Introduction

For over a century, the 12-lead electrocardiogram (ECG) has been recognized as an accessible and reliable gold standard for non-invasive cardiac assessment1,2, offering comprehensive insights into cardiac electrical activity and its association with arrhythmias, heart failure, stroke, and even mortality3,4,5,6,7. Moreover, long-term and continuous ECG assessment has emerged as a promising approach increasingly endorsed in clinical guidelines8,9. However, the current healthcare infrastructure, particularly in resource-limited settings, is not yet capable of supporting scalable monitoring10,11. Wearables like wrist-worn devices and chest patches offer convenient long-term ECG monitoring but are usually constrained by limited-lead configurations and potential skin irritation12,13. These limitations highlight the urgent need for scalable, continuous, and user-friendly solutions for cardiovascular monitoring14.

In this work, we introduce CardioRadar, a compact radar-based system for contactless and continuous 12-lead ECG monitoring at a distance. As illustrated in Fig. 1, the device comprises two compact printed circuit boards (PCBs) housed within a protective plastic shell, each approximately 6 cm in diameter. The upper PCB incorporates a radar chip for signal acquisition, while the lower PCB includes wireless communication modules for efficient data transmission. By integrating off-the-shelf radar and IoT chips with custom-designed components, such as power integrated circuits (ICs) and storage modules, the system achieves an efficient and portable design. The radar chip is priced at around $20, keeping the total system cost under $30. This makes it a cost-effective solution, especially considering that 12-lead ECG machines typically cost $1000 and more15. This compact and low-cost architecture may enable scalable deployment in both clinical and home settings, with potential value in low-resource environments where cardiovascular disease (CVD) imposes major economic burdens16.

Fig. 1: Overview of the CardioRadar system.
Fig. 1: Overview of the CardioRadar system.The alternative text for this image may have been generated using AI.
Full size image

a Modular radar device with two stacked PCBs: the upper integrates the radar chip for signal acquisition, and the lower handles wireless communication. Each PCB is 6 cm in diameter and combines commercial components with custom circuitry. b The system uses an LFMCW radar at 60 GHz with 4 GHz bandwidth and a 3 × 4 MIMO virtual array, producing 15 × 12 I/Q data per frame. c The radar is positioned 0.5 m from the subject for continuous contactless cardiac monitoring. d Contactless sensing captures chest vibrations from both respiration and heartbeat; 3D beamforming isolates cardiac motion for precise detection. e Filtered phase and I/Q signals at a selected spatial point with synchronized ECG lead. f Example of reconstructed 12-lead ECG showing atrial fibrillation features.

Radar-based technology provides the foundation for this approach, offering great potential for contactless and burden-free cardiac monitoring17,18,19. By capturing subtle chest motion, radar can detect respiration and heartbeat20,21,22, track heart-rate variability18, and assist in cardiac assessment19,23. Recent studies further demonstrate that radio-frequency signals can reconstruct ECG waveforms17,24,25. This capability arises from the intrinsic coupling between electrical excitation and mechanical contraction in the heart26. For instance, mechanical and electrical phase singularities co-exist during fibrillation27, and mechanical activity can modulate electrical signals through mechano-electrical feedback28,29. Such coupling has also inspired recent advances in mechano-electrical inference30,31,32 and ECG synthesis using deep learning33,34. Despite recent progress, current radar-based ECG approaches remain constrained to single-lead systems, controlled settings, and small-scale cohorts, while generalizable reconstruction of full 12-lead ECG remains a major challenge.

Our CardioRadar system was validated in a cohort of 6974 individuals across diverse clinical environments, with dataset details provided in Supplementary Note 1. It achieves an average correlation of 0.75 (median: 0.77), a root mean square error (RMSE) of 0.09 mV (median: 0.07 mV), and a mean inter-beat interval (IBI) error of 23.39 ms (median: 5.60 ms) in reconstructing 12-lead ECG signals. The system accurately reproduced various cardiac waveforms and detected abnormalities with performance comparable to conventional ECG. It also demonstrated feasibility for overnight, contactless monitoring. These findings highlight its potential as a transformative, clinically viable solution for continuous cardiac assessment and as a foundation for future studies of long-term cardiac dynamics and disease progression.

Results

System design and capabilities

The proposed radar device illustrated in Fig. 1a dopts a modular design integrating a radar chip and wireless communication modules in a compact 6 cm portable unit. As shown in Fig. 1b, it employs Linear Frequency Modulated Continuous Wave (LFMCW) radar for range measurement and a 3 × 4 virtual Rx-Tx MIMO array for enhanced spatial resolution35. Operating at 60 GHz with 4 GHz bandwidth, each transmitter emits one chirp per frame at 100 Hz, generating 15 × 12 I/Q samples per frame (15 range bins and 12 virtual antennas). The system enables real-time uncompressed data transmission at 627.2 Kbps, comparable to a 360 p video stream. The device is positioned about 0.5 m in front of the subject for contactless monitoring (Fig. 1c) and can operate continuously overnight for long-term observation.

Our radar captures chest vibrations caused by respiration and heartbeat (Fig. 1d). Using 3D beamforming, the system reconstructs signals from multiple spatial locations over time, generating temporal voxels that represent chest motion. Respiratory movement typically occurs at 0.2–0.5 Hz (12–30 breaths per minute), whereas cardiac motion appears at higher frequencies of 1–2 Hz (60–120 beats per minute). This frequency separation allows filtering methods to suppress respiratory components and enhance subtle heartbeat-induced variations, with detailed radar configurations and representative signal examples provided in Supplementary Note 2.

The ECG waveform comprises three principal components: the P wave (atrial depolarization), the QRS complex (ventricular depolarization), and the T wave (ventricular repolarization). Prior studies have demonstrated a strong coupling between electrical activation and mechanical function26,36, observable in pulsed-wave Doppler and echocardiography37. As shown in Fig. 1d, the R wave coincides with peak ventricular contraction during the isovolumic phase, while the T-wave termination corresponds to minimal ventricular volume. Each cardiac cycle also exhibits distinct blood-flow patterns. This tight temporal alignment between electrical and mechanical events provides a physiological basis for reconstructing electrical activity from mechanical motion. In our system, radar-captured mechanical signals reflect this coupling, with the R peak showing the strongest correspondence to the heartbeat and weaker, yet discernible, alignment observed for other waveform components. By integrating motion patterns across all spatial locations on the chest, the system can thus infer the underlying cardiac electrical activity. Figure 1e shows the temporal alignment between radar-captured mechanical motion and ECG signals. Both the phase and I/Q components of the filtered radar data at the optimal spatial point exhibit clear synchronization with the reference ECG, providing physiological evidence for inferring electrical activity from mechanical motion. Figure 1f further presents a contactless 12-lead ECG reconstructed from a patient with atrial fibrillation (AF), accurately capturing the irregular and rapid atrial activity characteristic of this condition.

Reconstruction performance of 12-lead ECG

We first evaluated the performance of 12-lead ECG reconstruction. As shown in Fig. 2a, a standard 12-lead ECG captures cardiac electrical activity from multiple perspectives through limb leads (I, II, III, aVR, aVL, aVF) and precordial leads (V1–V6). Limb leads measure the frontal plane axis, while precordial leads record horizontal plane activity. Unlike single-lead approaches, our system reconstructs the full spatiotemporal panorama of cardiac motion. Figure 2b shows the correlation distribution between reconstructed and reference ECGs using Kernel Density Estimate (KDE) and Cumulative Distribution Function (CDF). The KDE peaks at 0.80 with an interquartile range of 0.14 (25th: 0.69, 75th: 0.83), indicating strong agreement between reconstructed and gold-standard signals. Figure 2c presents the detailed performance of the predicted ECG, evaluated using correlation, RMSE, and IBI error. For correlation, we compare the results from a window-based approach and a sample-based approach. The window-based method uses a 2.56-s window, approximately covering 3–4 heartbeats, while the sample-based method follows the standard 10-se duration typically used in routine ECG examinations. Sample-based correlation methods exhibit heightened sensitivity to localized data anomalies (e.g., transient failures), as their covariance estimation process nonlinearly amplifies such perturbations’ impact on holistic statistical measures. The proposed method achieves a mean window-based correlation of 0.75 (median: 0.77), with the same performance observed in sample-based evaluations at 0.75 (median: 0.77). These results not only validate the method’s capability to estimate ECG signals accurately but also underscore its robustness and consistency. For RMSE, the method achieves a mean error of 0.09 mV and a median of 0.07 mV. In IBI error, it achieves a mean of 23.39 ms and a median of 5.60 ms. Considering the 100 Hz recording rate (10 ms resolution), the median IBI error of 5.60 ms indicates that most predictions closely match the ground truth. Figure 2d illustrates the reconstruction results for an individual exhibiting premature atrial contractions (PAC). The system accurately tracks transient fluctuations in heartbeat intervals, with RR interval errors typically within 10 ms–20 ms. These deviations correspond to prolonged subsequent RR intervals consistent with PAC patterns. Across the reconstructed 12 leads, mean absolute errors remain consistently low, and even in the magnified Lead II trace, waveform morphology and timing closely match the reference ECG, confirming the fidelity of the reconstruction. Comparative results against baseline methods and additional detailed statistical analyses are provided in Supplementary Note 3.

Fig. 2: Overall performance of reconstructed 12-lead ECG.
Fig. 2: Overall performance of reconstructed 12-lead ECG.The alternative text for this image may have been generated using AI.
Full size image

a The 12-lead ECG records cardiac electrical vectors from various angles across the frontal plane (using limb leads I, II, III, aVR, aVL, aVF) and horizontal plane (through precordial leads V1-V6). b Correlation distribution with KDE peak at 0.80 and IQR = 0.14 (P25 = 0.69, P75 = 0.83), shown with KDE (blue) and CDF (red). c Different metrics for reconstructed 12-lead ECG. d An example of reconstructed 12-lead ECG.

Evaluation of prediction robustness across demographics, pathologies, and leads

We evaluated the reconstruction performance across demographic groups, cardiac pathologies, and ECG leads to assess the robustness and generalizability of the proposed approach under diverse conditions.

Figure 3 a shows the prediction performance across age and gender groups. The highest correlations are observed in younger individuals, likely due to stronger cardiac motion and higher signal quality during early adulthood38. With aging, cardiac contraction weakens, leading to a gradual decline in signal strength and correlation performance. In contrast, RMSE and IBI errors show minimal age dependence, as they are primarily influenced by high-amplitude peaks with strong signal-to-noise ratios (SNR). Overall, performance remains stable across demographics, with the largest correlation reduction of 8.1% across age groups and 3.8% between genders, indicating robust generalization across populations. Figure 3b illustrates the prediction performance across different cardiac conditions. Ventricular premature beats (VPB), complete right bundle branch block (CRBBB), and AF eshow the lowest performance, largely due to their limited representation in the dataset (each comprising less than 2.5% of samples) and their distinct signal morphologies. VPB is associated with high-amplitude QRS signals, which disproportionately increase RMSE and result in higher prediction errors. CRBBB involves conduction abnormalities that typically do not lead to significant mechanical changes, making radar-based reconstruction more challenging. AF is characterized by chaotic electrical activity and irregular rhythms39, which will lower the SNR and complicate the prediction accordingly. Despite these variations, the method maintains performance comparable to normal sinus rhythm across most abnormal ECGs, demonstrating robust adaptability under pathological conditions. Figure 3c compares reconstruction performance across ECG leads. Lower accuracy is observed for leads III, aVL, and aVF. As shown in Fig. 2a, this is primarily because these leads are more perpendicular to the direction of electrical activity propagation along the long axis of the left ventricle (from the apex to the base), resulting in inherently weaker signal amplitudes. Additionally, while leads III and aVF are aligned with the direction of electrical activity in the inferior wall, their accuracy is affected by the influence of diaphragmatic motion and the volume conductor effects of abdominal organs. This observation is consistent with findings from autoencoder models trained on ECG signals, suggesting that the intrinsic characteristics of these leads play a significant role in prediction variability40. Nonetheless, the statistical analysis, with over 97% of samples showing a p-value below 0.001, underscores the robustness of the method in reconstructing ECG signals across leads despite inherent variability.

Fig. 3: Prediction performance across demographics, cardiac conditions, and ECG leads.
Fig. 3: Prediction performance across demographics, cardiac conditions, and ECG leads.The alternative text for this image may have been generated using AI.
Full size image

a Performance varies by age and gender, with male youth and females in prime working age achieving the highest correlation. Both metrics remain stable across groups. b Performance across cardiac conditions shows lower results for VPB, CRBBB, and AF due to their limited number of samples, but predictions still remain comparable to others. c Performance across different leads with strong statistical reliability across leads (P < 0.001 for 97% of samples).

Demonstration of diagnostic capability across cardiac abnormalities

To demonstrate the system’s capability to preserve key electrophysiological signatures essential for clinical diagnosis, we analyzed serveral representative ECG cases covering rhythm-, waveform-, and conduction-related abnormalities in Fig. 4.

Fig. 4: Examples of reconstructed ECG signals for various abnormalities.
Fig. 4: Examples of reconstructed ECG signals for various abnormalities.The alternative text for this image may have been generated using AI.
Full size image

a SB: accurate rhythm abnormality detection with predicted IBIs closely matching ground truth. b STach: consistent rhythm tracking with accurate IBI estimation. c AF: irregular RR intervals, absent P waves, and undulating f-wave baseline. d STC: horizontal ST depression reflecting ischemia. e TAb: bifid T-wave morphology in Lead III. f CRBBB: RSR' morphology and prolonged QRS reproduced in Lead V3. g Feature separability: radar- and ECG-derived signals show comparable performance.

Rhythm-related abnormalities

We first examined rhythm-related abnormalities, including sinus tachycardia (STach), sinus bradycardia (SB), and atrial fibrillation (AF). STach is defined by a sinus rhythm with a heart rate above 100 bpm, while SB is characterized by a sinus rhythm below 60 bpm41,42. Both are common arrhythmias of sinus origin and can be either normal or pathological. For example, intense physical activity can cause the heart rate to exceed 100 bpm, whereas well-trained athletes typically have resting heart rates below 60 bpm. On the other hand, STach is often seen in individuals with congestive heart failure, while SB is frequently present during the early stages of acute myocardial infarction. This emphasizes the need for continuous monitoring to distinguish normal variations from pathological conditions. Figure 4a, b illustrate examples of SB and STach. The predicted IBIs are approximately 1 s for SB and 0.5 s for STach, closely matching the ground truth ECG. Additionally, since these two types are primarily related to the depolarization frequency of the sinus node, they are relatively easy to capture.

AF is one of the most common and clinically significant sustained arrhythmia, increasing the risk of ischemic stroke by 4–5 times, with an estimated 11% of cases remaining undiagnosed43. In ECG, AF is recognized by irregular QRS complexes without discernible P waves. It may often be accompanied by fine fibrillatory waves (f-waves) that create an undulating baseline due to chaotic atrial electrical activity. Figure 4c displays a typical case of AF in V1, where irregular RR intervals can be observed through the IBI, and the predicted ECG shows an absence of P waves. Furthermore, the baseline in the predicted ECG appears slightly undulating, which indicates the presence of f-waves between the R-peaks. Due to the limited amount of AF data and its chaotic rhythm, particularly the unpredictable f-waves caused by multiple tiny reentrant circuits, the average performance is slightly lower than in other cases. However, its key characteristics can still be captured.

Waveform-related abnormalities

We next evaluated waveform-related abnormalities, focusing on ST-segment changes (STC) and T-wave abnormalities (TAb). These categories represent critical electrocardiographic manifestations of myocardial repolarization disturbances with distinct pathophysiological and diagnostic implications.

ST-segment depression, one of the most prevalent ST-segment abnormalities on ECG, manifests as a downward deviation of the ST segment below the isoelectric baseline. Clinically, it is strongly associated with myocardial ischemia, serving as a hallmark finding in conditions such as coronary artery disease. Additionally, it may reflect ventricular overload, electrolyte disturbances, or medication effects. ST depression is categorized into three morphological patterns: horizontal, downsloping, and upsloping, each carrying distinct diagnostic implications. Figure 4d demonstrates horizontal ST depression, where the algorithmically reconstructed ECG aligns precisely with the ground truth, faithfully replicating both the magnitude and morphology of the depressed ST segment.

TAb encompasses various morphological alterations, including low amplitude, inversion, and structural distortions. A representative example in Fig. 4e demonstrates a distinctly abnormal T-wave configuration in Lead III, characterized by bifid morphology with separated peaks. This repolarization abnormality may be associated with myocardial ischemia, hypokalemia, or autonomic dysregulation, which requires further clinical evaluation to exclude acute coronary syndrome.

The findings demonstrate a strong capability in detecting myocardial repolarization abnormalities, such as horizontal ST-segment depression and bifid T-waves, with clinical-grade accuracy. This enables comprehensive cardiac monitoring that extends beyond rhythm analysis and may provide valuable support for ECG-based diagnostics in relevant clinical contexts.

Conduction-related abnormalities

To evaluate conduction-related abnormalities, we focus on CRBBB, a clinically prevalent conduction disorder. CRBBB results from impaired right ventricular activation due to right bundle branch obstruction. The hallmark ECG feature of CRBBB is the presence of an RSR’ pattern (characterized by a secondary R’ wave), accompanied by a QRS duration ≥ 120 ms. As shown in Fig. 4f, our system’s reconstructed Lead V3 tracing demonstrates precise reconstruction fidelity for CRBBB, accurately reproducing both the diagnostic R’ wave morphology and QRS prolongation.

These findings validate the technology’s capacity to resolve critical conduction system abnormalities, which are often difficult to capture due to the subtle internal cardiac changes involved. This further demonstrates its potential as a valuable tool for comprehensive cardiac diagnostics.

Quantitative diagnostic accuracy

We further quantified the diagnostic reliability of the reconstructed ECGs by evaluating their ability to support automated disease classification. As shown in Fig. 4g, features derived from radar and ECG signals were independently extracted and assessed using a classification model. Radar-derived features performed comparably to those obtained from reference ECGs across most abnormality types. Notably, for sinus tachycardia, radar-based features even outperformed ECG-derived ones, reflecting the system’s high sensitivity to rhythmic mechanical variations. In contrast, slightly lower performance was observed for CRBBB, likely due to its subtle mechanical manifestations and limited training samples. To further assess clinical applicability, we evaluated a lightweight neural network trained on the reconstructed ECGs for atrial fibrillation versus normal sinus rhythm classification. The model demonstrated strong performance, achieving an ROC AUC of 0.997, an average precision of 0.896, an accuracy of 0.989, and an F1-score of 0.813. The corresponding sensitivity and specificity were 0.867 and 0.993, respectively (see Supplementary Note 4), confirming that the reconstructed signals retain sufficient diagnostic information for automated analysis. These findings further support the robustness of the system and its potential as a contactless, general-purpose 12-lead ECG platform for comprehensive cardiovascular monitoring.

Long-term contactless daily life ECG monitoring

To assess the feasibility and significance of long-term ECG monitoring, we conducted a focused evaluation during the overnight sleep of a healthy adult male in a home environment, using Holter ECG as the ground truth reference. To ensure reliable overnight assessment, we applied an automated output-based artifact exclusion procedure that removes segments in which the predicted ECG amplitude remains continuously below a predefined threshold, as such low-amplitude periods are often associated with motion-related artifacts. Increasing the threshold results in more segments being excluded while retaining higher-quality predictions. The excluded segments typically occur as short, intermittent intervals rather than long continuous stretches. To balance data retention and signal reliability, we selected a threshold of 0.2 mV, which corresponded to excluding approximately 10% of the data and retaining 4.8 h of valid monitoring. Additional details are provided in Supplementary Note 5.

Figure 5a presents the reconstructed ECG signals over the entire night of monitoring. We separately show a limb lead (II) and a precordial lead (V4), along with the correlation coefficient at each moment (averaged over 2 s). The mean and median correlation for Lead II are 0.81 and 0.84, respectively, while for V4, they are 0.85 and 0.89. These values demonstrate the stable performance of the system throughout the night. Additionally, we present the variation of IBI across the night. As illustrated in Fig. 5b, the predicted IBI closely tracks the trend of the true IBI and effectively captures the transient IBI changes during sleep. The mean IBI value is 23.67 ms, while the median is 10.00 ms. An additional example is provided in Supplementary Note 6, demonstrating that the proposed method remains functional even when certain leads are lost. Figure 5c shows the correlation between the reconstructed and ground truth ECG across different leads. Our model achieved a mean correlation of 0.81 and a median of 0.84, demonstrating strong performance across all leads. Figure 5d presents an example of a 20-s segment of reconstructed ECG signals. The algorithm effectively reconstructs ECG waveforms across different leads, including those exhibiting subtle morphological deviations (Supplementary Note 7).

Fig. 5: Long-term monitoring performance.
Fig. 5: Long-term monitoring performance.The alternative text for this image may have been generated using AI.
Full size image

a Reconstructed ECG signals over an entire night of monitoring for a 32-year-old male, showing a limb lead (II) and a precordial lead (V4), with momentary correlation coefficients averaged over 2 s. b Predicted IBI closely matches the true IBI, capturing transient changes during sleep. c Correlation between reconstructed and ground truth ECG across leads, with a mean correlation of 0.81 and a median of 0.84. Modified limb leads used by the Holter ECG result in amplitude differences compared to reconstructed signals. d An example of a 20-s segment of reconstructed ECG signals.

These results suggest the potential of our method to enable cardiac monitoring in challenging home environments. They should be regarded as proof-of-concept demonstrations that contactless radar-based ECG can sustain long-term, continuous monitoring in a healthy subject, while its broader applicability, including for chronic disease management and remote healthcare, remains to be established through systematic validation in larger and more diverse populations.

Discussion

This study presents a novel contactless 12-lead ECG system that integrates radar sensing with deep learning to reconstruct cardiac electrical activity from radio-frequency signals with high fidelity. Validated on a large cohort of 6974 individuals, the system achieved a mean Pearson correlation of 0.75 and RMSE of 0.09 mV against gold-standard ECG, while reliably detecting arrhythmias and other cardiac abnormalities. It enables reliable overnight monitoring without the discomfort or signal loss typical of contact-based systems, supporting strong potential for home and clinical use.

Further analysis in Supplementary Note 3 shows that the reconstruction preserves more than R-peak activity. Component-wise suppression produces distinct and physiologically consistent changes when P, R, or T waves are attenuated, indicating that the model draws on information beyond beat timing. Additional interval-level evaluation, including comparison with existing wearable devices, demonstrates high accuracy in heart rate and RR estimation, while QT intervals, despite lower correlation, exhibit small mean differences. These results suggest that the system retains meaningful waveform and interval characteristics rather than merely detecting heartbeats.

Despite these strengths, the reconstructed signals exhibit lead-dependent variation, particularly in those less aligned with the dominant cardiac vector, and rare pathologies are less represented, underscoring the need for continued evaluation across broader clinical conditions. Long-term monitoring studies are also required to assess performance and stability over time, extending the present proof-of-concept to real-world deployment contexts.

To contextualize the system’s translational potential, we outline here a feasible pathway for regulatory validation and clinical deployment. A three-stage process is envisioned. The first stage focuses on preliminary safety and electromagnetic interference verification, together with assessments of wireless reliability and data integrity, to ensure stable long-term operation. The second stage consists of single-center prospective studies designed to evaluate day-to-day stability, workflow compatibility, and agreement with standard 12-lead ECG measurements. The third stage expands to multi-center studies to examine cross-site generalizability across diverse populations and clinical environments. Long-term single-center reliability assessment is currently underway, and multi-center evaluations are planned for future development.

Regarding system integration, the device’s wireless module already supports real-time transmission of raw radar data to a server-based reconstruction backend, enabling reconstructed ECG signals to be forwarded to hospital telemetry systems or IoT hubs through a lightweight gateway. For future deployment, an important objective is to progressively localize the reconstruction pipeline so that the system no longer depends on centralized computing resources. Integration with existing ECG infrastructures will also require alignment of data formats and communication protocols, as well as the implementation of appropriate data-security and access-control mechanisms. These considerations outline a practical route for incorporating the system into both hospital monitoring workflows and home-based IoT environments.

Methods

Radar principle and signal processing

Our radar system captures chest movements by analyzing phase and amplitude variations in reflected radio signals44,45. Using a 3-transmit, 4-receive MIMO FMCW radar (Fig. 1), the transmitted chirps follow a linearly increasing frequency:

$$f(t)={f}_{\!0}+\frac{B}{{T}_{{\rm{chirp}}}}t.$$
(1)

where f0 is the initial frequency, B is the bandwidth, and Tchirp is the chirp duration. The received signals are mixed with the transmitted signals to produce an intermediate frequency (IF) signal, encoding chest displacement r(t) as:

$${\phi }_{{\rm{IF}}}(t)=4\pi \frac{r(t)+d}{\lambda },$$
(2)

where d is the subject’s distance, λ is the wavelength, and r(t) contributes to phase variations. Subtle movements are extracted by unwrapping the phase. The on-chip Fast Fourier Transform (FFT) processes the data and retains Nrng = 15 range bins for efficiency. The radar data are then reconstructed into spatial reflections (r, θ, ϕ) using 3D beamforming:

$${{\bf{S}}}_{t}(r,\theta ,\phi )={{\boldsymbol{a}}}_{t}(r,\theta ,\phi )\exp \left(j{{\boldsymbol{\omega }}}_{t}(r,\theta ,\phi )\right)\in {{\mathbb{C}}}^{{N}_{{\rm{rng}}}\times {N}_{{\rm{azi}}}\times {N}_{{\rm{ele}}}},$$
(3)

where at and ωt respectively encode the amplitude and phase information. Nazi and Nele represent the number of spatial angular samples in the azimuth and elevation directions, respectively. Chest movements, primarily driven by respiration, modulate both ωt and at, while heartbeat signals introduce subtle, higher-frequency variations.

A noise-robust differentiator46 enhances heartbeat signals by filtering low-frequency components and amplifying high-frequency features:

$${f}^{{\prime\prime} }=\frac{\left(\,{f}_{3}+{f}_{-3}\right)+2\left(\,{f}_{2}+{f}_{-2}\right)-\left(\,{f}_{1}+{f}_{-1}\right)-4{f}_{0}}{16{\Delta }^{2}}.$$
(4)

This improves the visibility of R peaks, even under noise, as shown in Fig. 1e. We apply filtering along the unwrapped phase ωt and the I/Q components of St to create multi-channel data.

The processed radar data is represented as a 4D tensor \({\bf{X}}\in {{\mathbb{R}}}^{C\times {N}_{{\rm{rng}}}\times {N}_{{\rm{azi}}}\times {N}_{{\rm{ele}}}\times T}\), where C = 3 corresponds to the phase and I/Q components, and T denotes the temporal dimension. Concurrently, synchronized ECG signals \({\boldsymbol{y}}\in {{\mathbb{R}}}^{L\times T}\) are captured, with L = 12 representing the number of leads. The goal is to learn a mapping from X to y, enabling contactless sensing of cardiac activity.

Deep learning models

Our framework employs a learning-based prediction model to map radar signals X to synchronized 12-lead ECG signals y.

To enhance robustness, we pretrain a VQ-VAE model47,48 that encodes ECG signals into a hierarchical latent space with multiple levels of abstraction. The VQ-VAE is pretrained exclusively on external data (Supplementary Dataset 1), using six publicly available ECG datasets: CPSC 201849, PTB-XL50, G12EL51, ECG-Arrhythmia52, SPH53, and CODE-1554. No ECG recordings from our study cohort are used during pretraining, ensuring that the VQ-VAE learning process is entirely independent of the test subjects.

Given the ECG signal \({\boldsymbol{y}}\in {{\mathbb{R}}}^{L\times T}\), the model encodes it into multi-scale latent representations \({{\boldsymbol{z}}}^{(l)}\in {{\mathbb{R}}}^{D\times \frac{T}{4}}\) and \({{\boldsymbol{z}}}^{(h)}\in {{\mathbb{R}}}^{D\times \frac{T}{8}}\), whose column vectors are then further quantized into discrete vectors drawn from learned codebooks. The VQ-VAE also includes a decoder that can reconstruct y from the quantized multi-scale latent representations \({{\boldsymbol{q}}}^{(l)}\in {{\mathbb{R}}}^{D\times \frac{T}{4}}\) and \({{\boldsymbol{q}}}^{(h)}\in {{\mathbb{R}}}^{D\times \frac{T}{8}}\).

Our radar model also adopts an encoder-decoder architecture, similar to the pre-trained VQ-VAE model. The encoder encodes 4D radar data into a latent space, while the decoder reconstructs it into the 12-lead ECG.

Voxel encoder

The voxel encoder processes radar signals Xt at each timestamp t to extract spatial features:

$${{\boldsymbol{h}}}_{t}^{(r)}={{\mathcal{F}}}_{r}({{\bf{X}}}_{t})\in {{\mathbb{R}}}^{D},t\in [T],$$
(5)

where \({{\mathcal{F}}}_{r}(\cdot )\) uses 3D convolution layers, each reducing spatial resolution of the feature map while increasing feature channels.

Temporal encoder. The temporal dynamics of h(r) are encoded through a multi-scale convolutional architecture:

$${{\boldsymbol{h}}}^{(l)}={{\mathcal{F}}}_{\!l}({{\boldsymbol{h}}}^{(r)})\in {{\mathbb{R}}}^{D\times \frac{T}{4}},\quad {{\boldsymbol{h}}}^{(h)}={{\mathcal{F}}}_{h}({{\boldsymbol{h}}}^{(l)})\in {{\mathbb{R}}}^{D\times \frac{T}{8}},$$
(6)

where temporal resolutions are reduced by factors of 4 and 8, respectively. The temporal encoder further enhances feature representations through Transformer-based55 context modeling:

$${{\boldsymbol{v}}}^{(l)}={{\mathcal{G}}}_{l}({{\boldsymbol{h}}}^{(l)})\in {{\mathbb{R}}}^{D\times \frac{T}{4}},\quad {{\boldsymbol{v}}}^{(h)}={{\mathcal{G}}}_{h}({{\boldsymbol{h}}}^{(h)})\in {{\mathbb{R}}}^{D\times \frac{T}{8}},$$
(7)

where \({{\mathcal{G}}}_{l}\) and \({{\mathcal{G}}}_{h}\) are Transformer encoders to establish global temporal dependencies at respective scales. Each Transformer encoder consists of 8 layers with an embedding size of 128, where this relatively compact dimension was found to provide stable training and robust reconstruction. These refined features form the basis for subsequent ECG reconstruction.

Decoder

We first transform the latent features v(l) and v(h) as follows:

$${\tilde{{\boldsymbol{q}}}}^{(h)}={{\bf{W}}}_{h}{{\boldsymbol{v}}}^{(h)}+{{\boldsymbol{b}}}_{h},\quad {\tilde{{\boldsymbol{q}}}}^{(l)}={{\bf{W}}}_{l}\left[{\tilde{{\mathcal{D}}}}_{h}({\tilde{{\boldsymbol{q}}}}^{(h)})\oplus {{\boldsymbol{v}}}^{(l)}\right]+{{\boldsymbol{b}}}_{l},$$
(8)

where {Wh, bh, Wl, bl} are linear projection parameters, and ⊕ represents channel-wise concatenation. These features are then used for 12-lead ECG reconstruction:

$$\tilde{{\boldsymbol{y}}}={\tilde{{\mathcal{D}}}}_{l}\left({\tilde{{\boldsymbol{q}}}}^{(l)}\oplus {\mathcal{U}}\left({\tilde{{\boldsymbol{q}}}}^{(h)}\right)\right),$$
(9)

where \({\mathcal{U}}\) is upsampling. Both \({\tilde{{\mathcal{D}}}}_{h}\) and \({\tilde{{\mathcal{D}}}}_{l}\) are stacked convolutional layers that follow the same structure as the decoder of the pre-trained VQ-VAE model.

Objective function

The radar model minimizes the reconstruction error between the predicted ECG signals \(\tilde{{\boldsymbol{y}}}\) and the ground truth y using an L1 loss:

$${\ell }_{{\rm{rec}}}=\parallel \tilde{{\boldsymbol{y}}}-{\boldsymbol{y}}{\parallel }_{1}.$$
(10)

Additionally, a regularization term aligns radar features with the pre-trained ECG latent representations:

$${\ell }_{{\rm{reg}}}={\lambda }_{h}\parallel {\tilde{{\boldsymbol{q}}}}^{(h)}-{{\boldsymbol{q}}}^{(h)}{\parallel }_{2}+{\lambda }_{l}\parallel {\tilde{{\boldsymbol{q}}}}^{(l)}-{{\boldsymbol{q}}}^{(l)}{\parallel }_{2},$$
(11)

where λh and λl control the regularization strength, q(h) and q(l) are quantized latent representations for the ECG signal, enforcing feature-level alignment. The overall objective function is:

$${\mathcal{L}}=\parallel \tilde{{\boldsymbol{y}}}-{\boldsymbol{y}}{\parallel }_{1}+{\lambda }_{h}\parallel {\tilde{{\boldsymbol{q}}}}^{(h)}-{{\boldsymbol{q}}}^{(h)}{\parallel }_{2}+{\lambda }_{l}\parallel {\tilde{{\boldsymbol{q}}}}^{(l)}-{{\boldsymbol{q}}}^{(l)}{\parallel }_{2},$$
(12)

where the L1 (MAE) loss was used to preserve sharp ECG morphology, as L2 (MSE) tends to over-smooth. The regularization weights λh and λl were fixed at 1.0 to balance the loss terms.

Training details

Training is performed for 500 epochs using the AdamW optimizer56 with a learning rate of 5 × 10−4, weight decay of 1 × 10−5, and a batch size of 32. The regularization coefficients are set to λh = λl = 1. Each input segment contains T = 2048 samples. To avoid subject overlap, we adopt five-fold cross-validation at the recording level. All experiments are conducted on a single NVIDIA RTX 4090 GPU, and each fold requires approximately 3.3 h to complete 500 training epochs. Additional implementation details are provided in Supplementary Note 8.

Evaluation metrics

To account for the sub-second timing discrepancies between radar and hospital-grade ECG signals, waveform alignment is evaluated using maximum normalized cross-correlation (NCC) between the predicted signal x and ground truth y as follows:

$$\,{\text{Correlation}}\,(\tau )=\frac{{\sum }_{t\in T}(y[t+\tau ]-{\bar{y}})(x[t]-{\bar{x}})}{\sqrt{{\sum }_{t\in T}{(y[t+\tau ]-{\bar{y}})}^{2}{\sum }_{t\in T}{(x[t]-{\bar{x}})}^{2}}},$$
(13)

where the maximum value across all lags τ is reported. Waveform deviation is further quantified using the root mean square error (RMSE).

For Inter-Beat Interval (IBI) analysis, R-peak locations are detected with NeuroKit257 using Lead-II as the reference, with IBI sequences computed as the time differences between consecutive R-peaks. Temporal consistency is assessed using the Mean Absolute Error (MAE) between the predicted and ground truth IBI sequences.

For disease prediction tasks, we evaluate the model’s performance using the Area Under the Receiver Operating Characteristic Curve (AUC). AUC quantifies the model’s ability to distinguish between positive and negative cases by plotting the True Positive Rate (TPR) against the False Positive Rate (FPR) at various classification thresholds.

Ethics approval

This project was conducted following approval from the Medical Research Institutional Review Board (IRB) of the First Affiliated Hospital of USTC. All study participants provided informed consent for the use of their data in scientific research and were informed of their right to withdraw consent at any time.