Source-linked AI summary
Contactless Electrocardiogram Monitoring with Millimeter Wave Radar
Jinbo Chen, Dongheng Zhang, Zhi Wu, Fang Zhou, Qibin Sun, Yan Chen
TL;DR
Contactless ECG monitoring is needed because electrode-based monitoring is uncomfortable and remains infeasible or unreliable in some settings. The paper combines millimeter-wave radar sensing, 4D cardiac-motion extraction, and deep learning to reconstruct ECG from RF input. The system demonstrates contactless, continuous, and accurate ECG monitoring potential, while evaluation remains limited to healthy individuals and interference can degrade accuracy.
Problem
Contactless ECG monitoring remains unaddressed because electrode-based monitoring can be uncomfortable, unreliable for continuous use, and infeasible for some patients.
Method
A millimeter-wave radar senses cardiac mechanical activity through 4D motion extraction, and an encoder-decoder network transforms RF-derived measurements into ECG.
Results
The system demonstrates contactless, accurate, and continuous ECG monitoring potential through extensive experiments.
Takeaways & Limitations
The results support the feasibility and potential of using radar and learned mechanical-to-electrical domain transformation for contactless ECG monitoring.
Takeaways & Limitations
The evaluation focused on healthy individuals, and interference can distort cardiac-motion measurements and reduce reconstruction accuracy.
Abstract
from arXiv · showhide
The electrocardiogram (ECG) has always been an important biomedical test to diagnose cardiovascular diseases. Current approaches for ECG monitoring are based on body attached electrodes leading to uncomfortable user experience. Therefore, contactless ECG monitoring has drawn tremendous attention, which however remains unsolved. In fact, cardiac electrical-mechanical activities are coupling in a well-coordinated pattern. In this paper, we achieve contactless ECG monitoring by breaking the boundary between the cardiac mechanical and electrical activity. Specifically, we develop a millimeter-wave radar system to contactlessly measure cardiac mechanical activity and reconstruct ECG without any contact in. To measure the cardiac mechanical activity comprehensively, we propose a series of signal processing algorithms to extract 4D cardiac motions from radio frequency (RF) signals. Furthermore, we design a deep neural network to solve the cardiac related domain transformation problem and achieve end-to-end reconstruction mapping from RF input to the ECG output. The experimental results show that our contactless ECG measurements achieve timing accuracy of cardiac electrical events with median error below 14ms and morphology accuracy with median Pearson-Correlation of 90% and median Root-Mean-Square-Error of 0.081mv compared to the groudtruth ECG. These results indicate that the system enables the potential of contactless, continuous and accurate ECG monitoring.
1 INTRODUCTION
Contactless ECG monitoring remains difficult because electrode-based systems are uncomfortable, unreliable for continuous use, and infeasible in some patients. This paper addresses the gap by sensing cardiac mechanical activity with millimeter-wave radar and transforming it into ECG.
- Motivation: Electrode-based ECG monitoring can be uncomfortable, lose continuity and reliability, and be infeasible for burned patients, highly infected patients, and premature babies.Battery limitations and electrode falling further deteriorate monitoring continuity and reliability.
- Motivation: Cardiac electrical and mechanical activities are related through the excitation-contraction coupling mechanism, motivating their use for contactless ECG reconstruction.The paper describes a mapping between these two representations of cardiac activity.
- Approach: The proposed system uses millimeter-wave radar to sense cardiac mechanical activity and transform it into ECG without body contact.Radar reflections are modulated by cardiac motion, providing the input for ECG reconstruction.
- Evaluation: The evaluation uses 200 trials involving 35 participants across 4 physiological statuses, with 10 hours of radar and corresponding ECG data.The dataset contains 7,200,000 radar frames and corresponding ECG ground truth.
- Approach: Signal-processing algorithms extract 4D cardiac motion signals from raw RF measurements to represent cardiac mechanical activity comprehensively.The pipeline includes beamforming, micro-motion amplification, cardiac signal focusing, and spatial filtering.
- Approach: A hierarchical deep neural network learns the transformation from cardiac mechanical-domain measurements to cardiac electrical-domain ECG representations.The network exploits temporal and spatial RF features while incorporating physiological domain knowledge.
2 RELATED WORK
Prior RF and millimeter-wave sensing studies mainly recovered heart rate, heartbeat periods, or selected cardiac signals rather than direct ECG. This work extends wireless sensing by combining comprehensive 4D cardiac motion measurements with deep neural ECG reconstruction.
- RF-based Sensing: RF sensing research has addressed localization, walking speed, gesture recognition, event detection, and human pose estimation, alongside physiological sensing tasks.These studies establish RF signals as a basis for sensing human motion and physiological activity.
- RF-based Sensing: Earlier cardiac RF studies primarily extracted average heart rate or heartbeat period rather than direct ECG signals.Other work measured diastole and systole periods or recovered the dominant heart sounds S1 and S2.
- Contribution: This paper proposes 4D cardiac motion measurements to represent cardiac mechanical activity comprehensively.The approach targets richer cardiac motion information than prior measurements focused on selected temporal or acoustic events.
- Contribution: A deep neural network transforms radar-based cardiac motion representations into ECG using RF input.The model addresses the cardiac domain transformation problem between mechanical and electrical activity.
3 SYSTEM DESIGN
The system reconstructs ECG without electrodes by sensing cardiac mechanical motion with millimeter-wave radar and transforming it into the cardiac electrical domain. It extracts spatially resolved 4D motion signals through staged processing before applying a deep network for RF-to-ECG mapping.
- Working Principle and System Overview: The system exploits the nonlinear relation between cardiac mechanical and electrical activity to infer ECG from indirect, contactless mechanical measurements.Cardiac excitation-contraction coupling bridges the two representations, motivating domain transformation from body-surface motion to electrical potential.
- Cardiac Motion Measurements in Radar: Micro-motion amplification improves cardiac-motion SNR while suppressing breathing effects and noise, especially for voxels near the heart.For distant voxels, cardiac-motion energy is consumed along the conduction path, leaving amplified signals dominated by meaningless noise.
- Cardiac Signals Focusing: Dynamic time warping-based pattern matching identifies voxel signals with periodic morphology similar to successive cardiac cycles and rejects signals when matching fails.The process segments repeating cardiac patterns and synchronizes their lengths with ECG R-peaks.
- Spatial Filtering for Cardiac Signals: Power-weighted spatial filtering merges focused signals within clusters, emphasizing high-SNR motion and suppressing noise interference.The resulting representation contains cardiac motion sequences paired with their 3D centroid locations as 4D measurements.
- Domain Transformation: An encoder-decoder deep neural network maps 4D cardiac motion measurements to ECG by learning temporal-spatial features and intermediate cardiac hidden representations.The architecture separately learns mappings from RF signals to hidden features and from hidden features to ECG measurements while incorporating physiological domain knowledge.
4.1 Experimental Setup
The system is evaluated in a clinically relevant supine, quasi-static setup with the radar positioned above the chest and ground-truth ECG collected simultaneously.
- 4.1 Experimental Setup: Participants lie supine and remain quasi-static during radar and ECG acquisition.The radar is placed 0.4–0.5 m above the torso chest, with its antenna main lobe directed approximately toward the sternum.
- 4.1 Experimental Setup: Ground-truth ECG is recorded simultaneously with radar measurements using a TI ADS1292 evaluation board.
- 4.1 Experimental Setup: The experimental setting is designed to reflect common hospital ECG-monitoring scenarios.
4.2 Implementations Details
The implementation combines a 12-channel millimeter-wave radar array, offline 3D signal processing, and a neural-network pipeline trained on segmented radar data.
- 4.2 Implementations Details: The radar uses 3 transmitters and 4 receivers to form a virtual 2D antenna array with 12 channels.Time-division multiplexing separates transmissions from the multiple transmit antennas.
- 4.2 Implementations Details: Radar signal processing is performed offline in Matlab on a cropped 9×17×9 sensing grid containing 1377 voxels.The grid covers x and y from −0.4 m to 0.4 m and z from 0.35 m to 0.6 m.
- 4.2 Implementations Details: The first CNN uses four repeated blocks of padded 1D convolutions, ReLU, batch normalization, and max pooling.
- 4.2 Implementations Details: The temporal encoder maps 50 × 1 × 640 inputs to 50 × 32 × 80 features, followed by a three-block transformer spatial encoder.
- 4.2 Implementations Details: Training uses Adam with a 0.001 learning rate, mini-batches of 64, and length-640 segments sampled with step 30.At 200 Hz, every three minutes of data provides approximately 1100 training samples.
4.3 System Performance
The system evaluates contactless ECG monitoring through cardiac-event timing and waveform morphology across physiological conditions, participants, age groups, and interference settings. Performance is generally consistent, but radar resolution and interference constrain accuracy, while clinical morphology evaluation remains incomplete.
- Evaluation scope: The evaluation measures cardiac-event timing and waveform morphology, both identified as fundamental accuracy dimensions for ECG diagnosis.Timing covers Q, R, S, and T waves; morphology is assessed against ground-truth ECG waveforms.
- Overall performance: 0.5%, 1.2%, 2%, and 3.5% are the 90-percentile false monitoring ratios during sleep, normal-breath, irregular-breath, and post-exercise conditions, respectively.A monitoring interval is classified as false when consecutive waves cannot be delineated within 150 ms of the ground truth.
- Morphology accuracy: 90% median correlation and 0.081mv median RMS error summarize overall morphology accuracy against the ground-truth ECG.Across four physiological statuses, median correlation ranges from 87% to 91% and median RMS error from 0.075mv to 0.096mv.
- Clinical usage potential: Across 35 unseen participants, timing errors remain below 54ms, RMS error below 0.17mv, and correlation exceeds 0.8 for 26 participants.Participant-level results vary, with one participant reaching 95% correlation and another showing 65% correlation with larger timing and RMS errors.
- Clinical usage potential: Timing and morphology remain closely similar across four age groups, with median correlations of 86%, 92%, 89%, and 93%.The reported median normalized timing errors also remain close across age groups for Q, R, S, and T peaks.
5 CONCLUSION
The system provides contactless ECG monitoring from RF signals by sensing cardiac mechanical activity and reconstructing cardiac electrical activity. Daily-life analysis examines performance across interference levels and measurement distances.
- The proposed system offers contactless monitoring of cardiac electrical activity from RF signals.It combines cardiac mechanical-activity sensing, signal processing, and deep-learning-based ECG reconstruction.
- Daily-life usage analysis evaluates event-timing RMS error and mean correlation under varying interference levels and measurement distances.
- The system reconstructs ECG output from cardiac mechanical activity using an end-to-end RF-input mapping.