Introduction

Seismocardiographic (SCG) signals are chest wall vibrations induced by cardiac contractions, cardiac valve motion, and blood momentum changes1,2,3,4. SCG is often acquired using accelerometers attached to the chest surface5 or using non-contact methods6. SCG captures both audible and subaudible frequency heart sounds5,7. Castiglioni et al. showed that bandpass filtered SCG from 0.6 to 20 Hz has similarities to ballistocardiogram (BCG) whereas the high frequency components beyond 20 Hz were thought to be related to valves closure4. BCG is the body recoil due to cardiac blood ejection and movement into great vessels8. Previous studies suggested SCG clinical utility for evaluating cardiac contractility, hemodynamic parameters, sleep apnea and aortic valve stenosis9,10,11,12. Moreover, SCG is non-invasive and can be measured using portable devices13,14, which would make it compatible with ambulatory monitoring.

Sources of SCG variability include: respiration, sensor location, exercise, and posture15,16,17,18,19,20,21. SCG variability may impede the extraction of accurate clinically useful features5,22,23. Therefore, there have been studies focusing on understanding sources of variability and developing techniques to reduce it15,17,22,24,25,26,27,28.

Cardio-pulmonary interactions might lead to changes in myocardial contractions and blood ejection. For example, since breathing modulates lung volume and intrathoracic pressure, breathing can in turn modulate cardiac preload and afterload29. Preload is defined as the initial stretching of the cardiac myocytes before contraction (the end-diastole ventricular volume or pressure are often used as indirect measures of preload)30. During inspiration, the diaphragmatic descent reduces intrathoracic pressure and increases intra-abdominal pressure29. As a result, venous return and right ventricular (RV) preload increase29. Due to this increase in RV preload, the RV pumps more blood to the lungs (i.e., due to higher stroke volume). The inflated lungs reserve more blood, and accordingly the left ventricle (LV) filling may be decreased30. Moreover, the interventricular septum may move to the left during inspiration, in response to the increase in RV preload, which limits LV filling and thereby reduces preload29.

Additionally, the LV afterload increases during inspiration because of an accompanying increase in the aortic transmural pressure (i.e., the difference between aortic intravascular systolic pressure and the intrapleural pressure)29. Here, afterload refers to the load against which the heart must contract to eject blood30. The RV afterload may also increase if lung volume changes (increase or decrease) beyond its functional residual capacity (e.g., during deep breathing) due to the associated increase in pulmonary vascular resistance (PVR)29.

Cardiopulmonary interactions could affect SCG through modulating the intensity of the first (S1) and second (S2) heart sounds, cardiac time intervals, and spectral energy distribution31,32. S1 intensity was reported to decrease during inspiration due to the low LV preload while S2 intensity increased due to the high LV afterload31. In addition, heart sound intensities are likely dependent on the relative location of chest-mounted accelerometer relative to the heart location, which varies during breathing31. Pandia reported a decrease in the S1–S2 time interval during inspiration and suggested that this might be attributed to the relatively low LV preload which would shorten LV systole31. It was also observed that the S1–S1 interval shortened with inspiration, which was thought to be associated with respiratory sinus arrhythmia (RSA), where the heart rate increases during inspiration31. Periodic modulation of SCG spectrum during respiratory cycle was also reported32.

To reduce SCG variability due to respiration, two techniques were reported in the literature. The first approach was grouping SCG beats, acquired during normal breathing, into clusters of similar morphologies using, for example, unsupervised machine learning (ML), while the second approach was to acquire SCG during breath hold (BH)22,25. Compared to SCG acquired during normal breathing without clustering, Azad et al.25 reported that clustering reduced intra-cluster variability by about 19% while SCG acquired during BH reduced intra-cluster variability by about 42%. One of the limitations of that study was the small number of test subjects which may render some of its findings inconclusive25. On the other hand, BH could induce SCG morphological changes compared to normal breathing. Pandia et al.31 reported that the S1/S2 intensity ratio and S1–S1 time interval were significantly higher during apnea than normal breathing (inspiration and expiration). However, the methodology in that study was not clear regarding the timing of BH relative to the respiration cycle (e.g., end inspiration, end expiration, etc.)31. Different BH timings will correspond to different intrathoracic pressures, lung volumes and heart locations, which might affect cardiac dynamics and SCG waveform.

To the best of the authors’ knowledge, the effect of the intrathoracic pressure on SCG variability has not been sufficiently investigated. The current study aims to investigate SCG variability during normal breathing (NB) and breath holding (BH). BH happened at end inspiration (end INS) and end expiration (end EXP) at different airway pressures, which would also correspond to different intrathoracic pressures. The SCG spectrum was also investigated to further understand how changing the intrathoracic pressure during BH can affect spectral energy distribution and the corresponding changes in variability.

Besides increasing our understanding of SCG variability sources, results of this study may provide guidance regarding breathing maneuvers that may increase the clinical utility of SCG for diagnosing heart diseases.

Methods

Data acquisition

SCG, electrocardiography (ECG) and spirometry were acquired simultaneously at a sampling rate of 1 kHz. ECG was recorded using a biopotential recorder (IX-B3G, iWorx Systems, Inc., Dover, NH). SCG was measured using a tri-axial accelerometer (Model: 356A32, PCB Piezotronics, Depew, NY) attached to the 4th intercostal space (ICS) close to the left lower sternal border using double-sided medical grade tape. Later, the analysis focused on the dorsoventral component of SCG for feature extraction15. The respiratory flow rate was measured using a spirometer (Model: A-FH-300, iWorx Systems, Inc., Dover, NH).

Experimental protocol

After approval of IRB University of Central Florida, 15 healthy subjects (11 Females and 4 males, 21 ± 2 y) participated after providing their consent. Subjects rested on a horizontal exam table with their back raised at 45°. Data were collected during NB and BH with the glottis open. NB sessions were 5 min long. BH was performed at end INS and end EXP at airway pressure of 0, ± 2–4 and ± 15–20 cm water. Three trials were performed, for each BH condition, with 2–3 min of rest in between. Each BH trial lasted around 20 s. For pressures other than zero, a face mask covering the mouth and nose was attached to a manometer to monitor the airway pressure. Airway pressure was displayed on the computer screen to help the subject maintain the target pressures.

Preprocessing

SCG was filtered using a Chebyshev type 2 bandpass filter23 (0.05–200 Hz) to remove background noise and keep as much information of the signal as possible. ECG R-waves were identified using Pan-Tompkins algorithm33. SCG was segmented such that each beat started 100 ms before the corresponding R-peak and ended 100 ms before the following R-peak15,22,34. The segmented beats were down sampled to 500 Hz to reduce the analysis computational cost22. Each SCG beat was normalized by subtracting the mean and dividing by the peak-to-peak amplitude. This normalization helps making waveform comparisons be more dependent on SCG morphology rather than amplitude.

Clustering normal breathing SCG

Normal breathing SCG beats were clustered into two groups with similar waveform morphologies using the k-medoid algorithm24,25. The choice of two clusters was based on35, where different methods (namely, the elbow method and average Silhouette value) were used to test different numbers of clusters. Two clusters division was found to be an optimum solution. Implementation of the k-medoid algorithm involved using the dynamic time warping (DTW) algorithm to measure the dissimilarity (also called DTW distance) among SCG beats24,25. K-medoid algorithm was used over the K-means algorithm because finding the mean in the K-means method requires SCG beats of the same length36. Due to heart rate variability, segmented SCG beats are not the same length. As such, K-means application will require trimming longer SCG beats to equate the lengths of all beats. This can cause a loss of information from the trimmed SCG beats.

Intra- and Inter-state variabilities

In the current study, a state refers a NB condition or a certain lung volume and airway pressure at BH where a group of SCG beats are acquired. Each group of SCG events (at a certain state) can be represented by their medoid. For example, the medoid (\(C_{j}\)) of a state (\(j\)) is an SCG beat from the group that represents that state, where \(C_{j}\) is defined as the SCG beat the has minimum sum of distances from the other SCG beats (\(X_{{i_{j} }}\)) in the group. Therefore, \(C_{j}\) can be calculated according to Eq. 124. In Eq. 1, \(X_{{i_{j} }}\) refers to the ith SCG beat, and \(n_{j}\) is the number of beats for the state (\(j\)). The function “\(dtw\)” refers to the DTW distance.

$$C_{j} = argmin_{{y\epsilon \left\{ {X_{{1_{j} }} ,X_{{2_{j} }} , \ldots ,X_{{i_{j} }} , \ldots ,X_{{n_{j} }} } \right\}}} \left( {\mathop \sum \limits_{i = 1}^{{n_{j} }} dtw\left( {y,X_{{i_{j} }} } \right)} \right)$$
(1)

Normal breathing before clustering and BH at a certain airway pressure and lung volume are individual states that are each represented by one group of SCG beats. On the other hand, normal breathing after clustering is a state that comprises two groups (i.e., cluster 1 and 2; each has its own medoid.).

The intra-state variability was calculated using Eq. 2a if the state has one group and Eq. 2b if the state has two groups. The intra-state variability focused on measuring the variability within the same state. The inter-state variability was calculated using Eq. 3 and focused on measuring the variability between different states.

$$Intra - state Variability = \frac{1}{n}\left( { \mathop \sum \limits_{i = 1}^{n} dtw\left( {C,x_{i} } \right)} \right)$$
(2a)
$$Intra - state\;Variability = \frac{1}{{n_{1} + n_{2} }}\left( {\mathop \sum \limits_{i = 1}^{{n_{1} }} dtw\left( {C_{1} ,x_{i1} } \right) + \mathop \sum \limits_{j = 1}^{{n_{2} }} dtw\left( {C_{2} ,x_{j2} } \right)} \right)$$
(2b)
$$Inter - state\;Variability = \frac{1}{{n_{1} + n_{2} }}\left( {\mathop \sum \limits_{i = 1}^{{n_{1} }} dtw\left( {C_{2} ,x_{i1} } \right) + \mathop \sum \limits_{j = 1}^{{n_{2} }} dtw\left( {C_{1} ,x_{j2} } \right)} \right)$$
(3)

Frequency domain analysis

The power spectral density (PSD) of SCG was calculated and the ratio of subaudible (i.e., below 20 Hz) and audible energy (i.e., above 20 Hz) of the spectrum (SAR) was determined, using Eq. 4. In Eq. 4, “\(f\)” refers to the frequency in Hz. SAR is significant as it represents the ratio of the contribution of subaudible and audible components of SCG signal which are in turn related to certain physiological effects (i.e., blood ejection and movement and valve closure respectively)4.

$$Subaudbile\;to\;audible\;energy\;ratio\left( {SAR} \right) = \left( {\mathop \smallint \limits_{0}^{20} PSD \cdot df/\mathop \smallint \limits_{20}^{200} PSD \cdot df} \right)*100$$
(4)

Human ethics and consent to participate declarations

  • Experiments were done after approval of IRB of University of Central Florida (in accordance with guidelines and regulations of the United States Department of Health and Human Services).

  • Study subjects provided their consent to participate in the study.

Results and discussion

Figure 1 shows an example of the distribution of SCG beats of clusters 1 and 2 in the flow rate-lung volume (FL-LV) space. Each SCG beat in Fig. 1 appears as one point (filled circle) represented by two coordinates: the standardized flow rate and lung volume at the timing of the corresponding ECG-R wave. Standardization was done by removing the mean and dividing by the standard deviation. The coordinates of the square marks in Fig. 1 designate the median standardized flow rate and lung volume of each cluster. Cluster 1 was found to have a consistently higher median lung volume than cluster 2 in all subjects.

Figure 1
figure 1

Example of the distribution of SCG beats of clusters 1 and 2 in the standardized flow rate-lung volume space (FL-LV) plane. Each SCG beat appears as one point at the timing of the corresponding ECG-R wave. The large squares designate the median standardized flow rate and lung volume of each cluster.

Full size image

Figure 2 shows the intra-state SCG variability of all subjects at all states. Compared with the unclustered NB state, the intra-state variability decreased for the clustered NB by 27% and for the BH by 62% (at zero airway pressure, whether at end INS or end EXP). This reduction is comparable to reported values of 19% and 42%, respectively25. During BH, the heart location and lung volume do not change compared to NB which might explain the observed stronger reduction in the intra-state variability for BH at zero airway pressure. In addition, Fig. 2 shows that as the airway pressure deviated from zero (increased or decreased), the intra-state variability for BH (for end INS and end EXP states) increased (p < 0.05, t-test). This suggests that the airway pressure (which correlates with the intrathoracic pressure) is a likely source of SCG variability. The mechanism of the increased intra-state variability of BH states as the airway pressure deviated from zero is unclear. However potential physiological mechanisms that may play a role include: (1) deviation of the non-zero pressure from typical physiological conditions that might have triggered a certain physiological response (2) induced changes in the cardiac preload and afterload as mentioned above in the Introduction section. The increase in the intra-state variability may also reflect the heart effort to adapt and compensate for pressure changes.

Figure 2
figure 2

The intra-state variability of NB and BH states at different airway pressures. The values shown represent the mean + standard deviation over all subjects.

Full size image

Figure 3 shows the ratio of subaudible to audible SCG spectral energy (SAR) for all subjects. SAR values in Fig. 3 are for the two clusters extracted from NB data and for BH states at different airway pressures. The data of Fig. 3 shows that SAR for these two clusters were comparable. In addition, SAR for cluster 1 was significantly higher (p < 0.05, t-test) than BH at end INS at zero airway pressure. BH SAR (for both end INS and end EXP) at—(15–20) cm H2O was significantly higher than BH at zero airway pressure (p < 0.05, t-test). But BH SAR at + (15–20) cm H2O was not significantly higher. It can then be concluded that SAR tended to increase for negative intrathoracic pressure. As discussed in the Introduction, the subaudible SCG components may be related to blood ejection into great vessels while the audible components are more related to valve closures. In addition, low intrathoracic pressures were reported to increase LV afterload which may require increased myocardium contraction force to overcome the higher afterload. Moreover, cardiac muscle fibers contraction velocity is known to decrease for higher afterload30. Therefore, the increase in low frequency energy (and hence SAR increase in Fig. 3) with intrathoracic pressure reduction might be a manifestation of the potential increase in the myocardium contraction force and decrease in the contraction velocity. This suggests potential utility of SAR for monitoring and diagnosis of cardiac conditions that involve myocardial contraction changes.

Figure 3
figure 3

Subaudible to audible energy ratio of SCG for NB clusters and BH states at different airway pressures. The values shown represent the mean + standard deviation over all subjects.

Full size image

Figure 4 shows the mean intra-state variability of BH end INS and end EXP states (with the same airway pressure) compared to the intra-state variability when combining these states into one state. Combining BH end INS and end EXP states was done by considering them as a one group, then the intra-state variability was calculated using Eq. 2a. As shown in Fig. 4, combining BH end INS and end EXP states as one group resulted in a higher intra-state variability (p < 0.05, t-test). The main anatomic difference between end INS and end EXP is the changes in lung volume and heart location. Therefore, the increase in intra-state variability when combining end INS and EXP suggests that the lung volume and heart location are significant sources of SCG variability. This increase in the intra-state variability seemed to be independent of the airway pressure (Fig. 4), which suggests a relatively small “factor interaction” between the airway pressure on one hand and lung volume and heart location on the other hand.

Figure 4
figure 4

The intra-state variability of BH end INS and end EXP states at the same airway pressure compared to the intra-state variability when combining BH end INS and end EXP states as a one group. The values shown represent the mean + standard deviation over all subjects.

Full size image

Figures 5a and b show the effect of varying the pressure at end INS and end EXP, respectively. Here, a reference state of − (15–20) cm H2O pressure was chosen. Then the inter-state variability was calculated between this reference state and the states at other pressure levels. The values shown in Fig. 4 represent the mean + standard deviation over all subjects.

Figure 5
figure 5

Inter-state variability for BH (a) end INS (b) end EXP. The reference state is the lowest pressure (i.e., Negative 15–20 cm H2O). The values shown represent the mean + standard deviation over all subjects.

Full size image

The linear trend between pressure and inter-state variability was evaluated according the criteria provided by an earlier study37, which states that a linear trend is considered statistically meaningful if the coefficient of determination (r2) is greater than 0.65 and the p value is less than 0.05. Results showed that the trend of an increased inter-state variability with pressure is statistically meaningful for BH at end EXP (r2 = 0.91, p = 0.048) but not for BH at end INS (r2 = 0.69, p = 0.172). This observed trend suggests a progressive change in the SCG waveform with airway (and intrathoracic) pressure changes. The lack of statistical difference may be because only four pressure ranges were considered. More pressures levels may be needed to confirm this finding.

Limitations of the study include a relatively small number of participants and small number of airway pressure levels. The intrathoracic pressure was not measured directly due to the invasiveness of the typical methods used for that measurement. Therefore, the airway pressure was measured instead, which would directly correlate with intrathoracic pressure, especially at breath hold.

Conclusions

SCG variability was investigated for breath hold at different airway pressures (i.e., 0, ± 2–4, and ± 15–20 cm H2O) and two different lung volumes (i.e., end inspiration and end expiration). Variability was also studied for normal breathing with and without the implementation of unsupervised machine learning (i.e., K-medoid algorithm) to group SCG beats into two clusters of similar waveform morphologies.

  • Clustering normal breathing SCG and breath holding at zero airway pressure reduced waveform variability. Variability reduction in the latter case was stronger possibly because the intrathoracic pressure was more constant during BH than during normal breathing.

  • As the airway pressure deviated from zero for the BH states, the intra-state variability increased. The mechanism underlying this trend needs more investigation but may be related to the effect of the intrathoracic pressure on cardiac preload and afterload.

  • For BH states, the subaudible to audible energy ratio increased for low airway pressures (and intrathoracic pressure). This may be attributed to a higher afterload and the accompanying possible increase in the muscle contraction force and the reduction in the muscle contraction velocity.

  • Combining BH end INS and end EXP states as a one group increased the intra-state variability, which suggests that lung volume and heart location are possible sources of SCG variability.

  • There was a significant linear trend between the airway pressure and the inter-state variability for BH end EXP states (but not for end INS), which suggests a progressive change in the SCG waveform with airway (and intrathoracic) pressure changes.

  • Since breathing maneuvers affect SCG waveforms, such maneuvers may provide features that correlate with cardiac contractility and may be useful for diagnosis and monitoring of heart conditions.

  • Future studies may include investigating the effect of intrathoracic pressure at more pressure levels and at respiration phases other than the ones considered in the current study (e.g., mid-inspiration, mid-expiration, etc.), and confirming observed tends in a larger number of subjects.

  • The autonomic nervous system (ANS) regulates circulation by controlling the heart rate and contractility38, which may be related to trends seen in the current study. However, the specific mechanism of the ANS response to intrathoracic pressure variation needs more investigation and would be performed in future studies.