Abstract
Cardiovascular disease (CVD) remains the leading cause of global mortality, highlighting the need for scalable, continuous, and contactless diagnostics. Traditional electrocardiogram (ECG) systems, though essential for detecting cardiac abnormalities, are constrained by contact-based, intermittent measurements. In this work, we introduce CardioRadar, a clinically validated radar-based system for contactless 12-lead ECG monitoring. Compact custom radar hardware captures chest micro-motions induced by cardiac activity, while data-driven inverse modeling reconstructs electrical signals from mechanical motion. Validated in 6,974 individuals across diverse clinical settings, CardioRadar achieves a mean correlation of 0.75 (median: 0.77), RMSE of 0.09 mV (median: 0.07 mV), and inter-beat interval error of 23.39 ms (median: 5.60 ms) compared to ground-truth ECG. It also reliably detects arrhythmias and supports continuous overnight monitoring. CardioRadar represents a transformative, low-cost solution for equitable, long-term cardiovascular care.
Introduction
The 12-lead electrocardiogram (ECG) is the gold standard for diagnosing a wide range of cardiac conditions, from arrhythmias to myocardial ischemia. However, conventional ECG systems require skin-contact electrodes, which can cause skin irritation, require clinical expertise to place, and are impractical for long-term, continuous monitoring. While wearable devices like smartwatches have improved accessibility, they typically provide only single-lead data, lacking the comprehensive diagnostic depth of a full 12-lead setup.
CardioRadar addresses these limitations by introducing a contactless, multi-lead monitoring system. By leveraging 60 GHz millimeter-wave radar and deep learning, the system captures subtle chest micro-motions and "translates" them into electrical signals, providing a comprehensive "3D view" of the heart without physical contact.
Methods and System Architecture
Hardware Design
The CardioRadar hardware is a compact, circular device (6 cm diameter) designed for low-cost scalability, with a total component cost under $30. It utilizes a 60 GHz Frequency Modulated Continuous Wave (FMCW) radar equipped with a 3×4 MIMO antenna array. This setup allows for sub-millimeter precision in capturing chest vibrations from a distance of approximately 0.5 meters.
Signal Processing and Deep Learning Pipeline
The system employs a sophisticated "inverse modeling" AI pipeline to bridge the gap between mechanical motion and electrical activity:
- 3D Beamforming: Isolates cardiac-induced micro-motions from larger physiological signals like respiration.
- VQ-VAE Pre-training: A Vector Quantized Variational Autoencoder was trained on over 500,000 public ECG records to learn the fundamental "latent language" of cardiac electrical signals.
- Radar-to-ECG Translation: A specialized model, utilizing a Voxel Encoder for spatial features and Transformers for temporal dependencies, maps the mechanical radar signals into the pre-learned ECG latent space.
- Waveform Reconstruction: A decoder transforms these latent representations into high-fidelity 12-lead waveforms.
Experimental Validation and Results
Clinical Performance
The system was validated on a massive cohort of 6,974 individuals, one of the largest studies of its kind. CardioRadar demonstrated clinical-grade accuracy with a mean Pearson correlation of 0.75 and a Root Mean Square Error (RMSE) of 0.09 mV. Timing precision was also high, with a median inter-beat interval (IBI) error of only 5.60 ms, making it suitable for heart rate variability (HRV) analysis.
Diagnostic Utility
Beyond simple heart rate monitoring, CardioRadar successfully identified various clinical abnormalities. The system detected Atrial Fibrillation (AF), Sinus Tachycardia, and subtle ST-segment depressions, which are critical indicators of myocardial ischemia. Long-term monitoring capabilities were further demonstrated through successful overnight sleep studies using Holter ECGs as ground truth.
Discussion and Future Work
Mechano-Electrical Coupling
The success of CardioRadar relies on the physiological link between electrical excitation and mechanical contraction. However, the researchers noted lead-dependent accuracy; leads aligned with the heart's dominant electrical vector (e.g., Lead II) were more accurately reconstructed than those perpendicular to it (e.g., Lead III). This suggests physical limits to sensing electrical activity via motion that are inherent to the heart's orientation.
Limitations and Outlook
While the system effectively filters respiration, large body movements remain a challenge for signal integrity. Future iterations aim to improve motion robustness and localize the processing pipeline to the device itself, reducing the need for high-bandwidth server connections. CardioRadar represents a significant step toward equitable, long-term cardiovascular care, offering a diagnostic tool that is both highly capable and accessible for low-resource settings.
Summary updated