Heartbeat Vibrations Show Promise for Wearable Hemodynamic Monitoring, but Clinical Readiness Varies Sharply by Endpoint

Every heartbeat sends a faint mechanical shudder through the chest wall and the rest of the body. For more than two decades, researchers have tried to turn those vibrations into clinical numbers: stroke volume, cardiac output, ejection fraction, blood pressure. The appeal is obvious. Sensors based on microelectromechanical systems, or MEMS, cost pennies, need no trained operator, and could in principle track the heart’s pumping performance continuously during daily life, something no catheter, echocardiogram, or cardiac MRI can do. A new critical review published in the Annals of Biomedical Engineering by Peshala Gamage, Mehmet Kaya, and colleagues at the Florida Institute of Technology now delivers the field’s most sobering reality check, and its verdict is strikingly uneven: some applications are close to validation, while others remain years of rigorous testing away from any clinical claim.

The review organizes the sprawling literature by sensing modality. Seismocardiography, or SCG, records local translational accelerations of the sternum produced by valve motion and myocardial contraction. Ballistocardiography, or BCG, captures the whole-body recoil forces generated when the heart ejects blood into the aorta, and can be measured from a simple weighing scale, a bed frame, or a seat cushion. Gyrocardiography, or GCG, adds the rotational component of chest-wall motion, while kinocardiography, or KCG, fuses sternal and lumbar inertial sensors into a twelve-channel description of cardiogenic body motion. Each modality produces characteristic waveforms with fiducial points tied to specific cardiac events, such as aortic valve opening and closure, and these timing anchors form the raw material for every downstream estimation model.

The authors’ central methodological complaint is that the field has repeatedly confused physiological plausibility with clinical validation. Correlation coefficients, the most commonly reported statistic, are a particular target of their criticism. A device that reads one liter per minute too high in every patient can achieve a correlation of 0.95 with a reference method and still be clinically useless, because correlation measures linear association rather than agreement. The appropriate tool for method comparison is the Bland-Altman analysis, which reports mean bias and the 95 percent limits of agreement between the new method and the reference standard. The review also establishes a hierarchy of reference standards, giving the greatest weight to direct quantitative measurements such as the Direct Fick technique, thermodilution during right heart catheterization, invasive arterial pressure, and cardiac MRI-derived volumes, and treating echocardiographic and cuff-based comparators as weaker evidence.

By these standards, stroke volume and cardiac output estimation emerge as the field’s strongest suit. The most convincing evidence comes from a multi-signal cardiac scale validated against the Direct Fick method in 32 patients with unexplained dyspnea, which reported a stroke volume bias of just 1.58 milliliters with limits of agreement spanning roughly 40 milliliters, a cardiac output bias of 0.31 liters per minute, and a trending concordance of 96.7 percent, all without requiring per-patient invasive calibration. A separate deep learning study estimated cardiac output from 30-second triaxial SCG recordings combined with ECG and body mass index in heart failure patients undergoing right heart catheterization, reporting a bias of 0.35 liters per minute with tight limits of agreement. Both studies, however, were single-center and cross-sectional, and neither establishes that vibration-based monitoring outperforms established non-invasive alternatives in its intended setting.

The technical machinery behind these results is worth appreciating. Scale-based BCG pipelines filter the force channel in a narrow 1 to 15 hertz band for amplitude features, deliberately using linear-phase filters to avoid distorting the timing of fiducial points, and average eight to sixteen consecutive R-wave-triggered beats to suppress respiratory noise. Body weight and basal impedance serve as correction inputs that compensate for inter-individual differences in signal propagation, eliminating the need for a calibration catheterization in every new user. On the SCG side, convolutional neural networks process raw triaxial waveforms directly, learning temporal and cross-axis patterns without explicit annotation of valve events. Yet the review notes that even the best percentage errors hover near or above the 30 percent benchmark historically cited for interchangeability with thermodilution, a figure that itself reflects the uncertainty of the reference method rather than an absolute threshold of clinical meaningfulness.

Ejection fraction tells a very different story. Despite abundant publications, no study has reported Bland-Altman bias and limits of agreement for continuous EF estimation from vibration signals against echocardiographic or MRI-derived values. What does exist is phenotyping: kinocardiographic kinetic energy metrics, computed as time integrals of kinetic energy and maximum instantaneous power across twelve sensor channels, separated heart failure patients with reduced ejection fraction from those with preserved EF in a cohort of 126 patients, achieving an area under the curve above 0.85 and independent prognostic value. That is a clinically meaningful mechanical biomarker, the review argues, but some secondary literature has wrongly conflated it with continuous EF estimation, which it is not. A convincing EF estimator would need to specify target agreement bounds and validate them prospectively in cohorts spanning the clinical EF range, something that has never been done.

Early diastolic SCG features, which researchers hope could one day assess diastolic dysfunction non-invasively, remain even more exploratory. Correlations of roughly 0.71 between diastolic SCG morphology and tissue Doppler e-prime velocity suggest the signal genuinely encodes relaxation mechanics, and preload perturbation experiments show the waveform responds directionally to changes in venous return. But this sensitivity cuts both ways: the same features that track relaxation also respond to loading state independently of intrinsic diastolic function, a confound that any estimation model ignoring volume status would inherit. The review identifies multivariate models with loading-state covariates, tested in heart failure with preserved ejection fraction cohorts, as the obvious next experiment.

Blood pressure estimation earns the most cautious assessment of all. Vibration-based approaches replace the ECG R-wave with a BCG or SCG event as the proximal timing reference for pulse transit time, on the theory that this reduces the electromechanical delay ambiguity that limits conventional ECG-PPG methods. In practice, the vibration landmarks are themselves sensitive to preload and contractility, partially reintroducing the problem they were meant to solve. Worse, a study recording wrist BCG alongside invasive arterial pressure during exercise showed that posture, workload, and heart rate each independently disrupt the transit-time calibration, because sympathetic activation alters arterial compliance independently of pressure. No vibration-based BP study has exceeded 25 subjects, all recruited healthy volunteers, every method required per-subject calibration, and none approaches the ISO 81060-2 standard of a mean error of 5 millimeters of mercury or less with a standard deviation of 8 or less in at least 85 representative subjects.

Against this uneven landscape, one recent development stands out as the field’s most clinically ambitious step. The SEISMIC-HF I study evaluated a wearable patch combining seismocardiography, ECG, and photoplethysmography in 310 heart failure patients with reduced ejection fraction undergoing clinically indicated right heart catheterization, aiming to estimate pulmonary capillary wedge pressure, the invasive gold standard of cardiac filling pressure and congestion. On a held-out test set, predicted pressures correlated with catheterization values at 0.74, with a mean difference of about 1 millimeter of mercury and limits of agreement within roughly 10 millimeters. The review treats this as an emerging endpoint that extends vibration sensing from flow and timing toward pressure-guided heart failure management, while cautioning that ambulatory performance, longitudinal calibration stability, and outcomes-guided trials remain untested.

The review closes with a diagnosis that applies well beyond cardiac vibrations. Four technical problems recur across every modality: sensor placement, where shifting an accelerometer two or three centimeters can distort waveforms more than the hemodynamic signal itself; posture and respiration, which modulate beat-to-beat stroke volume by up to 15 percent; motion artifact, which during walking exceeds the cardiac signal by an order of magnitude; and calibration drift, whose behavior over weeks to months has never been prospectively quantified. The authors also flag a neglected engineering dimension, noting that deep learning pipelines requiring tens of millions of operations per inference can exceed the compute and battery budget of a wearable device, and that no reviewed study reports inference latency or energy consumption. Their prescriptions are refreshingly inexpensive: adopt Bland-Altman reporting, recruit disease populations instead of healthy volunteers, measure calibration drift, and exploit newly available open datasets that pair wearable recordings with invasive references. The barrier, they conclude, is not hardware or algorithms but an incentive structure that rewards demonstrating success in ideal conditions over establishing where a method fails.

Subject of Research: Clinical validation readiness of wearable cardiac vibration signals for estimating central hemodynamic parameters

Article Title: Clinical Readiness of Hemodynamic Estimation from Cardiac Vibration Signals: An Endpoint-Specific Critical Review

Article References: Gamage, P. T., Weerasinghe, Y., Weerasinghe, P., & Kaya, M. (2026). Clinical Readiness of Hemodynamic Estimation from Cardiac Vibration Signals: An Endpoint-Specific Critical Review. Annals of Biomedical Engineering. https://doi.org/10.1007/s10439-026-04366-5

Image Credits: AI Generated

DOI: 10.1007/s10439-026-04366-5

Keywords: seismocardiography, ballistocardiography, gyrocardiography, kinocardiography, hemodynamic monitoring, stroke volume, cardiac output, ejection fraction, cuffless blood pressure, wearable sensors, Bland-Altman analysis, heart failure