LifWavNet Explained: How a Lifting Wavelet Network Enables Non-contact ECG Reconstruction from Radar

Key Takeaways

• LifWavNet employs a 4-level lifting wavelet decomposition to transform radar SCG signals into clean ECG, enabling robust non-contact monitoring.
• The architecture links wavelet-domain representations to ECG, producing interpretable intermediates and exhibiting superior noise resilience compared to raw features.
• It provides a concrete architecture, a trainable pipeline, and explicit metrics suitable for replication.
• The market shows strong demand for non-contact ECG: the overall ECG devices market was valued at $7.7B in 2024 with a 4.2% CAGR (2025–2034), while the ECG Patch/Holter segment was $2.1B with an 11.0% CAGR (2025–2033).
• Expert perspective suggests that radar-based SCG recovery is less studied than traditional ECG due to ECG’s broader diagnostic information (e.g., atrial activity), highlighting the importance of radar-to-ECG translation.
• For implementation, use db4 wavelets, 4 lifting levels, 64–128 feature channels per block, and a mix of synthetic and real data (e.g., 5,000 synthetic sequences + 100 real sequences).
• Privacy is addressed through on-device inference, anonymization, and data minimization to support clinical safety.

Technical Deep Dive: LifWavNet Architecture, Training, and Implementation

Lifting Wavelet Network: Core Idea and Why It Works for Radar-based SCG

Radar-based SCG (Seismocardiography) signals contain vital information about the heartbeat, but converting them into clear, ECG-like waveforms is challenging due to interference from breathing and body motion. LifWavNet addresses this by integrating a lifting-based wavelet transform with a learnable mapping. This approach generates a multi-resolution view that accurately reflects the heart’s true signals.

Core Ideas at a Glance

• LifWavNet combines a lifting-based wavelet transform with a learnable mapping to create a multi-resolution representation of radar SCG signals, allowing for robust ECG reconstruction despite breathing and motion variations.
• The 4-level wavelet decomposition extracts detail and approximation coefficients, which the network then maps to ECG-domain waveforms, preserving key P/QRS/T-wave structures.
• Processing in the wavelet domain helps to denoise and stabilize non-stationary radar signals, reducing aliasing and motion-induced artifacts before ECG reconstruction.

How It Works: Multi-Resolution ECG Reconstruction

Four successive wavelet levels decompose the input radar SCG signal into a hierarchy of coefficients. These coefficients effectively separate coarse trends from fine details. The LifWavNet module then learns to translate these coefficients into ECG-domain waveforms, ensuring that essential features like the P, QRS, and T components are preserved in the reconstructed ECG signal. This multi-resolution pathway enables the model to be robust to breathing patterns and body motion, as it processes information at scales relevant to both global heart-rate dynamics and local waveform specifics.

Why Wavelet-Domain Processing Matters for Radar SCG

Radar signals used for SCG are inherently non-stationary and susceptible to motion and environmental noise. Wavelet-domain processing offers a natural method for separating noise and motion artifacts from the heart’s signal content. By denoising and stabilizing signals prior to reconstruction, LifWavNet mitigates aliasing and artifact contamination that could otherwise distort the ECG waveform. The outcome is a cleaner, more reliable ECG-like reconstruction suitable for downstream analysis and monitoring in dynamic, real-world settings.

Expert Context: Why Focus on Radar-to-ECG Translation

Radar-based SCG recovery is a less explored area compared to traditional ECG research. ECG provides more comprehensive physiological information, such as atrial activity, which is crucial for diagnosis and monitoring. This gap highlights the need for solutions like LifWavNet, which aims to translate radar-derived SCG into clinically useful ECG waveforms, thereby bridging the divide between non-contact sensing and actionable cardiac insights.

Market Relevance: Why This Matters Now

The market landscape underscores the significance of LifWavNet:

Market Segment	2024 Revenue	CAGR
ECG devices (overall)	$7.7B	4.2%
Patch/Holter monitors	$2.1B	11.0%

The expanding ECG devices market and the rapid growth in the patch and Holter segments provide strong incentives for advancing radar-based ECG reconstruction. LifWavNet presents a practical pathway to extract ECG-quality signals from radar, enabling more convenient, non-contact monitoring solutions.

In Summary: By integrating a lifting-based wavelet decomposition with a learnable mapping, LifWavNet establishes a stable, multi-resolution processing pipeline from radar SCG to ECG-domain waveforms. It effectively preserves critical cardiac signal structures, reduces non-stationary noise, and addresses a significant market demand for robust radar-to-ECG translation in a rapidly evolving field.

Architecture Details: Layers, Activation, and Connectivity

The transformation of radar-derived time-series data into a clean, clinically meaningful ECG signal is achieved through a streamlined, wavelet-powered architecture. This design prioritizes temporal alignment, supports real-time streaming, and ensures robustness via careful normalization and regularization.

Aspect	Key Details
Input representation	Preprocessed radar-derived time-series features with shape [batch, time, channels]. Includes phase, amplitude, and micro-Doppler information. Designed to maintain temporal alignment and channel-wise information for interpretation.
Lifting blocks	4 sequential lifting-based wavelet blocks. Each block comprises a forward wavelet lift, a small CNN (2–3 conv layers, 64–128 filters), and a residual connection feeding back into the wavelet domain.
Wavelet specifics	Daubechies-4 (db4) lifting steps are used for multi-resolution coefficients, enabling stable reconstruction from compressed representations and facilitating clean transitions between time and wavelet domains.
Nonlinear mapping & stability	GELU activations, LayerNorm after each block, and dropout (0.1–0.2 range) enhance generalization without compromising signal fidelity.
Output	A single-channel, time-aligned ECG waveform, ready for downstream clinical interpretation or multi-lead augmentation via domain adapters.
Regularization & stability	Weight decay around 1e-5. Ablation studies confirm that four lifting levels offer an optimal balance between reconstruction fidelity and computational cost.
Real-time potential	Streaming-capable blocks with lightweight CNNs allow for near real-time inference on modern GPUs or edge devices.

Input Representation: Preparing the Radar-to-ECG Bridge

The model commences with preprocessed radar-derived time-series features. Maintaining the shape [batch, time, channels] and incorporating phase, amplitude, and micro-Doppler components ensures the preservation of the temporal structure essential for mapping to an ECG waveform. Normalization and thoughtful feature selection stabilize learning across diverse subjects and conditions.

Lifting Blocks: Four Stages of Multi-Resolution Processing

The architecture’s core consists of a stack of four sequential lifting-based wavelet blocks. Each block follows a consistent pattern:

1. Forward Wavelet Lift: Transitions the signal into a multi-resolution representation using db4 lifting steps.
2. Wavelet-Domain CNN: A small convolutional neural network (2–3 layers, 64–128 filters) operates on the wavelet coefficients.
3. Residual Connection: The output of the block is fed back into the wavelet-domain representation, preserving information and promoting stable reconstruction.

This stacking of four blocks creates a compact, hierarchical representation capable of capturing the fine-grained and coarse temporal structures needed for accurate ECG reconstruction, while remaining computationally efficient.

Wavelet Specifics: Why db4 Lifting Steps

The use of Daubechies-4 (db4) lifting steps provides a principled approach to signal decomposition across multiple resolutions with stable reconstruction. The lifting scheme simplifies the inversion process, helps control numerical errors, and supports a streaming-friendly workflow where coefficients at different scales can be progressively refined as new data arrives.

Nonlinear Mapping and Normalization: Shaping the Feature Space

Each block incorporates GELU activations to introduce smooth non-linearity, enabling more effective modeling of complex data relationships. Layer Normalization is applied after every block to stabilize training by normalizing features across channel and time dimensions. Optional dropout (0.1–0.2 range) assists in preventing overfitting, particularly when training with varied radar datasets.

Output and Alignment: Time-Synchronous ECG

The network generates a single ECG channel that is precisely time-aligned with the input signal. This ensures suitability for immediate clinical interpretation or as a basis for multi-lead augmentation using domain adapters, offering flexibility in various clinical workflows.

Regularization and Stability: Keeping Learning Robust

Weight decay is set around 1e-5 to regularize model complexity without hindering learning capacity. Ablation experiments indicate that four lifting levels strike an effective balance: higher levels might improve fidelity but increase computation, while fewer levels risk suboptimal reconstruction.

Real-Time Potential: Streaming-Ready Design

The architecture’s lifting blocks and lightweight CNNs are designed for streaming applications. As data streams in, wavelet-domain computations can be incrementally updated, facilitating near real-time inference on modern GPUs and edge devices with minimal latency.

In essence, this architecture merges a principled wavelet framework with compact neural blocks, modern non-linearities, and meticulous regularization to produce stable, real-time ECG waveforms from radar-derived time-series inputs.

Training Protocol: Data, Losses, and Hyperparameters

The training protocol for this radar-to-ECG translator is designed to be focused and practical. It utilizes synthetic data to cover a wide spectrum of ECG morphologies, a small subset of real-subject data for fine-tuning, a balanced suite of loss functions to preserve timing and morphology, and carefully selected hyperparameters to capture heart rhythms and higher-frequency details. The following protocol details the experimental setup.

Data Strategy

Source	Quantity / Examples	Notes
Synthetic simulations	5,000 sequences	Ground-truth ECG derived from simulated radar interactions.
Real-subject subset	100 sequences	Used for fine-tuning to enhance real-world robustness.

Preprocessing

• Normalize all inputs to the [-1, 1] range to ensure consistent scaling across sensors and subjects.
• Segment radar–ECG pairs into 2-second windows with 50% overlap to increase sample diversity and stabilize learning.

Loss Composition

• Mean Squared Error (MSE) between the reconstructed ECG and the ground-truth ECG.
• Spectral loss on the STFT magnitude to preserve energy in key components (QRS complex and T-wave).
• Dynamic Time Warping (DTW)-based temporal alignment loss to accommodate small latency shifts between radar and ECG signals.
• Optional perceptual ECG-feature loss using a pre-trained ECG extractor to encourage physiologically meaningful representations.

Optimization

• Optimizer: Adam with learning rate lr = 1e-4 and betas = (0.9, 0.999).
• Learning-rate scheduling: Cosine decay or plateau-based scheduling based on validation performance.
• Batch size: 32.
• Total training steps: Approximately 100,000.

Hyperparameters

• Model architecture: 4 lifting levels.
• Channels per level: 64–128 (adjustable based on dataset and compute budget).
• Per-block receptive field: Tuned to capture typical ECG rhythm frequencies (~0.6–1 Hz) and higher-frequency components like noise and murmurs.

Augmentation

• Breathing-rate jitter (±20%) to simulate natural physiological variability.
• Random motion jitter to mimic subject movement and sensor misalignment.
• Signal-to-noise ratio (SNR) variations to improve robustness in different environments.

Evaluation Protocol

• Cross-subject validation using metrics such as RMSE, MAE, Pearson correlation, and Bland-Altman analysis to assess agreement and bias across individuals.
• Inference latency is reported on target hardware to quantify real-time feasibility.

Privacy, Compliance, and Reproducibility

Radar-based non-contact ECG monitoring has the potential to reveal sensitive physiological signals from a distance. This capability necessitates robust privacy safeguards, a solid plan for reproducibility, and thoughtful deployment strategies, particularly for edge devices. The following guidance offers practical measures for researchers and development teams.

1. Privacy Safeguards for Radar-Based Non-Contact ECG

As noted, “Radar-based non-contact ECG reconstruction raises privacy concerns; LifWavNet documentation recommends on-device inference and data minimization, anonymization of radar traces, and secure data handling.” Practical implementation should include:

• On-device inference: Process raw radar traces and inferences locally on the user’s device whenever feasible to minimize network exposure.
• Data minimization: Collect and retain only the essential data required for ECG signal computation, purging intermediate data promptly.
• Anonymization of radar traces: Remove or obscure any identifiers in radar traces that could link measurements to individuals before sharing or aggregation.
• Secure data handling: Employ encryption for data at rest and in transit, enforce strict access controls, and maintain audit logs for all data handling activities.

2. Reproducibility Plan

Reproducibility is built on transparency and verifiability. A comprehensive plan involves:

• Fixed random seeds: Specify explicit seed values for all stochastic processes to ensure consistent results across multiple runs.
• Explicit data splits: Document all training, validation, and test splits (and report cross-validation folds if used).
• Experiment tracking with versioned datasets: Log hyperparameters, dataset versions, and results using a central tracking system (e.g., MLflow, Weights & Biases).
• Release of model architecture and training scripts: Provide clear, accessible code for the model, loss functions, optimization procedures, data loading, and preprocessing steps to facilitate replication.
• Synthetic data generator for replication: Share a synthetic data generator capable of producing radar traces similar to real data, accompanied by clear documentation on its assumptions and limitations.

3. Releasing Code and Data: Licensing, Caveats, and Domain Adaptation

• Licensing: Clearly define the licensing terms for shared code and data, specifying reuse rights and any attribution requirements.
• Synthetic-data caveats: Explicitly state the potential differences between synthetic and real radar data and their implications for evaluation and generalization.
• Domain adaptation guidance: Offer practical advice for adapting the model from synthetic to real radar data, such as fine-tuning with a small real-data subset, employing domain-invariant features, or testing on a distinct real-world dataset.

4. Edge Deployment Considerations

Designing for edge deployment ensures privacy and performance on devices with constrained computational resources and memory. Key considerations include:

• Quantization-friendly blocks: Structure model components to allow for quantization to lower bitwidths with minimal accuracy loss, enabling faster inference on smartphones or embedded systems.
• Efficient lifting-wavelet operations: Implement fast, memory-efficient wavelet transforms suitable for real-time ECG reconstruction from radar data.
• On-device privacy-preserving inference pathways: Prioritize keeping data local and leverage secure enclaves or trusted execution environments when necessary to protect sensitive signals during processing.

Closing Thought: Privacy, reproducibility, and edge efficiency are foundational elements that enhance the trustworthiness and impact of radar-based health sensing technologies. Integrating these practices throughout the research and product development lifecycle allows for confident innovation while upholding user rights and scientific integrity.

LifWavNet vs. Traditional Radar-to-ECG Pipelines

Here’s a comparative overview:

Aspect	LifWavNet	Traditional Radar-to-ECG Pipelines
Model core	Utilizes a 4-level lifting-wavelet transform with a learnable mapping.	Often relies on CNNs/RNNs operating directly on raw radar features.
Feature representation	Operates on wavelet coefficients for multi-resolution ECG-like reconstruction.	Primarily uses raw radar frames or hand-crafted features.
Data efficiency	Achieves strong results with approximately 5,000 synthetic sequences plus minimal real-set fine-tuning (e.g., 100 sequences).	Many traditional methods require extensive paired real ECG-radar datasets.
Interpretability	Wavelet-domain intermediates offer insights into ECG component reconstruction across frequency bands.	Raw CNN/RCNN methods generally provide less transparent frequency decomposition.
Reproducibility and transparency	Emphasizes explicit architecture, data splits, and open-source code.	Some traditional pipelines exhibit limited hyperparameter reporting or code availability.
Performance targets	RMSE around 0.04–0.08 mV on synthetic benchmarks; real-time inference in the 2–10 ms per sample range on modern GPUs for streaming radar data.	Specific targets vary significantly by model and hardware, not explicitly defined here.
Privacy posture	Supports privacy-friendly monitoring if on-device inference and data minimization are enforced.	May involve cloud processing or data transfer, potentially raising privacy concerns.
Market context alignment	Aligns with ECG device market growth, positioning radar-based non-contact ECG as a scalable monitoring modality.	Market predominantly centers on contact ECG devices; radar-based non-contact is an emerging alternative.

Practical Implementation Guide and Reproducibility

Pros

• Clear and repeatable pipeline
• Explicit LifWavNet architecture details
• Well-defined data-generation strategy
• Specific loss components and hyperparameters
• Comprehensive plan for cross-subject validation
• Guidance for privacy-conscious deployment

Cons

• Relies partly on synthetic data for scalability.
• Domain adaptation may be necessary for certain real-world radar setups.
• Potential computational overhead from wavelet transforms.
• Regulatory considerations for clinical deployment need to be addressed.