Acoustic Heart Sound Recording and Computer Analysis

Number: 0692

Table Of Contents

Policy
Applicable CPT / HCPCS / ICD-10 Codes
Background
References

---

Policy

Scope of Policy

This Clinical Policy Bulletin addresses the following: auscultation jacket, ballistocardiography, optical vibrocardiography, phonocardiography,  vectorcardiography, and acoustic heart sound recording and computer analysis.

1. Experimental, Investigational, or Unproven

   Aetna considers the following interventions experimental, investigational, or unproven because the effectiveness of these approaches has not been established:

   1. Acoustic heart sound recording, computer analysis and interpretation;
   2. Auscultation jacket, ballistocardiography, optical vibrocardiography, phonocardiography, and vectorcardiography;
   3. SENSORA AI Heart Disease Detection Platform for detection of cardiac dysfunction (e.g., atrial fibrillation, cardiac murmurs, reduced ejection fraction, and valvular heart disease).

---

Table:

CPT Codes / HCPCS Codes / ICD-10 Codes

Code	Code Description
CPT codes not covered for indications listed in the CPB:
0716T	Cardiac acoustic waveform recording with automated analysis and generation of coronary artery disease risk score
0962T	Assistive algorithmic analysis of acoustic and electrocardiogram recording for detection of cardiac dysfunction (eg, reduced ejection fraction, cardiac murmurs, atrial fibrillation), with review and interpretation by a physician or other qualified health care professional
HCPCS codes not covered for indications listed in the CPB:
S3902	Ballistocardiogram
ICD-10 codes not covered for indications listed in the CPB (not all-inclusive):
I00 - I99.9	Diseases of the circulatory system
R00.8 - R00.9	Other and unspecified abnormalities of heart beat
R01.0 - R01.2	Cardiac murmurs and other cardiac sounds
R09.89, R09.A0, R09.A1, R09.A2, R09.A9	Other specified symptoms and signs involving the circulatory and respiratory systems
R93.1	Abnormal findings on diagnostic imaging of heart and coronary circulation
R93.811 - R93.89	Abnormal findings on diagnostic imaging of other specified body structures
R94.30 - R94.39	Abnormal results of cardiovascular function studies
Z13.6	Encounter for screening for cardiovascular disorders

---

Background

Acoustic Heart Sound Recording, Computer Analysis and Interpretation

Acoustic heart sound recording with computerized analysis involves the digital acquisition of heart sounds using an electronic stethoscope or other acoustic sensor, followed by software-based analysis of the recorded signals. These systems use signal processing techniques and, in some cases, artificial intelligence algorithms to identify acoustic patterns associated with cardiac abnormalities, such as heart murmurs, arrhythmias, valvular heart disease, or left ventricular dysfunction. The technology is intended to augment traditional cardiac auscultation by providing objective analysis and documentation of heart sounds and may be used as a screening or diagnostic support tool in primary care and cardiology settings.

Auscultation Jacket

An auscultation jacket is a wearable garment that incorporates multiple acoustic sensors or electronic stethoscope components to simultaneously record heart, lung, or vascular sounds from various anatomical locations. The technology is designed to facilitate standardized collection of physiologic acoustic data and may be used in conjunction with computerized signal analysis systems. Potential applications include cardiac screening, remote monitoring, and research involving automated assessment of cardiopulmonary function. However, the technology primarily serves as a method of data acquisition rather than a stand-alone diagnostic test.

Visagie and colleagues (2009) presented the findings of a study using an auscultation jacket with embedded electronic stethoscopes, and a software classification system capable of differentiating between normal and certain auscultatory abnormalities. The aim of the study is to demonstrate the potential of such a system for semi-automated diagnosis for under-served locations, for instance in rural areas or in developing countries where patients far out-number the available medical personnel. Using an "auscultation jacket", synchronous data was recorded at multiple chest locations on 31 healthy volunteers and 21 patients with heart pathologies. Electrocardiograms were also recorded simultaneously with phonocardiographic data. Features related to heart pathologies were extracted from the signals and used as input to a feed-forward artificial neural network. The system is able to classify between normal and certain abnormal heart sounds with a sensitivity of 84% and a specificity of 86%. Although the number of training and testing samples presented were limited, the system performed well in differentiating between normal and abnormal heart sounds in the given database of available recordings. The results of this study showed the potential of such a system to be used as a fast and cost-effective screening tool for heart pathologies.

Ballistocardiography

Ballistocardiography (BCG) is a non-invasive technique that measures the mechanical recoil and body movements generated by cardiac contraction and blood flow through the cardiovascular system. The signals may be acquired using sensors embedded in beds, chairs, scales, wearable devices, or other monitoring platforms. BCG is intended to provide information regarding cardiac output, heart rate, ventricular function, and hemodynamic status without directly imaging the heart. Recent advances in sensor technology and digital signal processing have renewed interest in BCG for remote cardiac monitoring and heart failure surveillance.

BGC had been investigated for potential use in measuring cardiac output (CO) and other aspects of cardiac function. Taylor and Sheffer (1990) stated, however, that thermodilution is the preferred method of measuring CO, and has achieved universal appeal in the clinical setting. The ballistocardiogram and other means of detecting CO, such as the impedance-cardiogram, have been developed and are being used clinically, but the development of the flow-directed thermodilution catheter has profoundly affected the universal acceptance of the thermodilution method. Thermodilution techniques, when performed properly, are capable of obtaining accurate and reproducible results.

1. bias (predicted greater than thermodilution) = 0.03 ml (95% confidence interval [CI]: -4.2 to 4.8 ml);
2. lower limit of agreement = -21 ml (95% CI: -29 to -13 ml); and
3. upper limit of agreement = 22 ml (95% CI: 14 to 29 ml).

Of derived stroke volumes, 82% were within 15 ml of pulmonary artery thermodilution-derived values. The authors concluded that sternal acceleration ballistocardiogram combined with hemodynamic and demographic data in a probabilistic model shows promise of providing a less invasive measure of CO than thermodilution.

Morra and colleagues (2019) examined whether micro-accelerometers and gyroscopes may provide useful information for the detection of breathing disturbances in further studies. A total of 43 healthy volunteers performed a 10-second end-expiratory breath-hold, while ballistocardiography (BCG) and seismocardiography (SCG) determined changes in kinetic energy and its integral over time (iK, J · s). BCG measures overall body accelerations in response to blood mass ejection into the main vasculature at each cardiac cycle, while SCG records local chest wall vibrations generated beat-by-beat by myocardial activity. This minimally intrusive technology evaluates linear accelerations and angular velocities in 12 degrees of freedom to calculate iK during the whole cardiac cycle; iK produced during systole and diastole were also computed. The iK during normal breathing was 87.1 [63.3; 132.8] µJ · s for the SCG and 4.5 [3.3; 6.2] µJ · s for the BCG. Both increased to 107.1 [69.0; 162.0] µJ · s and 6.1 [4.4; 9.0] µJ · s, respectively, during breath-holding (p = 0.003 and p < 0.0001, respectively). The iK of the SCG further increased during spontaneous respiration following apnea (from 107.1 [69.0; 162.0] µJ · s to 160.0 [96.3; 207.3] µJ · s, p < 0.0001). The ratio between the iK of diastole and systole increased from 0.35 [0.24; 0.45] during apnea to 0.49 [0.31; 0.80] (p < 0.0001) during the restoration of respiration. The authors concluded that a brief voluntary apnea generated large and distinct increases in SCG and BCG waveforms. These researchers stated that iK monitoring during sleep may prove useful for the detection of respiratory disturbances.

Morra and associates (2020) examined whether modern BCG and SCG are useful in the detection of hemodynamic changes occurring during simulated obstructive apneic events. A total of 47 healthy volunteers performed a voluntary maximum Mueller maneuver (MM) for 10 seconds, and BCG and SCG signals were simultaneously recorded. The kinetic energy of a set of cardiac cycles before and during the apneic episode was automatically computed from the rotational and linear channels of the SCG and BCG waveforms, and iK was derived (unit of measure: microjoules per second (µJ·s)). The estimated transmural pressure (eP TM) was assessed as the difference between systemic blood pressure (BP) and maximal inspiratory pressure (MIP). The Wilcoxon signed-rank test was used to evaluate differences in energy measurements between normal respiration and the loaded inspiration maneuver. Cardiac kinetic energies and the MIP produced during the MM were compared by linear regression analysis following log transformation in order to examine the correlation between variables. The iK during normal breathing increased from 1.1 (0.8; 1.4) to 1.9 (1.4; 4.3) µJ·s during MM (p < 0.001). When eP TM was considered, this association became positive (r: +0.58, p < 0.001 and r: +0.60, p < 0.001). When the iKLIN was considered as the comparative factor, correlations with the MIP and eP TM were weak and insignificant. Men had higher values of iK than women. The authors concluded that simulated obstructive apnea elicited large rotational iK swings, which were related to the intensity of the inspiratory effort and, as such, to the intensity of the left ventricular afterload. These researchers stated that computation of cardiac kinetic energy through BCG and SCG may shed further light on the impact of obstructive respiratory events on the cardiovascular system.

Zaid et al. (2021) noted that the time interval between the peaks in the ECG and BCG waveforms, thoracic electrical bio-impedance (TEB), has been associated with the pre-ejection period (PEP), which is an important marker of ventricular contractility. However, the use of BCG-related markers in clinical practice is limited by the difficulty of obtaining a replicable and consistent signal from patients. These researchers examined the feasibility of BCG measurements within a complex clinical setting, by means of an accelerometer placed under the head pillow of patients admitted to the Surgical Intensive Care Unit (SICU). The proposed technique proved capable of capturing TEB based on the R peaks in the ECG and the BCG in its head-to-toe and dorsoventral directions. TEB detection was found to be consistent and repeatable both in healthy individuals and SICU patients over multiple data acquisition sessions. The authors concluded that this work provided a promising starting point to examine how TEB changes may relate to the patients' complex health conditions and give additional clinical insight into their care needs.

Sumali et al. (2022) stated that cardiovascular disease (CVD) is the number one cause of death globally. In the case of cardiopulmonary arrest due to myocardial infarction (MI), the survival rate is as low as 13.3% one month after resuscitation, which necessitates the need for continuous heart monitoring. These researchers developed a BCG measurement system using a load cell installed on a chair and a heart rate (HR) estimation algorithm that is robust to waveform changes, with the objective of constructing a non-contact HR acquisition system. The proposed system was examined by employing data obtained from 13 healthy subjects and one subject with abnormal ECG who were simultaneously measured with ECG. The output of the BCG system was confirmed to change with the same period as the ECG data obtained as the correct answer, and the synchronization of the R-peak positions was confirmed for all cases. As a result of comparing the HR intervals estimated from BCG and those obtained from ECG, it was confirmed that the same HR variability (HRV) features could be obtained even for the subject with abnormal ECG.

These researchers stated that there are several suggestions for future investigation. First, a larger sample size with varying age distribution is needed to validate the effectiveness of the results of this study for a more generalized sample size. Second, a larger sample size of patients with various CVDs is needed to confirm the effectiveness of the proposed system for non-healthy subjects. Third, the use of signal fusion technologies instead of signal selection might improve the overall performance of the system. Fourth, the use of machine-learning (ML) models specialized for peak detection might also improve performance; as in this study, convolutional neural networks (CNN) improve the system’s effectiveness compared to conventional studies using traditional peak detection algorithms.

Rajput et al. (2022) noted that managing hypertension (HPT) remains a significant challenge. Despite advancements in BP-measuring systems and the accessibility of safe and effective anti-hypertensive medicines, HPT is a major public health concern. Headaches, dizziness, and fainting are common symptoms of HPT. In HPT patients, normalcy may be observed at one instant, and abnormality may prevail during a long duration of 24-hour ambulatory BP monitoring. This may cause difficulty in identifying patients with HPT, and hence there is a possibility that individuals may be untreated or administered insufficiently. More importantly, uncontrolled HPT can result in severe complications (stroke, heart attack, kidney disease, and heart failure [HF]), mainly ignoring the signs in nascent stages. HPT in the beginning stages may not present distinct symptoms and may be difficult to diagnose from standard physiological signals. Hence, BCG signal was used in this study to detect HPT automatically. The processed signals from BCG were converted into scalogram images using a continuous wavelet transform (CWT) and were then fed into a 2-D CNN model, which was trained to learn and recognize BCG patterns of healthy controls (HC) and HPT classes. The authors concluded that the proposed model obtained a high classification accuracy of 86.14% with a 10-fold cross-validation (CV) strategy. Hence, this was the first use of a 2D-CNN model (deep-learning algorithm) to detect HPT using BCG signals. Moreover, these researchers stated that in the future, they will validate the developed model with more data belonging to different stages (mild, moderate, and severe) of HPT from different races.

The authors stated that this study had two main drawbacks. First, the BCG signal collection using a smart device requires the subject to be continually lying on the mattress. This method could not examine a subject’s physiological state while performing routine activities, and obtaining high-quality signals proves to be a problem. To overcome this issue, the inclusion of other smart monitoring devices is suggested, some of which may include smartwatches or smart chairs to obtain signals during daytime activities. Second, CNN takes a long duration to train a large dataset and generally requires a graphics processing unit (GPU) to accelerate the training process. As a result, this increases the cost and complexity of the model.

Steffensen et al. (2023) stated that BCG features are of interest in wearable cardiovascular monitoring of cardiac performance. These investigators examined the feasibility of wrist acceleration BCG during exercise for estimating pulse transit time (PTT), enabling broader cardiovascular response studies during acute exercise and improved monitoring in individuals at risk for CVD. In addition, they investigated the relationship between PTT, BP, and stroke volume (SV) during exercise and posture interventions. A total of 25 subjects underwent a bike exercise protocol with four incremental workloads (0 W, 50 W, 100 W, and 150 W) in supine and semi-recumbent postures. BCG, invasive radial artery BP, tonometry, photoplethysmography (PPG), and ECG were recorded. Ensemble averages of BCG signals determined aortic valve opening (AVO) timings, combined with peripheral pulse wave arrival times to calculate PTT. These researchers tested for significance using the Wilcoxon signed-rank test. BCG was successfully recorded at the wrist during exercise. PTT exhibited a moderate negative correlation with systolic BP (ρSup = -0.65, ρSR = -0.57, ρAll = -0.54). PTT differences between supine and semi-recumbent conditions were significant at 0 W and 50 W (p < 0.001), less at 100 W (p = 0.0135), and 150 W (p = 0.031). Systolic BP and diastolic BP were lower in semi-recumbent posture (p < 0.01), while HR was slightly higher. ECG confirmed the association of BCG features with AVO and indicated a positive relationship between BCG amplitude and SV (ρ = 0.74). The authors concluded that wrist BCG may allow convenient PTT and possibly SV tracking during exercise, enabling studies of cardiovascular response to acute exercise and convenient monitoring of cardiovascular performance.

The authors stated that a drawback of this study was the use of steady-state exercise. Steady exercise conditions allow multiple heartbeats to be averaged, mitigating signal quality issues to a degree. However, this approach removes some of the variance in the averaged data (e.g., breathing cycles) that may be of interest. Furthermore, the steady states examined in this trial may not be directly comparable to many uncontrolled exercise conditions where exertion levels might vary more dynamically. Application in more realistic exercise situations requires further investigation in instrumentation as well as in signal processing to handle motion artifacts and enable the use of shorter ensemble windows. The sample size of this trial was small, and extrapolation to larger populations may also be complicated by the homogenous makeup of the participant group. In addition, these investigators noted that the four exercise load levels, set to absolute levels of delivered work, allowed for intra-individual comparison across a linearly increasing load; however, individual responses at each level will vary significantly. In the absence of quantitative measures of perceived exertion, i.e., Borg scale ratings, inter-individual comparisons must be made with caution.

Feng et al. (2023) noted that under the influence of COVID-19 and the in-hospital cost, the in-home detection of CVD with smart sensing devices is becoming more popular recently. In the presence of qualified signals, BCG can not only reflect the cardiac mechanical movements but also detect HF in a non-contact manner. However, for potential HF patients, the additional quality ECG-aided assessment requires more procedures and brings inconvenience to their in-home HF diagnosis. To enable the detection of HF in many real applications, these researchers proposed an ML-aided scheme for HF detection, where the BCG signals recorded from the force sensor were used without the heartbeat location, and the respiratory effort signals separated from force sensors provided more HF features due to the connection between the heart and lung systems. Lastly, the effectiveness of the proposed HF detection scheme was verified in comparative experiments. First, a piezoelectric sensor was employed to record a signal sequence of the two-dimensional (2D) vital sign, which included the BCG and the respiratory effort. Then, the linear and non-linear features with respect to BCG and respiratory effort signals were extracted to serve the HF detection. Finally, the improved HF detection performance was verified via the leave-one-out (LOO) and the leave-one-subject-out (LOSO) cross-validation settings with different ML classifiers. The proposed ML-aided scheme achieved robust performance in HF detection by using four different classifiers and yielded an accuracy of 94.97% and 87.00% in the LOO and LOSO experiments, respectively. Furthermore, experimental results showed that the designed respiratory and cardiopulmonary features were beneficial to HF detection (left ventricular ejection fraction [LVEF] less than or equal to 49%). The authors concluded that compared with existing studies focusing on BCG signals, the proposed scheme fully exploited the relationship between the heart and lung systems. The experimental results verified that these features could significantly improve the accuracy performance and robustness of HF detection. These researchers stated that in the further step of this study, quantitative analysis for possible classification between HF patients with LVEF less than or equal to 40% and LVEF greater than 40% will be considered. They noted that it is thus expected that the proposed scheme has the potential for patients with limited mobility performing in-home HF detection.

Computer-Aided Electronic Auscultatory Devices

Computer-aided electronic auscultatory devices have been developed that acquire, record, and analyze the acoustic signals of the heart. Using acoustic signal processing algorithms, a computer analyzes the recording to identify specific heart sounds that may be present, including S1, S2, and suspected murmurs. A graphic user interface displays the results. These computer-aided electronic auscultatory devices are intended to provide support to the physician in the evaluation of heart sounds for the identification of suspected murmurs, a potential sign of heart disease.

Zargis Medical Corporation (Princeton, NJ) has received U.S. Food and Drug Administration (FDA) clearance to market the Zargis Acoustic Cardioscan, a computer-aided electronic auscultatory device intended to support physicians in analyzing heart sounds in patients. As this clearance was based on a 510(k) premarket notification, the manufacturer was not required to provide the evidence of efficacy that is necessary to support a premarket approval application (PMA). The FDA indications for use state that the device is not intended as a sole means of diagnosis and that interpretation of heart sounds with the Zargis Acoustic Cardioscan is only significant when used in conjunction with physician over-read, as well as consideration of all other patient data.

There is inadequate evidence of the validity of computer-aided electronic auscultatory devices or their impact on clinical outcomes in the peer-reviewed published medical literature. Clinical studies are necessary to determine the performance characteristics (sensitivity, specificity, and predictive values) of computer-aided electronic auscultatory devices and their impact on clinical management and patient outcomes.

Another brand of computer-aided electronic auscultatory device (correlated audio-electric cardiography device) is the Audicor System (Inovise Medical, Inc., Portland, OR). It received marketing clearance from the FDA through the 510(k) process in 2003. According to the product labeling, the Audicor Upgrade System, when used with Audicor Sensors in the V3 and V4 positions on the chest wall, is for use in attaining, analyzing, and reporting electrocardiographic and phonocardiographic data, and to provide interpretation of the data for consideration by physicians.

Phonocardiography (PCG) is the graphic or digital recording of heart sounds and murmurs generated during the cardiac cycle. Heart sounds are captured using a microphone, electronic stethoscope, or other acoustic transducer and displayed as waveforms that can be analyzed visually or by computer algorithms. PCG was historically used to aid the detection and characterization of valvular heart disease and other cardiac abnormalities before the widespread adoption of echocardiography. More recently, advancements in digital stethoscopes, signal processing, and artificial intelligence have renewed interest in phonocardiography as an adjunctive tool for cardiac screening and automated murmur detection.

Mansy et al. (2005) found that computerized analysis of vascular sounds may be useful in vessel patency surveillance. Moreover, they stated that further testing using longitudinal studies may be warranted.

Watrous and colleagues (2008) noted that as many as 50% to 70% of asymptomatic children referred for specialist evaluation or echocardiography because of a murmur have no heart disease. These researchers hypothesized that computer-assisted auscultation (CAA) can improve the sensitivity and specificity of referrals for the evaluation of heart murmurs. In this study, seven board-certified primary care physicians were evaluated both with and without the use of a computer-based decision-support system using 100 pre-recorded patient heart sounds (55 innocent murmurs, 30 pathological murmurs, and 15 without a murmur). The sensitivity and specificity of their murmur referral decisions relative to American College of Cardiology/American Heart Association guidelines, as well as the sensitivity and specificity of murmur detection and characterization (innocent versus pathological), were measured. Sensitivity for the detection of murmurs significantly increased with the use of CAA from 76.6% to 89.1% (p < 0.001), while specificity remained unaffected (80.0% versus 81.0%). Computer-assisted auscultation improved the sensitivity of correctly identifying pathological murmur cases from 82.4% to 90.0%, and the specificity of correctly identifying benign cases (with innocent or no murmurs) from 74.9% to 88.8% (p < 0.001). Referral sensitivity increased from 86.7% to 92.9%, while specificity increased from 63.5% to 78.6% using CAA (p < 0.001). The authors concluded that CAA appears to be a promising new technology for informing the referral decisions of primary care physicians.

Computer-aided electronic auscultatory devices differ from phonocardiograms in that only the former incorporate computer analysis of the heart sounds. The Center for Medicare and Medicare Services (CMS, 1997) has determined that phonocardiography and vectorcardiography diagnostic tests are "outmoded and of little clinical value." Vectorcardiography (VCG) is a method of recording and displaying the magnitude and direction of the heart’s electrical activity in three-dimensional space throughout the cardiac cycle. VCG data are typically derived from electrocardiographic signals and displayed as vector loops representing atrial and ventricular depolarization and repolarization. The technique has been investigated for the evaluation of cardiac conduction abnormalities, myocardial infarction, ventricular hypertrophy, arrhythmias, and cardiac resynchronization therapy. Vectorcardiography is generally considered an extension of conventional electrocardiography that provides additional spatial information regarding cardiac electrical activation.

Luciani and colleagues (2019) noted that acoustic cardiography (AC) is a hybrid technique that couples heart sound recording with ECG, providing insights into the electrical-mechanical activity of the heart in an unsupervised, non-invasive, and inexpensive manner. During myocardial ischemia, hemodynamic abnormalities appear in the first minutes, and these researchers hypothesized a putative diagnostic role of acoustic cardiography for prompt detection of cardiac dysfunction to improve future patient management. A total of 10 female Swiss large white pigs underwent permanent distal coronary occlusion as a model of acute myocardial ischemia. Acoustic cardiography analyses were performed prior to, during, and after coronary occlusion. Pressure-volume analysis was conducted in parallel as an invasive method of hemodynamic assessment for comparison. Similar systolic and diastolic intervals obtained with the two techniques were significantly correlated [Q to min dP/dt versus Q to second heart sound (r² = 0.9583, p < 0.0001), PV diastolic filling time versus AC perfusion time (r² = 0.9686, p < 0.0001)]. Indexes of systolic and diastolic impairment correlated with quantifiable features of heart sounds [Tau versus fourth heart sound Display Value (r² = 0.2721, p < 0.0001), cardiac output versus third heart sound Display Value (r² = 0.0791, p = 0.0023)]. Furthermore, acoustic cardiography diastolic time (AUC 0.675, p = 0.008), perfusion time (AUC 0.649, p = 0.024), and third heart sound Display Value (AUC 0.654, p = 0.019) emerged as possible indicators of coronary occlusion. Finally, these three parameters, when combined with heart rate into a composite joint index, represent the best model in these investigators’ experience for ischemia detection (AUC 0.770, p < 0.001). The authors concluded that in the rapidly evolving setting of acute myocardial ischemia, acoustic cardiography provided meaningful insights into mechanical dysfunction in a prompt and non-invasive manner. They stated that these findings should propel interest in resurrecting this non-invasive technique for future translational studies, as well as reconsidering its re-introduction in the clinical setting.

The authors stated that this study had several drawbacks. These researchers presented a pre-clinical study with a relatively small cohort (n = 10) composed of healthy pigs under general anesthesia. They noted that despite the aforementioned limitations, their experience shed light on putative immediate mechanical alterations that would be otherwise infeasible to replicate in a real-world clinical scenario, and it might represent a reference point for future studies involving acoustic cardiography with more complex and heterogeneous conditions.

Optical Vibrocardiography

Optical vibrocardiography is a non-invasive technology that uses optical sensors, cameras, or laser-based systems to detect subtle chest wall vibrations produced by cardiac mechanical activity. By measuring microscopic surface movements associated with myocardial contraction and blood flow, the technology seeks to derive physiologic information about cardiac function. Optical vibrocardiography is being investigated for contactless cardiac monitoring, detection of arrhythmias, and assessment of cardiovascular status in clinical and remote-monitoring settings. Morbiducci et al. (2007) described the use of VCG for heart rate (HR) monitoring, based on the measurement of chest wall movements induced by the pumping action of the heart, which is eligible as a surrogate of electrocardiogram (ECG) in assessing both cardiac rate and HR variability (HRV). The method is based on the optical recording of the movements of the chest wall by means of laser Doppler interferometry. To this aim, the ECG signal and the velocity of vibration of the chest wall, named VCG, were simultaneously recorded on 10 subjects. The time series built from the sequences of consecutive R waves (on ECG) and vibrocardiographic (VV) intervals were compared in terms of HR. To evaluate the ability of VCG signals as quantitative marker of the autonomic activity, HRV descriptors were also calculated on both ECG and VCG time series. Heart rate and HRV indices obtained from the proposed method agreed with the rate derived from ECG recordings (mean percent difference less than 3.1%). The authors concluded that optical VCG provides a reliable assessment of HR and HRV analysis, with no statistical differences in term of gender are present. They noted that optical VCG appears promising as non-contact method to monitor the cardiac activity under specific conditions (e.g., in magnetic resonance environment, or to reduce exposure risks to workers subjected to hazardous conditions). The technique may be used also to monitor subjects (e.g., severely burned, for which contact with the skin needs to be minimized).

Mahnke (2009) noted that heart disease is a major cause of worldwide morbidity and mortality. Properly performed, the cardiac auscultatory examination is an inexpensive, widely available tool in the detection and management of heart disease. Unfortunately, accurate interpretation of heart sounds by primary care providers is fraught with error, leading to missed diagnosis of disease and/or excessive costs associated with evaluation of normal variants. Thus, automated heart sound analysis, also known as CAA, has the potential to become a cost-effective screening and diagnostic tool in the primary care setting.

SENSORA AI Heart Disease Detection Platform

The SENSORA AI Heart Disease Detection Platform (Eko Health) is an FDA-cleared software platform that integrates digital stethoscope-derived phonocardiogram and electrocardiogram data with artificial intelligence algorithms to assist in the detection of cardiac abnormalities. The platform is designed to identify acoustic and electrical signatures associated with conditions such as atrial fibrillation, structural heart murmurs suggestive of valvular heart disease, and reduced left ventricular ejection fraction. SENSORA is intended for use by healthcare professionals as a clinical decision-support tool to supplement routine cardiovascular assessment and facilitate identification of patients who may warrant further diagnostic evaluation. 

Bachtiger et al. (2022) stated that most patients who have heart failure (HF) with a reduced ejection fraction (EF), when left ventricular ejection fraction (LVEF) is 40% or lower, are diagnosed in the hospital. This is despite previous presentations to primary care with symptoms. In a prospective, observational, multi-center study, these investigators tested an AI algorithm applied to a single-lead ECG, recorded during ECG-enabled stethoscope examination, to validate a potential point-of-care (POC) screening tool for LVEF of 40% or lower. This trial employed a convolutional neural network (CNN; known as AI-ECG) that was previously validated for the detection of reduced LVEF using 12-lead ECG as input. These investigators used AI-ECG retrained to interpret single-lead ECG input alone. Patients (aged 18 years or older) attending for trans-thoracic echocardiography (TTE) in London (U.K.) were recruited. All subjects had 15 seconds of supine, single-lead ECG recorded at the four standard anatomical positions for cardiac auscultation, plus one hand-held position, using an ECG-enabled stethoscope. TTE-derived percentage LVEF was used as ground truth. The primary outcome was the performance of AI-ECG at classifying reduced LVEF (LVEF of 40% or less), measured using metrics including the area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity, with two-sided 95% confidence intervals (CIs). The primary outcome was reported for each position individually and with an optimal combination of AI-ECG outputs (interval range of 0 to 1) from two positions using a rule-based approach and several classification models. Between February 6, 2021, and May 27, 2021, these researchers recruited 1,050 patients (mean age of 62 years [SD 17.4], 535 [51%] men, 432 [41%] non-White); 945 (90%) had an EF of at least 40%, and 105 (10%) had an EF of 40% or lower. Across all positions, ECGs were most frequently of adequate quality for AI-ECG interpretation at the pulmonary position (979 [93.3%] of 1,050). Quality was lowest for the aortic position (846 [80.6%]). AI-ECG performed best at the pulmonary valve position (p = 0.02), with an AUROC of 0.85 (95% CI: 0.81 to 0.89), sensitivity of 84.8% (76.2% to 91.3%), and specificity of 69.5% (66.4% to 72.6%). Diagnostic odds ratios (DORs) did not differ by age, sex, or non-White ethnicity. Taking the optimal combination of two positions (pulmonary and hand-held positions), the rule-based approach resulted in an AUROC of 0.85 (0.81 to 0.89), sensitivity of 82.7% (72.7% to 90.2%), and specificity of 79.9% (77.0% to 82.6%). Using AI-ECG outputs from these two positions, a weighted logistic regression with L2 regularization resulted in an AUROC of 0.91 (0.88 to 0.95), sensitivity of 91.9% (78.1% to 98.3%), and specificity of 80.2% (75.5% to 84.3%). The authors concluded that this study found that AI-ECG could identify patients with reduced LVEF (40% or less) from single-lead ECG inputs. Through the use of an ECG-enabled stethoscope, these investigators highlighted an AI algorithm embedded in a familiar clinical tool that fits into routine and universal clinical workflows. These researchers stated that given the frequent clinical encounters of undiagnosed patients before index hospital admission for HF, the stethoscope examination has the potential to be a POC screening opportunity, and through further AI algorithms, to become a tool for comprehensive detection of cardiovascular disease.

The authors stated that this trial had several limitations. First, the patient cohort was not fully representative of a screening population, where a lower prevalence of an LVEF of 40% or lower could influence performance characteristics, especially positive predictive values (PPVs). Although the disease profile of the study population probably reflected those who would benefit most from screening for reduced LVEF, further studies are needed, especially in primary care settings. Further investigations should also address the current paucity of data describing the prevalence of asymptomatic disease. Second, without comprehensive access to all participants' electronic health records to determine any previously normal LVEF, the authors were unable to precisely characterize how many subjects were flagged as positive by AI-ECG as part of an index diagnosis of HF with a reduced EF. Third, there was established inter-operator variability in the measurement of LVEF from ECG, giving rise to the possibility that some subjects close to the LVEF 40% borderline were misclassified. Further investigations with a higher number of operators across the wider clinical workforce are needed to examine if the device is universally easy to use. Fourth, these researchers have not tested for reproducibility. There would have been inter-operator variability in the precise position and angulation of the ECG-enabled stethoscope, although this variability reflected the reality of clinical practice.

Rancier et al. (2023) noted that clinically significant valvular heart disease (VHD) affects 10% of adults over the age of 65 years. Cardiac auscultation, the current POC clinical standard, has a 44% sensitivity for detecting VHD, leaving most patients undiagnosed. A deep learning (DL)-based AI platform that detects cardiac murmurs associated with clinically significant VHD (moderate severity or higher, confirmed by ECG) was used to examine its real-time impact on the POC. These researchers hypothesized that the use of AI during POC auscultation may result in higher discovery rates and earlier diagnosis of VHD compared to conventional practice. In a blinded, prospective study, these investigators examined the real-world performance of AI versus primary care physician (PCP) auscultation in identifying previously undiagnosed VHD. This trial enrolled 369 patients aged 50 years or older without a prior VHD diagnosis or history of murmur from three different primary care clinics. Four-point cardiac auscultation was carried out on each patient by a PCP with a standard stethoscope, and a coordinator with an Eko digital stethoscope collected phonocardiogram (PCG) data for analysis by the AI. Each patient received an echocardiogram to confirm whether clinically significant VHD was present, and their PCG recordings were reviewed by an outside expert panel to confirm whether audible murmurs were present. The PCP and expert panel were blinded to AI and echocardiogram results. Performance metrics were calculated for both PCP auscultation and the AI in detecting audible VHD. The AI showed over a two-fold improvement in sensitivity versus PCP auscultation for detecting audible VHD (94.1% versus 41.2%) with limited impact on specificity (84.5% versus 95.5%). The AI identified 22 patients with moderate-or-greater VHD who were previously undiagnosed, while PCPs identified 8 previously undiagnosed patients with VHD. The authors concluded that when applied in a primary care setting, a digital stethoscope with structural murmur detection AI showed meaningful impact on the new discovery of VHD as compared to conventional practice. These investigators stated that these findings suggested that AI-powered auscultation at the POC may significantly increase earlier VHD discovery, facilitate appropriate patient care, and improve outcomes.

Armoundas et al. (2024) stated that a major focus of academia, industry, and global governmental agencies is to develop and apply AI and other advanced analytical tools to transform healthcare delivery. The American Heart Association (AHA) supports the creation of tools and services that would further the science and practice of precision medicine by enabling more precise approaches to cardiovascular and stroke research, prevention, and care of individuals and populations. Nevertheless, several challenges exist, and few AI tools have been shown to improve cardiovascular and stroke care sufficiently to be widely adopted. This scientific statement outlined the current state of the art on the use of AI algorithms and data science in the diagnosis, classification, and treatment of cardiovascular disease. It also set out to advance this mission, focusing on how digital tools and, in particular, AI may provide clinical and mechanistic insights, address bias in clinical studies, and facilitate education and implementation science to improve cardiovascular and stroke outcomes. Finally, a principal objective of this scientific statement was to advance the field by identifying best practices, gaps, and challenges for interested stakeholders.

Guo et al. (2025) noted that asymptomatic left ventricular systolic dysfunction (ALVSD) affects 7 million worldwide, resulting in delayed diagnosis and treatment, high mortality, and substantial downstream healthcare costs. Current detection methods for ALVSD are inadequate, necessitating the development of improved diagnostic tools. Recently, ECG-based algorithms have demonstrated promise in detecting ALVSD. These researchers developed and validated a CNN model using single-lead ECG and PCG inputs captured by a digital stethoscope to examine its use in detecting individuals with actionably low EF in a large cohort of patients. A total of 2,960 adults undergoing ECG from four U.S. healthcare networks were enrolled in this observational, multi-center study. Patient data were captured by means of a digital stethoscope, and ECG was carried out within one week of data collection. The algorithm's performance was compared against echocardiographic EF (EF measurements, categorizing EF as normal and mildly reduced [greater than 40%] or moderate and severely reduced [40% or less]). The CNN model showed an AUROC of 0.85, with a sensitivity of 77.5%, specificity of 78.3%, positive predictive value (PPV) of 20.3%, and negative predictive value (NPV) of 98.0%. Among those with an abnormal AI screen but EF of greater than 40% (false positives), 25% had an EF between 41% and 49%; and 63% had conduction/rhythm abnormalities. Sub-group analyses revealed consistent performance across various demographics and comorbidities. The authors concluded that the CNN model, employing a digital stethoscope, offered a non-invasive and scalable method for early detection of individuals with EF of 40% or less. Moreover, these researchers stated that while these findings highlighted the promise of this technology for improving detection rates, further investigations are needed to examine its impact on clinical outcomes and cost-effectiveness. They stated that future work should focus on adapting the model for three-lead input and optimizing algorithm performance using transformer models. Furthermore, longitudinal studies examining patient outcomes are needed to assess the clinical utility and cost-effectiveness of the diagnostic tool. Such studies would provide valuable insights into the long-term effectiveness and impact on patient management strategies.

The authors stated that this study had several drawbacks. First, while the discriminatory power, sensitivity, and specificity of the algorithm were good, there is substantial room for improvement. These researchers believe that the addition of multiple ECG leads and/or the addition of clinical characteristics or other biomarkers such as brain natriuretic peptide (BNP) and N-terminal pro-B-type natriuretic peptide (NT-proBNP) could further optimize the identification of individuals with an undiagnosed low EF by potentially improving sensitivity and PPV. Second, shared by almost all machine learning (ML) models, is that the exact features of the ECG and PCG leading to the prediction of a low versus normal EF are not readily identifiable. Third, the accuracy of the ground truth EF data from ECG and potential biases therein could influence the performance and reliability of the algorithm, although the influence of any variability in EF determination would likely only impact those with EFs close to (e.g., ± 5%) 40%. Fourth, as the study population was comprised of patients already referred for an ECG, the generalizability of these findings to the broader population may be limited.

---

References