A wearable cardiac ultrasound imager

The study addresses a long-standing gap in cardiovascular monitoring: current non-invasive imaging modalities (for example, MRI, CT and conventional ultrasound probes) are bulky and unsuitable for continuous monitoring, while existing wearable devices measure only surface signals (ECG, PPG) and cannot visualize cardiac structure or function directly. The authors aim to enable continuous, real-time echocardiographic imaging and automated assessment of cardiac function indices (stroke volume, cardiac output, ejection fraction) during daily activities and exercise. They propose a soft, skin-conformal, wearable ultrasound imager with orthogonal transducer arrays and improved skin coupling, combined with a deep learning pipeline to automatically extract left ventricular volume from continuous image streams. The purpose is to deliver clinically relevant echocardiographic views and quantitative metrics comparable to commercial systems, but in a truly wearable, motion-tolerant form factor, thus expanding echocardiography to continuous monitoring, stress testing during activity, and broader outpatient applications.

The paper situates its contribution against prior non-invasive imaging modalities that are not conducive to continuous monitoring due to size and operational constraints (MRI, CT, nuclear imaging) and the new generation of handheld/POCUS devices that still require an operator to hold the probe. Wearable technologies to date largely capture skin-surface signals (ECG, ballistocardiography, PPG, impedance tomography) and cannot image cardiac structures or measure volumetric indices directly. Conventional stress echocardiography captures images before and after exercise, missing transient abnormalities during activity and relying on subjective endpoints. Prior ultrasound array and beamforming strategies (plane-wave, mono-focus, compounding) inform the imaging design; coherent compounding is known to boost frame rate and image quality. Standard echocardiographic practice emphasizes orthogonal views and left-ventricular segmentation models (17-segment), providing a clinical framework for wall motion assessment. Finally, literature highlights inter- and intra-observer variability in manual contouring and ejection fraction estimation, motivating automated analysis via deep learning.

Device design and materials: The wearable imager integrates two orthogonally oriented piezoelectric transducer arrays on a soft SEBS substrate to obviate manual probe rotation and expand the sonographic window. Each element uses an anisotropic 1-3 piezoelectric composite with a silver-epoxy backing. The center frequency is 3 MHz (balancing penetration and resolution), with 0.4 mm pitch (~0.78 λ) to enhance lateral resolution and suppress grating lobes. Multilayer, high-density stretchable electrodes are fabricated from eutectic gallium-indium liquid metal dispersed in SEBS, patterned to ~30 µm minimum widths, and layered for interconnects and electromagnetic shielding to reduce RF noise. The composite electrode is ~8 µm thick. Lap shear testing shows bonding strengths of ~250 kPa (transducer-SEBS) and ~236 kPa (transducer-electrode), exceeding commercial adhesives. The device exhibits low Young’s modulus (~921 kPa, comparable to skin), high stretchability (~110% biaxial), and mechanical compliance under bending, wrapping, poking, and twisting.

Beamforming and imaging characterization: Three transmit strategies were compared on wire and inclusion phantoms: plane-wave, mono-focus, and wide-beam compounding (multiple divergent transmissions at different angles, coherently combined). Receive beamforming further improves image quality. Wide-beam compounding yields synthetic focusing and elevated acoustic intensity across the insonation area, providing the best SNR and spatial resolution. Spatial resolutions (axial, lateral, elevational) were quantified from point spread functions; location accuracies were assessed by comparing imaged vs. ground-truth wire positions. Dynamic range and contrast-to-noise ratio (CNR) were quantified using cylindrical inclusions with varying acoustic impedances.

Chest curvature compensation: A 3D scanner captured each subject’s chest curvature. Element positional shifts within the soft array were compensated in transmit/receive beamforming to correct phase distortion from non-planar mounting, enabling standard echocardiographic views (apical four-, apical two-chamber; parasternal long- and short-axis).

Wear protocol and couplant: A liquid silicone couplant was used for long-duration coupling (less evaporation than aqueous gels). Continuous recordings were collected during rest, exercise (stationary bicycle, intervals to maximal heart rate), and recovery, with simultaneous ECG for correlation. Device temperature and skin tolerance were monitored; reproducibility was assessed across subjects.

Deep learning pipeline: Pre-processed apical four-chamber image sequences trained several segmentation models; the FCN-32 architecture was selected based on qualitative and quantitative performance. Data augmentation expanded the training set. The trained model segments LV contours and infers LV volume per frame; volumetric indices (EDV, ESV, stroke volume, cardiac output, ejection fraction) are derived via slice integration and geometry processing (including rotation alignment and PCA). During recovery when deep breathing intermittently occluded the heart, an imputation algorithm was used to fill blocked image segments before inference. Agreement between model outputs and manual labels was assessed with Bland–Altman analysis on wearable and commercial datasets.

Comparators and analyses: Imaging performance was benchmarked against a commercial ultrasound system across four standard cardiac views. Wall motion was analyzed via displacement waveforms from basal, mid-cavity, and apical short-axis slices (mapped to the 17-segment model). M-mode echocardiography was extracted to quantify LVIDd, LVIDs, fractional shortening, and to align mechanical activity with ECG phases.

• Mechanical/electrical performance: The stretchable liquid-metal/SEBS electrodes maintained conductivity up to ~750% tensile strain; device stretchability ~110% enables robust skin conformity. Bonding strengths were ~250 kPa (transducer–SEBS) and ~236 kPa (transducer–electrode). Overall modulus ~921 kPa (skin-like).
• Imaging strategy: Wide-beam compounding with receive beamforming produced the highest SNR and best spatial resolutions across depth. Axial resolution remained nearly constant with depth (set by frequency/bandwidth); lateral resolution degraded only slightly with depth; elevational resolution degraded with depth but was improved by integrating six small elements into a longer element for better beam convergence.
• Accuracy and sensitivity: Axial and lateral location accuracies were 96.01% and 95.90%, respectively. Dynamic range was 63.2 dB (exceeding the 60 dB diagnostic threshold). CNR ranged from 0.63 to 2.07 across inclusions; even ±3 dB contrast inclusions were clearly visualized.
• Clinical views and comparability: The wearable system produced apical four-, apical two-chamber, parasternal long- and short-axis views comparable to a commercial system. Short-axis imaging enabled myocardial wall displacement waveforms at basal, mid, and apical levels following the 17-segment model, and M-mode captured chamber dimensions and valve motion synchronized to ECG phases.
• Continuous monitoring during exercise: The wearable imager recorded uninterrupted M-mode during rest, exercise, and recovery with minimal motion artefacts. During exercise, LVIDd and LVIDs decreased as the septum and posterior wall moved closer to the probe; fractional shortening increased, then returned toward baseline during recovery.
• Automated LV volume and indices: The FCN-32 model generated LV volume waveforms closely matching manual labels for both wearable and commercial data. Bland–Altman analysis indicated mean differences of ~−1.5 ml for both systems with ~95% of points within 95% limits of agreement. Derived stroke volume, cardiac output, and ejection fraction showed no marked differences between wearable and commercial systems.
• Recovery dynamics: In recovery, EDV and ESV increased as heart rate decreased; the diastasis phase became more evident over time. Stroke volume rose from ~60 ml to ~70 ml; ejection fraction declined from ~80% to ~60%. Cardiac output decreased from ~11 l/min to ~9 l/min as the drop in heart rate (from ~175 to ~130 bpm) outweighed the modest increase in stroke volume.
• Overall, image quality and quantitative outputs from the wearable device were comparable to a commercial imager, while enabling continuous, hands-free acquisition during motion.

The findings demonstrate that a soft, skin-conformal ultrasound array with orthogonal transducer geometry and advanced beamforming can deliver high-quality echocardiographic images continuously, including during vigorous exercise. By compensating for chest curvature and leveraging wide-beam compounding, the system achieves high SNR, accurate localization, and clinically adequate dynamic range and CNR. Automated segmentation with an FCN-32 model yields beat-to-beat LV volumes and derived indices (SV, CO, EF) that agree closely with manual labels and commercial system results. Continuous monitoring reveals dynamic physiological trends (e.g., exercise-induced reductions in LV dimensions, recovery-associated changes in EDV/ESV, HR, SV, EF, and CO) that are often missed by traditional pre/post stress echocardiography. These capabilities address the research objective of enabling non-invasive, real-time, and operator-independent assessment of cardiac function in naturalistic settings. The technology has immediate relevance for stress testing, critical care, perioperative management, and athletic monitoring and could broaden echocardiography’s reach beyond the clinic. The concept generalizes to other deep tissues and procedures (e.g., IVC, aorta, spine, liver imaging, and ultrasound-guided interventions) by providing simultaneous orthogonal views and hands-free operation.

This work introduces a wearable cardiac ultrasound imager that maintains intimate skin coupling, achieves high-fidelity, continuous echocardiographic imaging from multiple standard views, and automatically outputs key cardiac performance metrics via deep learning. It matches commercial systems in image quality and quantitative indices while uniquely enabling uninterrupted monitoring during motion and stress. Future research should focus on: (1) enhancing spatial resolution by developing algorithms that compensate for time-varying (dynamic) chest curvature; (2) miniaturizing and integrating the back-end hardware to eliminate tethered connections; (3) improving model generalizability by expanding training datasets and exploring few-shot or reinforcement learning approaches; and (4) extending and validating applications to other organs and clinical workflows, including hands-free image-guided procedures.

• Dynamic curvature compensation: Current phase distortion correction uses static 3D chest curvature; it does not yet adapt to time-varying chest geometry during breathing and movement, which can limit peak spatial resolution.
• System integration: The wearable array relies on a flexible tether to an external back-end for data acquisition and processing; full miniaturization and on-board processing are future goals.
• Model generalizability: The FCN-32 segmentation model is presently validated on subjects similar to those in the training dataset; broader population generalization requires larger, more diverse data and/or advanced learning strategies.
• Posture constraints: Accurate LV dimension estimation is difficult in the standing position due to anatomical limitations and lung interference; image imputation was sometimes needed during deep breathing.
• Elevational resolution at depth: Elevational resolution degrades with depth, mitigated but not eliminated by element integration into longer apertures.