Content uploaded by Jonathan Zia
Author content
All content in this area was uploaded by Jonathan Zia on Oct 18, 2019
Content may be subject to copyright.
Automated Identification of Persistent Time-Domain
Features in Seismocardiogram Signals
Jonathan Zia, Student Member, IEEE, Jacob Kimball, Student Member, IEEE
Md Mobashir Hasan Shandhi, Student Member, IEEE, and Omer T. Inan, Senior Member, IEEE
Abstract—In the field of cardiac monitoring, the seismocar-
diogram (SCG) measures the movement of the chest wall using
accelerometers and gyroscopes. A key limitation of SCG signals
is their sensitivity to transient signal disruptions primarily due
to motion artifacts. This work describes a method for automated
extraction of time-domain features in SCG signals in the presence
of such artifacts, using an iterative method of clustering and
re-sampling features to optimize consistency. The accelerometer
(axl) and gyroscope (gyr) features extracted with this method are
shown to correlate more strongly (median R2= 0.88 (axl), 0.88
(gyr)) with the reference standard for pre-ejection period (PEP),
impedance cardiography (ICG), than both peak-counting (R2=
0.29 (axl), 0.48 (gyr)) and manual labeling (R2= 0.44 (axl), 0.38
(gyr)) in the post-exercise period. This result has implications for
the feasibility of at-home SCG monitoring.
Index Terms—seismocardiography, feature extraction, cardiac
monitoring, automation
I. INTRODUCTION
The advent of wearable sensing has changed the landscape
of healthcare delivery by enabling the continuous and unob-
trusive monitoring of patients’ health [1]. One modality of
wearable sensing that is being investigated recently is the
seismocardiogram (SCG) measurement—a signal representing
the vibrations of the chest wall in response to the movement
of the heart and blood. Notably, SCG waveform features
have been demonstrated to correlate with cardiac cycle events,
including aortic valve opening (AO) and closing (AC) [2] [3].
Such data can be used to calculate important indicators such as
the pre-ejection period (PEP), defined as the interval between
ventricular depolarization and AO. Continuous monitoring
of such indicators may serve to improve care for patients
suffering from heart disease [4].
A key limitation of the SCG is the presence of transient
disruption, making it difficult to consistently identify sensitive
time-domain features of the waveform [5]. This is especially
true during periods of rapid cardiovascular adaptation—such
as during exercise recovery—or when the patient is moving.
Prior research estimating AO or other points of interest
from SCG waveforms includes abstraction methods such as
using the time-frequency spectrum [6], wavelet decomposition
[7], envelope functions [8], and ensemble-averaged templates
This material is based on work supported by the National Institutes of
Health under Grant 1R01HL130619-A1 and the National Center for Advanc-
ing Translational Sciences of the National Institutes of Health under Award
Number UL1TR002378.
J. Zia, J. Kimball, M.M.H. Shandhi, and O. T. Inan are with the School of
Electrical and Computer Engineering at the Georgia Institute of Technology,
Atlanta, GA 30332 USA (email: zia@gatech.edu).
[9] among others. Peak-counting methods have been quite
prevalent, namely selecting numbered peaks from the wave-
form to represent cardiac events [10]. Finally, sensor fusion—
especially across modalities—has emerged as a method of
improving SCG feature extraction [11]. While effective in
predicting cardiac cycle events, such methods may be imprac-
tical in noisy environments. In fact, prior work demonstrating
correlation between SCG and cardiac events has often relied
on manual labeling, which is prone to human error. Conse-
quently, methods that utilize supervised training algorithms are
inherently limited by the efficacy of manual labeling. However,
these results suggest that if SCG features could be extracted
consistently, changes in these features would strongly correlate
with changes in cardiac events or indicators such as PEP
[12]. By extension, this would improve the feasibility of SCG
monitoring in at-home environments.
This work presents a method for identifying time-domain
features of the SCG segments while ignoring transient disrup-
tions in waveform morphology. Rather than identifying fiducial
points, this method extracts features based on their consistency,
and optionally their covariance with PEP. The desired features
in this study include the first and second consistent minima and
maxima of the SCG interval—which are frequently correlated
with the AO point—however, the approach may generalize to
any time-domain features. Figure 1(a) illustrates that when
these features are extracted using a simple peak-counting
method—returning the first two local minima and maxima—
transient disruptions can result in feature corruption.
II. ME TH OD S
A. Experimental Protocol
A total of 17 healthy subjects (10 males, 7 females, Age:
26.1 ±4.1 years, Weight: 66.2 ±13.6 kg, and Height: 168.2 ±
8.9 cm) with no history of heart disease participated in a study
conducted under a protocol approved by the Georgia Institute
of Technology Institutional Review Board [13]. The purpose
of this study was to induce changes in PEP while monitoring
changes in SCG. To this end, a three-axis accelerometer and
gyroscope were placed on the subject’s sternum to collect SCG
waveforms, along with reference electrocardiogram (ECG) and
impedance cardiogram (ICG) sensors. After standing vertically
and motionless for five minutes, the subject performed three
minutes of walking at three miles per hour on a treadmill
followed by 90 seconds of a squatting exercise. Subsequently,
the subject again stood vertically and motionless for a five-
minute recovery period.
Raw Annotated Features using Peak-Counting
Time (ms)
25 50 75 100 125 150
Acceleration (m/s2)
0
- 20
- 15
- 10
- 5
0
5
10
15
20
25
- 25
Feature Selection using Algorithm 1
Time (ms)
25 50 75 100 125 150
0175 200
Acceleration (m/s2)
- 20
- 15
- 10
- 5
0
5
10
15
20
25
- 25
(a) (b)
Fig. 1. (a) Raw features for a SCG signal during post-exercise period. Features include first and second minima (black, blue), and first and second maxima
(red, green). (b) Updated features after using Algorithm 1 on signals in (a). Notably, finalized features ignore transient peaks in the range 0 - 50 ms.
B. Signal Processing and Feature Extraction
Signal Processing: The data used in this study was limited
to the recovery period due to the subjects’ rapidly-changing
hemodynamic state during this interval. All signals were fil-
tered with a finite impulse response (FIR) band-pass filter with
Kaiser window. Cutoff frequencies were: 1-40 Hz for SCG, 1-
30 Hz for ICG, and 0.5-40 Hz for ECG. These signals were
then segmented into beat-by-beat intervals using the R-peaks
on the reference ECG signal. Segments for each modality were
then smoothed with an exponential moving average filter with
a window of five beats [14].
ICG Feature Extraction: The reference standard for the
aortic opening (AO) point in this study is the B-point of the
ICG waveform, identified as the point of maximal inflection
of the wave in the period preceding the signal peak [15]. Since
ICG is a measure of intra-thoracic bloodflow, the inflection of
the waveform is coupled with ejection of blood through the
aortic valve during systole. Thus, the R-B period serves as a
correlate to the true PEP.
Manual SCG Labeling: Performance of the automated al-
gorithm presented below is compared against manual feature
extraction. For this task, a trained annotator was instructed to
mark the first and second minima and maxima of the SCG
waveform, choosing peaks consistently between intervals and
ignoring transient disturbances [1]. To aid in this, the annotator
was presented with a reference SCG waveform, which was the
ensemble-average of signal segments for each trial.
C. Automated SCG Feature Identification
For both the accelerometer- and gyroscope-based SCG data,
an algorithm was developed for identifying the set of time-
domain features of interest (the first and second minima and
maxima following the ECG R-peak), while ignoring transient
or aberrant peaks that are not present throughout the entirety
of the signal. The feature extraction algorithm is summarized
in Algorithm 1.
Each feature is first extracted using the peak-counting
method described above. In the presence of aberrant peaks,
the raw features may be shifted in time in adjacent seg-
ments, resulting in the apparent clustering of features into
distinct distributions as shown in Figure 2(a). To identify
these distributions, Gaussian mixture modeling (GMM) is
used to determine the probability that each sample belongs
in one of up to Mdistributions, with each sample being
clustered according to the distribution to which it most likely
belongs. Note that, though the data in Figure 2(a) is linearly
separable, this is not always the case — thus, a clustering
algorithm was chosen rather than a linear classifier due to
its superior generalizability. The optimal number of clusters
Algorithm 1 SCG Feature Extraction
1: procedure GET FEATU RE S(M , N, F)
2: for f∈Fdo Loop through features
3: Obtain raw feature vector rf
4: P←0Initialize performance
5: while ∆P > 0do
6: for n= 1 −→ Ndo Order of b.f. line
7: Cluster rfinto set Cs.t. |C|=M?≤M
8: Generate best-fit
bm,n for Cm∈C
9: Resample rfalong
bm,n to obtain sm,n
10: Calculate performance pm,n of sm,n
11: m?, n?←argmaxm,n p
12: ∆P←pm?,n?−P
13: if ∆P > 0then
14: rf←sm?,n?
15: P=pm?,n?
16: return R:={r1...r|F|}
Gaussian Mixture Model of Raw SCG Features
Normalized Segment
Normalized Amplitude
- 1.5
- 1.0
- 0.5
0
0.5
1.0
-1.5 -1.0 -0.5 00.5 1.0
-2 1.5 2.0
Updated Feature Vector after Re-Sampling
100 200 300 400
0500
0
1.0
0.2
0.4
0.6
0.8
Scaled Amplitude
Signal Segment
(a) (b)
Fig. 2. (a) Clustering of features corresponding to the first minimum of SCG segments for signals shown in Figure 1(a). Features assigned to clusters 1 and
2 are shown in black and red respectively, along with the contour of the Gaussian model associated with each cluster. (b) Updated feature vector (blue line)
after the first iteration of re-sampling the raw vector (blue area). Best-fit lines corresponding to the two clusters in (a) are overlaid (black, dashed), with the
updated vector corresponding to re-sampling along the uppermost line. The PEP derived from the ICG B-point is provided (green area).
M?is determined by performing GMM iteratively on the raw
feature vector, and selecting the number that yields the highest
average silhouette score, defined as:
SC=1
|C|
|C|
X
i=1
β(i)−α(i)
max(β(i), α(i)) (1)
for cluster Cin the set of all clusters C. In (1), β(i)is the
average distance between sample iand points in the next-
nearest cluster, and α(i)is the average distance between iand
points in the same cluster. As rfis a time-series, separation
in time does not contribute to distance in (1). An example
of performing GMM on the data in Figure 1(a) is shown in
Figure 2(a). Note that such clustering accounts for the natural
shift in feature values during the recovery period.
Features should be drawn from the distribution that results
in a feature vector with the lowest variability, assuming the
relative beat-by-beat consistency of cardiac parameters. To
determine the proper distribution, Nbest-fit lines of poly-
nomial order 1 to Nare fit to the samples in each cluster.
The original vector rfis then re-sampled such that, of the
candidate features in the signal, the feature closest to the
expected value given by the best-fit
bm,n is chosen. Naturally,
distributions that represent inconsistent features will result in
higher variability. The performance of each re-sampled vector
sm,n (for cluster mof best-fit order n) is thus calculated as:
pm,n =1
¯σm,n
(2)
where ¯σm,n is the sample-averaged expected standard devi-
ation of
bm,n. The highest-performing vector is then chosen
to update the original feature vector rf, and the process is
iteratively performed until convergence upon the optimal set
of features. Example results are shown in Figures 1(b) and
2(b) when the process is applied to the data in Figure 2(a).
III. RES ULTS
The process described in Algorithm 1 was used to extract
the aforementioned time-domain features from accelerometer-
and gyroscope-based SCG signals for 16 of the 17 subjects
in the study, as the reference ECG for one of the subjects
was not usable. Each set of features was then compared to the
PEP estimated from the corresponding ICG signal. Quality of
the feature was assessed using the coefficient of determination
(R2), which quantifies how much variance in one signal is
explained by the variance in the other. The R2equation used
in this study was adjusted for the number of observations:
R2= 1 −n−1
n−pSS E
SS T (3)
where nis the number of observations, pis the number of
polynomial coefficients, SSE is the sum squared error, and
SS T is the sum squared total.
The results of this analysis are shown in Figure 3. The
automated algorithm significantly outperformed both manual
labeling and raw (peak-counting) feature extraction in identi-
fying the desired features. For the accelerometer, median R2
was 0.88 (IQR: 0.77 - 0.92) for automated features, 0.44 (0.30
- 0.63) for manual features, and 0.29 (0.15 - 0.36) for raw
features. For the gyroscope, median R2was 0.88 (0.81 - 0.91)
for automated features, 0.38 (0.22 - 0.56) for manual features,
and 0.48 (0.24 - 0.77) for raw features.
IV. DISCUSSION AND CONCLUSION
Automated features notably had both a higher mean R2
and lower IQR than the other labeling methods. Furthermore,
gyroscope features had a comparable correlation coefficient
Fig. 3. Performance of Algorithm 1 in feature extraction, reported as average
R2between the time domain features and PEP derived from ICG. Green
boxes represent accelerometer data and blue boxes represent gyroscope data.
with the PEP estimate than accelerometer features. These
results suggest that the automated algorithm was more effec-
tive in identifying consistent features than manual labeling
and raw feature extraction, and supports recent works that
present gyroscopes as a viable alternative to accelerometers
for SCG. Since manual labeling is a very time consuming
task; automating such tasks would greatly benefit researchers
in this field.
As the purpose of this approach is to identify the most
consistent features in the signal set, it is not intended as a
method of AO point or PEP estimation, but rather identifying
consistent features that co-vary with these metrics of interest.
This is most useful in domains such as adverse event pre-
diction, where the patient’s relative deviation from baseline
PEP is more important than the value of PEP itself. Equation
(2) may thus be modified to select for feature vectors that best
achieve this, for example defining performance as R2between
the feature and the reference standard for AO.
Though the ICG B-point was used as the reference standard,
echocardiogram is the gold-standard for AO point identifica-
tion. Though SCG has been used extensively for this purpose,
a universal translation of SCG features to cardiac cycle events
has yet to be confirmed, and thus event prediction from SCG
often relies on additional mapping of feature vectors to event
timing that may not generalize. For this reason, perhaps the
most judicious use of SCG in cardiac monitoring is to predict
relative rather than absolute event timings.
A possible limitation of this study is that the automated
algorithm was tested against two trained annotators; future
work should test the automated approach against a larger set
of annotators to generalize results.
The method presented in this work has the benefit of
selecting features based on the entire five-minute span of data,
however it could be modified to perform this analysis in real-
time settings. Specifically, the Gaussian mixture model may be
updated dynamically as data becomes available, and feature
selection may incorporate both past and present data, such
as with Kalman filtering. This would be useful as part of
a larger system for continuous monitoring that also includes
signal quality verification and noise reduction components.
Future work should extend this method to real-time systems
for continuous monitoring in at-home environments. This
method should also be evaluated directly against prior methods
of SCG feature extraction.
REFERENCES
[1] O. T. Inan, P.-F. Migeotte, K.-S. Park, M. Etemadi, K. Tavakolian, R.
Casanella, J. Zanetti, J. Tank, I. Funtova, G. K. Prisk, and M. D. Rienzo,
“Ballistocardiography and Seismocardiography: A Review of Recent
Advances,” IEEE Journal of Biomedical and Health Informatics, vol.
19, no. 4, pp. 1414-1427, 2015.
[2] R. S. Crow, P. Hannan, D. Jacobs, L. Hedquist, and D. M. Salerno, “Re-
lationship between Seismocardiogram and Echocardiogram for Events
in the Cardiac Cycle,” American Journal of Noninvasive Cardiology,
vol. 8, no. 1, pp. 39-46, 1994.
[3] M. J. Tadi, E. Lehtonen, A. Saraste, J. Tuominen, J. Koskinen, M. Ters, J.
Airaksinen, M. Pnkl, and T. Koivisto, “Gyrocardiography: A New Non-
invasive Monitoring Method for the Assessment of Cardiac Mechanics
and the Estimation of Hemodynamic Variables,” Scientific Reports, vol.
7, no. 1, 2017.
[4] J. M. Zanetti and K. Tavakolian, “Seismocardiography: Past, present and
future,” 35th Annual International Conference of the IEEE Engineering
in Medicine and Biology Society (EMBC), 2013.
[5] A. Q. Javaid, H. Ashouri, A. Dorier, M. Etemadi, J. A. Heller, S.
Roy, and O. T. Inan, “Quantifying and Reducing Motion Artifacts in
Wearable Seismocardiogram Measurements During Walking to Assess
Left Ventricular Health,” IEEE Transactions on Biomedical Engineering,
vol. 64, no. 6, pp. 1277-1286, 2017.
[6] A. Taebi and H. Mansy, “Time-Frequency Distribution of Seismocardio-
graphic Signals: A Comparative Study,” Bioengineering, vol. 4, no. 4,
p. 32, 2017.
[7] T. Choudhary, L. N. Sharma, and M. K. Bhuyan, “Automatic Detection
of Aortic Valve Opening using Seismocardiography in Healthy Individ-
uals,” IEEE Journal of Biomedical and Health Informatics, pp. 1–1,
2018.
[8] F. Khosrow-Khavar, K. Tavakolian, A. Blaber, and C. Menon, “Auto-
matic and Robust Delineation of the Fiducial Points of the Seismocar-
diogram Signal for Noninvasive Estimation of Cardiac Time Intervals,”
IEEE Transactions on Biomedical Engineering, vol. 64, no. 8, pp. 1701-
1710, 2017.
[9] G. Shafiq, S. Tatinati, W. T. Ang, and K. C. Veluvolu, “Automatic
Identification of Systolic Time Intervals in Seismocardiogram,” Scientific
Reports, vol. 6, no. 1, 2016.
[10] M. D. Rienzo, E. Vaini, and P. Lombardi, “Use of seismocardiogram
for the beat-to-beat assessment of the Pulse Transit Time: A pilot
study,” 37th Annual International Conference of the IEEE Engineering
in Medicine and Biology Society (EMBC), 2015.
[11] C. Yang, S. Tang, and N. Tavassolian, “Annotation of seismocardiogram
using gyroscopic recordings,” IEEE Biomedical Circuits and Systems
Conference (BioCAS), 2016.
[12] H. Ashouri, S. Hersek, and O. T. Inan, “Universal Pre-Ejection Period
Estimation Using Seismocardiography: Quantifying the Effects of Sen-
sor Placement and Regression Algorithms,” IEEE Sensors Journal, vol.
18, no. 4, pp. 1665–1674, 2018.
[13] M. Shandhi, B. Semiz, S. Hersek, N. Goller, F. Ayazi, and O. T. Inan,
“Performance Analysis of Gyroscope and Accelerometer Sensors for
Seismocardiography-Based Wearable Pre-Ejection Period Estimation,”
IEEE Journal of Biomedical and Health Informatics, 2019, In Press.
[14] M. Etemadi, O. T. Inan, L. Giovangrandi, and G. T. A. Kovacs, “Rapid
Assessment of Cardiac Contractility on a Home Bathroom Scale,” IEEE
Transactions on Information Technology in Biomedicine, vol. 15, no. 6,
pp. 864–869, 2011.
[15] A. Sherwood, M. T. Allen, J. Fahrenberg, R. M. Kelsey, W. R. Lovallo,
and L. J. P. Vandoornen, “Methodological Guidelines for Impedance
Cardiography,” Psychophysiology, vol. 27, no. 1, pp. 1–23, 1990.
