key: cord-0267130-mombkcxp
authors: Yazdi, D.; Patel, S.; Sridaran, S.; Wilson, E.; Smith, S.; Centen, C.; Gillon, L.; Kapur, S.; Tracy, J. A.; Lewine, K. A.; Systrom, D. M.; MacRae, C. A.
title: Non-Invasive Scale Measurement of Cardiac Output Compared with the Gold-Standard Direct Fick Method: A Feasibility Study
date: 2021-03-31
journal: nan
DOI: 10.1101/2021.03.28.21254243
sha: b82addd4e089efb627fe8cd377c27db27a34e7d8
doc_id: 267130
cord_uid: mombkcxp

Background: Objective markers of cardiac function are limited in the outpatient setting and may be beneficial for monitoring patients with chronic cardiac conditions. Objective: We assess the accuracy of a scale, with the ability to capture ballistocardiography, electrocardiography, and impedance plethysmography signals from a patient's feet while standing on the scale, in measuring stroke volume and cardiac output compared to the gold-standard direct Fick method. Methods: Thirty-two patients with unexplained dyspnea undergoing level 3 invasive cardiopulmonary exercise test at a tertiary medical center were included in the final analysis. We obtained scale and direct Fick measurements of stroke volume and cardiac output before and immediately after invasive cardiopulmonary exercise test. Results: Stroke volume and cardiac output from a cardiac scale and the direct Fick method correlated with r = 0.81 and r = 0.85, respectively (P < 0.001 each). The mean absolute error of the scale estimated stroke volume was -1.58 mL, with a 95% limits of agreement (LOA) of -21.97 mL to 18.81 mL. The mean error for the scale estimated cardiac output was -0.31 L/min, with a 95% LOA of -2.62 L/min to 2.00 L/min. The change in stroke volume and cardiac output before and after exercise were 78.9% and 96.7% concordant, respectively between the two measuring methods. Conclusions: This novel scale with cardiac monitoring abilities may allow for non-invasive, longitudinal measures of cardiac function. Using the widely accepted form factor of a bathroom scale, this method of monitoring can be easily integrated into a patient's lifestyle.

Cardiovascular disease, in particular heart failure (HF), is a major health and economic problem worldwide, expected to increase in incidence and prevalence due to the aging population and rise in co-morbidities 1,2 . Novel approaches for easily monitoring cardiac function trends over time in the home environment may prove to be important in dealing with these conditions. Accelerated by the COVID-19 pandemic, the field of medicine is increasingly shifting towards telemedicine and remote patient monitoring, welcoming innovation 3, 4 . In this study, we investigate the accuracy of a connected cardiac scale with ballistocardiography (BCG), impedance plethysmography (IPG), and electrocardiography (ECG) sensors in measuring stroke volume and cardiac output compared to the direct Fick method.

BCG measures the effects of the cyclical hemodynamic forces transmitted from the heart with each cardiac systolic ejection 5 . The methodology for BCG was developed and popularized in the 1950s, but its use waned later in the century due to the impractical nature of the apparatus, limited reliability of the measurements in diseased states, and a focus on other measures of cardiovascular function, such as blood pressure recordings 6, 7 . However, over the past decade BCG has regained popularity as a result of the ability to obtain measurements from novel sensor platforms, such as bathroom scales, advances in data processing and machine learning algorithms, and the emergence of more rigorous studies demonstrating the utility of BCG recordings from patients with cardiac diseases [8] [9] [10] . IPG measures pulsatile blood flow via changes in electrical impedance. This measurement, along with the ECG, can identify important cardiac time intervals such as valvular opening and closing, measures of contractility such as left ventricular ejection time, and estimates for stroke volume and cardiac output 11 . The simultaneous extraction of ECG, IPG, and BCG signals from a single scale measurement can enhance the amount of cardiovascular information obtained for any individual signal in isolation.

In this study, Bodyport Inc., a company based in San Francisco, CA, and the Brigham and Women's Hospital collaborated to assess the accuracy of the Bodyport Cardiac Scale in non-invasively measuring stroke volume and cardiac output. The cardiac scale has the form of a bathroom scale and has a multi-sensor system that can capture a single lead ECG, IPG, and BCG signals when a patient stands on the scale (Figure 1 ). This is the first study to our knowledge to investigate the accuracy of such a scale in measuring stroke volume and cardiac output compared to the gold-standard direct Fick method from pulmonary arterial catheters. Our findings demonstrate a robust correlation and relatively accurate mean error between direct Fick measurements of stroke volume and cardiac output and those obtained from the cardiac scale.

These data support the potential clinical utility of such home-based sensors.

Fifty-six subjects undergoing a level 3 invasive cardiopulmonary exercise test (iCPET) at the Shapiro Cardiovascular Center at Brigham and Women's Hospital were recruited and consented between July 2018 and January 2019. The patients were each undergoing the iCPET to evaluate unexplained dyspnea. These patients had an array of underlying comorbidities including heart failure, pulmonary artery hypertension, peripheral vasomotor abnormalities, valvular pathologies, and pulmonary diseases. The study was not designed to discriminate the accuracy of the cardiac output measurements dependent on the patient's underlying medical condition. Of the 56 subjects enrolled in the study, 32 subjects (9 male, 23 female) were . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted March 31, 2021.

included in the final analysis. Twenty-four subjects were excluded from the final analysis.

Among the 24 patients excluded, eleven subjects experienced light-headedness at the completion of the iCPET making it unsafe for them to stand on the scale. Six subjects had missing or incorrect reference data due to instrumentation related issues (i.e. patient pulled face mask off before Fick measurement completed). Seven patients had poor balance during the post-exercise recovery measurement requiring assistance from the clinical staff, which caused excessive motion artifacts requiring exclusion of the data. The age range for participants was 26-78 years (mean 51.7 years, SD 14.5 years).

We obtained scale measurements before and immediately after the iCPET. Ultrasound guided pulmonary artery catheters were placed prior to patient arrival in the iCPET lab. Once in the iCPET lab, patients were asked to stand, and a baseline pulmonary artery measurement was obtained followed immediately by a 2-minute baseline measurement on the cardiac scale.

Patients then mounted an upright cycle ergometer to perform the exercise portion of the test.

The exercise workload was gradually increased in a ramped protocol until the patient reached exhaustion or developed objective evidence of hemodynamic instability or myocardial ischemia.

After the exercise limit was reached, patients dismounted from the bicycle as quickly as possible to obtain a recovery pulmonary artery catheter measurement. This was immediately followed by a 2-minute recovery cardiac scale measurement.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted March 31, 2021. ; https://doi.org/10.1101/2021.03.28.21254243 doi: medRxiv preprint

The measurements obtained from the metabolic cart, radial and pulmonary artery catheters before and immediately after the iCPET study include ventilation, pulmonary gas exchange, venous and arterial blood gases, a 12-lead ECG, heart rate, pulmonary artery pressure, blood pressure from a radial artery catheter. Cardiac output was calculated using the direct Fick method as the current gold standard. Stroke volume was simply derived from its relationship with the measured cardiac output and heart rate.

The Bodyport Cardiac Scale is a physical platform on which the patient stands with bare feet.

Using embedded sensors, the scale measures three cardiovascular biosignals that are used to extract various cardiac biomarkers. The BCG, ECG, and IPG signals are analyzed by Bodyport's software and proprietary algorithms to extract characteristic features in each waveform that are used to estimate parameters including heart rate, heart rate variability, cardiac time intervals, and signal morphological features used to derive estimates of stroke volume and cardiac output.

We developed a feature set derived from the BCG, ECG, and IPG signals from the combined pre-exercise and post-exercise measurements to function as inputs in the model development CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted March 31, 2021. ; https://doi.org/10.1101/2021.03.28.21254243 doi: medRxiv preprint including those of Kubicek and Starr 12 . The BCG J-wave amplitude, which correlates with pulse pressure, along with additional proprietary electromechanical parameters from the BCG, ECG, and IPG signals, specifically pre-ejection period and left ventricular ejection time, were used to develop the final feature set 5 . Heart rate was used as a correction factor applied to time interval features. Additional non-cardiovascular parameters collected directly from the scale, such as body weight and basal impedance, were also incorporated into the model to remove the need for calibration and compensate for individual anthropomorphic variability.

The scale signals (BCG, ECG, IPG) were filtered and interpolated prior to feature extraction and model validation. For this calibration exercise, signal regions containing motion artifacts or excessive noise were objectively identified and removed. Motion artifacts and other sources of signal interference were detected through adaptive thresholding and monitoring of the patient's center of pressure during the measurement. Linear phase digital lowpass and highpass filters were applied to the signals to prevent distortion. Cutoff frequencies were established independently for each signal and ranged from 0.5Hz to 50Hz. All three signals were simultaneously sampled at 250Hz. Ensemble averaged waveforms were constructed from the real-time signals. Specific features were identified on the averaged waveforms using proprietary algorithms optimized for the Bodyport device. These features included signal amplitudes, such as the BCG J-wave magnitude, and temporal relationships between each of the averaged waveforms, such as pre-ejection period and other systolic time intervals.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted March 31, 2021. ; https://doi.org/10.1101/2021.03.28.21254243 doi: medRxiv preprint Regression model training for stroke volume was accomplished using a Tree-based Pipeline Optimization Tool (TPOT) 13 . This tool evaluates the performance of individual regression models, while optimizing hyperparameters. The following models were preselected to be used by TPOT: Linear Regression, Lasso, Elastic Net, Ridge, Random Forest, Support Vector Regression, Multi-Layer Perceptron. Feature preprocessors were also predefined to be included as part of the TPOT process. TPOT iterates over combinations of models and preprocessors, as well as the hyperparameter space. Each trained model in this stage used K-Fold cross-validation using three folds. Mean absolute error was used to optimize model training accuracy. The final model used an ensemble regression pipeline consisting of a random forest and gradient boosting regression and was then evaluated using leave-one-out cross-validation.

This cross-validation technique fits the model on all but one measurement, which is then used as the test measurement. Each measurement was held out once and the final accuracy was determined based on the performance of all held out test measurements. Bodyport-derived heart rate and stroke volume were used to derive an estimate of cardiac output. The p-value for the correlation coefficients was calculated using the Wald test with t-distribution of the test statistic. The Bland-Altman limits of agreement analysis for the combined pre-exercise and post-exercise data set accounted for the multiple measurements from the same subject using the methodology described by Bland and Altman 14 .

The multivariate regression model demonstrated a strong relationship between the scale and The change in stroke volume and cardiac output were evaluated pre-and post-exercise test for both the scale-derived and Fick methods, to assess if the scale can trend directional changes in these perfusion markers. We measured the degree of agreement between these two methods by calculating their concordance: the fraction of patients for which the change in stroke volume or cardiac output (post-exercise minus pre-exercise) were either both positive or negative using both modalities. The concordance for stroke volume was 78.9% and 96.7% for cardiac output (P < 0.001, each, derived from the concordance correlation) (Figure 4 ). To reduce statistical noise from the analysis, we excluded data points (13 for stroke volume and 2 for cardiac output)

where the change in stroke volume or cardiac output was less than 15% of the mean value in the study 15 .

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The stroke volume and cardiac output concordance before and immediately after exercise were 78.9% and 96.7%, respectively. Current advice for trending ability in cardiac output studies is a concordance greater than 92%, which is achieved in this cardiac output analysis but not the stroke volume 18 . The increased concordance for the cardiac output is likely a result of an increased heart rate post-exercise, compared to the more variable response seen in the stroke volume. Because exercise is incorporated in the iCPET, cardiac output is expected to increase. This is why a majority of the data points fall in the upper right corner of the cardiac output concordance plot (Figure 4) . Another experimental design is necessary to better detect reductions in cardiac output.

Having all sensors integrated in one device is imperative for the future adoption of such a technology, since applying multiple sensor technologies simultaneously is cumbersome and would likely yield low patient adherence. The analysis of these orthogonal sensor signals is further enhanced by ongoing advancements in signal processing and machine learning techniques. Utilizing the form factor of a bathroom scale will also enhance patient adherence, since taking scale measurements is a behavior already adapted by many patients, especially those with cardiac conditions such as heart failure.

A non-invasive, scalable, and inexpensive method for assessing cardiac function could have widespread applications in medicine. Robust estimates of stroke volume and cardiac output may help monitor the cardiac performance of patients with chronic conditions, such as heart failure, and facilitate early detection of decompensation and the virtual titration of goal-directed medical therapy. Given the simplicity of use and existing user behaviors for self-weighing on a scale, longitudinal data can be obtained and trended in large populations to identify novel biomarkers of cardiovascular health.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted March 31, 2021. ; https://doi.org/10.1101/2021.03.28.21254243 doi: medRxiv preprint cardiac output were relatively accurate at -1.58 mL (-21.97, 18.81 mL) and -0.31 L/min (-2.62, 2.00 L/min), respectively. The scale and direct Fick estimates for cardiac output were strongly concordant pre-and post-exercise at 96.7%, demonstrating the ability for scale to trend increases in cardiac output. Future studies will gather additional data to improve the model and will also assess longitudinal scale measurements from individuals with in clinical settings to better understand how these biomarkers, when integrated, can be used to assess relevant changes in a range of disorders.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted March 31, 2021. ; https://doi.org/10.1101/2021.03.28.21254243 doi: medRxiv preprint A central exclusion zone (square) represents the data within 15% of the mean stroke volume or cardiac output in the study, as they contain a high level of random variation compared to changes in the cardiac output. The line of identity y = x is shown.

The global health and economic burden of hospitalizations for heart failure: lessons learned from hospitalized heart failure registries

Temporal trends and patterns in heart failure incidence: a population-based study of 4 million individuals

Implications for Telehealth in a Postpandemic Future: Regulatory and Privacy Issues

Feasibility of Remote Video Assessment of Jugular Venous Pressure and Implications for Telehealth

Ballistocardiogram: Mechanism and Potential for Unobtrusive Cardiovascular Health Monitoring

A critical appraisal of ballistocardiography

Essay on the ballistocardiogram

Ballistocardiography and seismocardiography: a review of recent advances

Preliminary results from BCG and ECG measurements in the heart failure clinic

Non-invasive cardiac output trending during exercise recovery on a bathroom-scale-based ballistocardiograph

Beyond the four quadrants: the critical and emerging role of impedance cardiography in heart failure

The Minnesota impedance cardiograph-theory and applications

Scaling tree-based automated machine learning to biomedical big data with a feature set selector

Agreement between methods of measurement with multiple observations per individual

Transtracheal Doppler cardiac output monitoring. comparison to thermodilution during noncardiac surgery

International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity

We would like to thank the staff of the Brigham and Women's Hospital cardiopulmonary exercise lab for helping collect the data.

Daniel Yazdi and Sarin Patel are employees at Bodyport Inc. Sarah Smith and Corey Centen are founders of Bodyport Inc. Calum A. MacRae is an advisor for Bodyport Inc.