signal contains overlapping respiratory and cardiac components, and the wavelets provide principled multi-resolution analysis that can isolate sharp transients such as QRS complexes while preserving low frequency morphology (P/T waves and baseline) [ 19 , 20 , 21 , 22 ] . More recently, Xu et al. [ 17 ] proposed WaveGRUNet, which incorporates Sym4 wavelet based multi-level decomposition as a pre-processing step, followed by CNN and gated recurrent units (GRUs) for feature extraction and ECG reconstruction. Similarly, Kim et al. [ 23 ] integrated wavelet transform with adaptive neuro‑fuzzy networks to preserve nonlinear dependencies while extracting ECG‑relevant frequency bands. Wavelets can effectively encode the multi‑scale nature of cardiac electrophysiology (wideband QRS vs. narrow band P/T). However, prior works have only used fixed, linear wavelet bases (e.g., Sym4, dB4) for pre-processing the radar signal. This underutilizes the potential of wavelets, pointing toward the need for learnable wavelet-based framework that can adaptively capture the nonlinear radar–ECG mapping. Moreover, the existing methods employ a single window STFT-based loss functions to constrain the reconstructed signal with ground truth ECG in both time and frequency domains. The PQRST waveform in ECG signal contain different frequencies, and a multi window-based STFT loss function can be more effective. Therefore, there is a need for a wavelet-based, learnable framework capable of capturing nonlinear cross-domain mappings. Additionally, incorporating an improved time–frequency constraints seem essential to better preserve the diverse temporal and spectral features of the ECG waveform. I-B Methodology Overview and Contributions Motivated by the insights described above, we first introduce a multi‑resolution analysis and synthesis (MRAS) model for the radar to ECG reconstruction task. Building on this formulation, we design LifWavNet, a novel end‑to‑end network that employs a learnable lifting wavelet scheme. The lifting scheme provides a flexible framework for constructing wavelets, distinct from classical fixed bases such as Haar [ 24 ] . Unlike these linear, pre‑defined filters, lifting enables the design of nonlinear, learnable filters that can adapt to the underlying structure of the data [ 25 ] . Prior studies have successfully applied lifting wavelet networks to image classification, compression, and denoising, where they demonstrated improved adaptability and interpretability compared to fixed transforms [ 25 , 26 , 27 ] . However, these applications involve reconstruction within the same modality. In contrast, radar‑to‑ECG reconstruction is a cross‑domain problem, requiring the transformation of radar echoes into physiologically meaningful ECG waveforms. In our MRAS model, the radar signal is decomposed into approximation and detail components through multi‑resolution analysis (MRA), and then these components are