Mapping Seismocardiogram Characteristics to Hemorrhage Status and

Vascular Pressure: A Novel Approach for Triage Assessment

Zeynep Deniz Gundogan

and Beren Semiz

2 a

Department of Electrical Engineering, Istanbul Technical University, Istanbul, Turkey

Department of Electrical and Electronics Engineering, Koc University, Istanbul, Turkey

Keywords:

Seismocardiogram, Hemorrhage, Vascular Pressure, Biomedical Signal Processing.

Abstract:

When a mass incident occurs, determining the severity of injuries and arranging the hospital triage are of great

importance to increase the survival rates. This study aims to develop a seismocardiogram (SCG)-based triage

assessment system by (i) distinguishing between different levels of exsanguination, and (ii) estimating the vas-

cular pressure values recorded from various body locations for prioritizing the triage processes and monitor-

ing vital parameters. In this project, publicly available Wearable and Catheter-based Cardiovascular Signals

During Progressive Exsanguination in a Porcine Model of Hemorrhage dataset, which includes cardiovascular

signals acquired through a catheter-based system and wearable sensors during progressive exsanguination, was

used. First, temporal and spectral features were extracted from the SCG signals taken at different blood-loss

levels from six Yorkshire swines. Hemorrhage severity assessment was then performed through multi-class

classiﬁcation leveraging one vs. all approach. As the second step, four different regression models were

trained for each of the right atria, aortic root, femoral artery and pulmonary capillary locations to estimate

the corresponding vascular pressure values. For hemorrhage severity assessment, the accuracy, sensitivity,

precision and f1-score values were all calculated to be 0.96 for the best performing model (XGBoost). For

the vascular pressure estimation, (mean-absolute-error and R

) pairs were calculated to be (1.54, 0.94), (2.76,

0.58), (1.29, 0.87) and (0.95, 0.90) for aortic root, femoral artery, right atrium and pulmonary capillary models,

respectively. Overall, this study introduced new use areas for the SCG signal, which can potentially be utilized

in the development of continuous and non-invasive monitoring systems to prioritize the triage processes and

track vital parameters.

1 INTRODUCTION

When a mass incident, such as an earthquake or pub-

lic transport accident, occurs, determining the sever-

ity of injuries and arranging the hospital triage (i.e.

order of treatment) are of great importance to increase

the survival rates. Considering that the number of ﬁrst

responders and provided resources are limited, pro-

viding immediate and timely treatment to all victims

is often not possible. Additionally, it is extremely

important to monitor and detect deterioration risks

in health status for preventing any follow-up com-

plications. Therefore, there is a need for innovative

approaches that can offer continuous hemodynamic

monitoring to enable appropriate triage assessment.

As physiological signals emerge directly from the

body, they hold valuable clinical information about

the underlying physiological conditions and irreg-

https://orcid.org/0000-0002-7544-5974

ularities. In wearable device design, three of the

most commonly used physiological signals can be

listed as the electrocardiogram (ECG), photoplethys-

mogram (PPG), and seismocardiogram (SCG) wave-

forms. The SCG represents the mechanical activity of

the heart resulting from the cardiac ejection and con-

traction, whereas the ECG acquires the electrical sig-

nal of the heart to assess rhythm and rate (Inan et al.,

2014). Recent studies have shown that SCG signal

can potentially be used in heart failure classiﬁcation

(Inan et al., 2018), myocardial contraction assessment

(Tavakolian et al., 2012), respiration phase analysis

(Imirzalioglu and Semiz, 2022; Pandia et al., 2012),

valvular heart disease assessment (Erin and Semiz,

2023), metabolic equivalent of task score estimation

(Tokmak and Semiz, 2022), systolic time interval and

hemodynamic parameter estimation (Shandhi et al.,

2019; Semiz et al., 2020). On the other hand, ﬂuc-

tuations in arterial blood volume during the cardiac

646

Gundogan, Z. and Semiz, B.

Mapping Seismocardiogram Characteristics to Hemorrhage Status and Vascular Pressure: A Novel Approach for Triage Assessment.

DOI: 10.5220/0012406700003657

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 1, pages 646-655

ISBN: 978-989-758-688-0; ISSN: 2184-4305

cycle cause variations in the light absorption within

the arteries, leading to the generation of PPG. Indeed,

previous studies have shown that blood pressure, oxy-

gen saturation and vascular resistance assessment can

highly beneﬁt from PPG analysis (cheol Jeong et al.,

2018).

Among these signals, the researchers have re-

cently found that the SCG signal is the most important

modality in predicting decomposition (Kimball et al.,

2021). Although there have been several studies on

electronic triage tag development leveraging breath-

ing rate, oxygen saturation, blood pressure, temper-

ature, and heart rate assessment (Sakanushi et al.,

2013; Rodriguez et al., 2014; Park, 2021; Gr

unerbel

et al., 2023), there is no study focusing on the assess-

ment of different levels of blood loss and correspond-

ing vascular pressure values using solely the SCG sig-

nals. Hence, in this work, the fundamental aims were

to investigate the relationship between blood loss lev-

els and SCG signals, and to map the SCG signal char-

acteristics to the vascular pressure values recorded

from various body locations for prioritizing the triage

processes and monitoring vital parameters.

The contributions of this study are threefold: For

the ﬁrst time, (i) SCG signal characteristics have been

leveraged to distinguish between different levels of

exsanguination, which can potentially accelerate the

triage processes, (ii) SCG signal characteristics have

been used to estimate the femoral artery, aortic root,

right atrium and pulmonary capillary wedge pressure

values to achieve continuous vascular pressure moni-

toring, (iii) it has been shown that the temporal char-

acteristics of the SCG signal has relatively higher im-

portance compared to the spectral ones in the case

of hemorrhage assessment. Overall, the study intro-

duces new use areas for the SCG signal, which can

potentially be utilized in the development of continu-

ous and non-invasive monitoring systems to prioritize

the triage processes and track vital parameters.

2 MATERIALS AND METHODS

2.1 Dataset Description

2.1.1 Experimental Protocol

In this project, publicly available Wearable and

Catheter-based Cardiovascular Signals During Pro-

gressive Exsanguination in a Porcine Model of

Hemorrhage dataset was used (Zia et al., 2020a).

The dataset contains cardiovascular signals acquired

through a catheter-based system and wearable sensors

during progressive exsanguination in a porcine model

of hemorrhage. The signals from the wearable sen-

sors (seismocardiogram, electrocardiogram and pho-

toplethysmogram) were acquired through BIOPAC

MP160 system (BIOPAC Systems, Inc., Goleta, CA,

USA). On the other hand, vascular pressure values

from the aortic arch, femoral artery, right atrium and

pulmonary capillary locations were collected with the

catheter-based setup. Data from the catheters were

acquired through the ADInstruments Powerlab 8/35

system. All signals were sampled at 2 kHz.

The study was conducted on 6 Yorkshire swines

(3 castrated female, 3 male, weight: 51.5-71.4 kg,

age: 114-150 days). Each swine underwent a health

assessment examination and no other exclusion crite-

ria was designated in the study. Anesthesia was ad-

ministered with xylazine and telazol, and sustained

via inhaled isoﬂurane during mechanical ventilation.

During the experimental protocol, blood was drained

through an arterial line at four levels (7%, 14%, 21%,

and 28%) to induce hypovolemia. After every stage

of blood loss, the process of exsanguination was tem-

porarily halted for nearly 5-10 minutes so that the

cardiovascular functions could return to normal. It

should be noted that the experimental protocol was

terminated at different blood levels for each pig as

cardiovascular collapse was reached at different lev-

els. More speciﬁcally, pig 5 reached 14% blood vol-

ume loss; pigs 1, 3, and 4 reached 21% blood volume

loss; and pigs 2 and 6 reached 28% blood volume loss.

These levels were visualized in Fig. 1(b).

2.1.2 Pre-Processing and Data Preparation

First, the signals were ﬁltered with ﬁnite impulse re-

sponse (FIR) band-pass ﬁlters with the following cut-

off frequencies: SCG (1-40 Hz), ECG (0.5-40 Hz),

PPG and catheter-based pressure signals (0.5-10 Hz).

The signals were then segmented into individual beats

using the R-peak locations of the ECG signal. Re-

sulting segments were truncated into a length of 1000

samples (500 ms), except the third pig (1500 samples,

750 ms) due to relatively longer left ventricular ejec-

tion time. The catheter-based pressure signals were

also truncated in a similar way.

Analysis pipeline is presented in Fig. 1(a). Two

different analyses, hemorrhage severity assessment

and vascular pressure estimation, were performed us-

ing the SCG signals taken from the dorso-ventral axis.

For hemorrhage severity assessment, blood-loss lev-

els at the time of each SCG beat were available as

a separate vector. For the vascular pressure estima-

tion analysis, the average of 1000 pressure samples

was taken to have one single value for each pressure

location. By this way, each SCG beat could be rep-

resented by one pressure value for each of the aortic

Mapping Seismocardiogram Characteristics to Hemorrhage Status and Vascular Pressure: A Novel Approach for Triage Assessment

647

Figure 1: (a) Analysis pipeline (b) Green color: Blood loss levels that the pigs could reach, Red color: Blood loss levels that

the pigs could not reach due to cardiovascular collapse.

arch, femoral artery, right atrium and pulmonary cap-

illary locations.

2.2 Feature Extraction

Based on the previous studies, SCG analysis can

highly beneﬁt from time domain analysis, as val-

leys and peaks correspond to speciﬁc cardiovascular

events in cardiac cycle (such as mitral closing (MC),

aortic opening (AO), etc.) (Semiz et al., 2020). On the

other hand, recent studies have revealed that spectral

content of the SCG signals also contains salient in-

formation regarding underlying physiological events

and pathologies (Erin and Semiz, 2023). Hence, each

SCG signal segment was analyzed both in temporal

and spectral domains. The sets of features are detailed

in the following subsections.

2.2.1 Temporal Domain Analysis

For temporal domain analysis, energy, zero crossing

rate, entropy and peak-location features were com-

puted for each SCG segment. While energy corre-

sponds to the total energy (i.e. sum of squared mag-

nitudes of the samples), energy entropy assesses the

abrupt changes observed in the signal energy. On

the other hand, zero crossing rate measures the sign-

change rate and used to evaluate the overall noise

level. In addition to these three temporal features, the

ﬁrst three minima and maxima locations and ampli-

tudes were computed. This step resulted in an ad-

ditional 12 features (6 amplitude and 6 location val-

ues). In the analysis, the valleys and peaks were

not associated with particular cardiac events such as

AO, MC, etc. This is due to two reasons - ﬁrst, the

SCG signals exhibit signiﬁcant variability among dif-

ferent subjects, making it challenging to detect those

points precisely. Second, the shape of the SCG sig-

nal changes depending on the sensor’s location or the

subject’s posture (Hersek et al., 2019).

Additionally, statistical features were calculated

for each segment. As statistical features, mean, root

mean square, variance, satandard deviation, skewness

and kurtosis were extracted. After computing the

aforementioned features, a matrix T

was generated

for time domain features where each row corresponds

to one segment, i, and each column represents one

feature for each segment i. Overall, there were 21

temporal features extracted from each segment.

2.2.2 Spectral Domain Analysis

Following temporal analysis, spectral analysis was

performed on the SCG segments. To that aim, spec-

tral entropy, rolloff, spread and centroid values were

computed for each segment i. While spectral entropy

represents the complexity of the spectrum rolloff cor-

responds to the frequency below which a certain per-

centage of the signal energy is accumulated. Con-

versely, spectral centroid and spread pertain to the

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

648

center of mass and frequency distribution of the spec-

trum, respectively.

In addition to the aforementioned four spectral

features, spectrogram analysis was employed. Spec-

trogram is used to investigate how the frequency

characteristics of a non-stationary signal change with

time. As the original function outputs frequency, time

and coefﬁcient values together, the average coefﬁ-

cients for each frequency band were computed over

time and then stored as a vector. In total 129 spec-

trogram coefﬁcients were extracted from each seg-

ment. After computing the aforementioned features, a

matrix S

was generated for spectral domain features

where each row corresponds to one segment, i, and

each column represents one feature for each segment

i. Overall, there were 133 spectral features extracted

from each segment.

2.2.3 Dataframe Generation

In total, there were 154 features (21 temporal and

133 spectral) extracted from each SCG segment. For

convenience of feature selection and model training,

one single data frame was created using these fea-

tures. In the dataframe, each row was correspond-

ing to one SCG segment i, whereas the columns

were including the extracted features (time domain

and spectral domain S

matrices). In addition to

the feature columns, ﬁve additional columns were

added to include blood-loss level labels (b

) and

the reference aortic root (p1

), femoral artery (p2

right atria (p3

) and pulmonary capillary wedge (p4

)

pressure values. Blood-loss levels were determined

as 0, 1, 2, 3, 4 representing 0% (baseline), 7%,

14%, 21%, and 28% blood-loss levels, respectively.

Overall, the dataframe was structured as follows:

, S

, b

, p1

, p2

, p3

, p4

2.2.4 Variance Thresholding

To prevent curse of dimensionality caused by high

number of features, feature selection was applied on

the 154 features extracted. As the feature selection

method, variance thresholding, which seeks to elim-

inate features with variance values that fall below a

speciﬁc threshold, was chosen. Since any feature’s

variance and level of predictive ability are correlated,

features with smaller variance convey relatively less

information (Bommert et al., 2020). In accordance

with the previous study, threshold value was deter-

mined as 0.0001 (Erin and Semiz, 2023). Features be-

low this threshold were masked as False and dropped,

whereas the remaining features (which were above

the threshold) were masked as True and kept in the

analysis.

2.3 Model Training

Under model training, two different tasks were im-

plemented: hemorrhage severity assessment and vas-

cular pressure estimation. As detailed in Sec. 2.2.3,

for hemorrhage severity classiﬁcation, blood-loss lev-

els were determined as 0, 1, 2, 3, 4 representing 0%

(baseline), 7%, 14%, 21%, and 28% blood-loss lev-

els, respectively. On the other hand, for vascular

pressure estimation, four different regression models

were trained for each of the aortic root, right atrium,

femoral artery and pulmonary capillary wedge pres-

sure estimation tasks. Both for hemorrhage sever-

ity assessment and vascular pressure assessment, four

different models were trained and corresponding per-

formance values were compared. It should be noted

that machine learning models were leveraged instead

of deep learning approaches due to the limited sample

size.

• Multi Layer Perceptron (MLP): MLP falls un-

der artiﬁcial neural networks composed of multi-

ple layers of neurons (i.e., nodes), which are inter-

connected via weighted connections. From input

layer to output layer, hidden layers perform com-

plex transformations on the data. More specif-

ically, each neuron in hidden layers computes

weighted sum of the inputs, employs non-linear

activation function and transfers the result to the

following layer. The weights of these neuron con-

nections are learned and updated during training

through back-propagation (Bishop et al., 1995).

• Support Vector Machines (SVM): SVM aims to

ﬁnd the best separating hyperplane while maxi-

mizing the margin between the classes. In this

context, the margin corresponds to the distance

between the hyperplane and closest data points

from different classes. SVM can also be used

in regression tasks where the aim is to ﬁnd the

best hyperplane that minimizes the sum of the

distances between the predicted and actual values

(Noble, 2006).

• Random Forest (RF): RF is a type of ensemble

methods where instead of using a single tree, it in-

volves generating multiple trees by randomly se-

lecting subsets from the original dataset. These

trees are trained separately and in parallel, and

their individual predictions are combined by av-

eraging them to produce the ﬁnal predicted target

value (Breiman, 2001).

• Extreme Gradient Boosting (XGBoost): XG-

Boost also belongs to the ensemble methods cat-

egory, more speciﬁcally the gradient boosting al-

gorithm. This approach involves using multiple

Mapping Seismocardiogram Characteristics to Hemorrhage Status and Vascular Pressure: A Novel Approach for Triage Assessment

649

estimators simultaneously instead of a single esti-

mator to predict a variable. The algorithm trains

numerous decision trees in a sequential manner,

enabling the model to forecast the leftover errors

from the previous round and enhance its perfor-

mance progressively (Chen and Guestrin, 2016;

Friedman, 2001).

2.3.1 Hemorrhage Severity Assessment

In the classiﬁcation task, the idea was to assess the

performance of the features extracted from the SCG

signal in distinguishing between different blood loss

levels. Using the features remained following vari-

ance thresholding, four different classiﬁcation mod-

els were trained using MLP, SVM, RF and XGBoost.

As explained in Section 2.1.1 and Fig. 1(b), the ex-

perimental protocol was terminated at different blood

levels as each pig reached cardiovascular collapse at

different levels. While pigs 1, 3, and 4 reached 21%

blood volume loss, pigs 2 and 6 could reach 28%

blood volume loss, and pig 5 could only reach 14%

blood volume loss. Thus, there was an imbalance

within the dataset: pigs 2 and 6 indeed had data corre-

sponding to each phase, whereas the other pigs were

missing one or two labels. Hence, instead of building

a leave-one-subject-out cross-validation framework,

k-fold cross validation was employed during model

training to prevent the bias that will occur across

folds.

The data was ﬁrst split into k number of sub-

groups (folds). In each iteration, one fold was left

out for testing and the model was trained with the

remaining (k-1) folds. The iterations continued un-

til every split was used for testing. In this project,

k was set to 5. On the other hand, the depth value

used in XGBoost and RF classiﬁers were chosen as

10 through grid search. Indeed this value was in ac-

cordance with the one used in (Zia et al., 2020b). For

severity classiﬁcation, one vs. all approach was lever-

aged and the models’ performance was assessed using

accuracy, recall, precision and f1-scores. The corre-

sponding equations were listed in Equations 1, 2, 3,

4, respectively (TN: true negatives, TP: true positives,

FN: false negatives, FP: false positives).

Accuracy =

T P + T N

T P + T N + FP + FN

(1)

Recall =

T P

T P + FN

(2)

Precision =

T P

T P + FP

(3)

score = 2 ∗

precision ∗ recall

precision + recall

(4)

2.3.2 Vascular Pressure Estimation

In the regression part, the relationship between the

catheter-based pressure values and SCG-based fea-

tures was examined. Similar to the classiﬁcation task,

MLP, SVM, RF and XGBoost were used to estimate

the vascular pressures from different locations. Model

validation was again employed through 5-fold cross

validation and the depth value used in XGBoost and

RF regressors were determined using grid search. The

best option for right atrium pressure, aortic root pres-

sure, and pulmonary capillary wedge pressure was

found to be 10, whereas the best depth option for

femoral artery pressure was determined as 8. The

performance of the regressors was evaluated through

mean absolute error (MAE) and coefﬁcient of deter-

mination (R

) metrics. The corresponding equations

were listed in Equations 5 and 6, respectively (y

true

actual values, y

pred

: predicted values, y

mean

: mean of

the actual values).

MAE =

∑

i=1

pred,i

− y

true,i

(5)

= 1 −

∑

i=1

true,i

− y

pred,i

)

∑

i=1

true,i

− y

mean

)

(6)

2.4 Feature Importance Ranking

To gain a better understanding of which features are

most important for classiﬁcation, it is necessary to

calculate the relative weight of each feature in the

model. This can directly be achieved through the use

of decision trees trained by the XGBoost classiﬁer as

the features to split are determined based on the great-

est reduction in loss. Each decision tree ranks the fea-

tures based on their importance in the resulting clas-

siﬁcation algorithm. After running all the trees in the

model, the relative importance of the features can thus

be determined by averaging the importance scores ob-

tained from each tree. These scores are then used to

obtain the ﬁnal relative feature importance ranking.

As the best performing model was XGBoost, it

was used to evaluate which features among the re-

maining ones were most relevant for distinguishing

between different blood loss levels. To that aim, a

second XGBoost classiﬁer was trained, this time us-

ing data from all pigs. The resulting model was then

used to calculate the feature importance scores. Since

the objective was not to assess the model’s ability to

generalize, no separate testing set was required.

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

650

Table 1: Remaining features after variance thresholding (max: maximum, min: minimum, amp: amplitude, loc: location, rms:

root mean square).

Remaining 17 Features

First Max Amp First Max Loc First Min Amp First Min Loc Energy Skewness

Second Max Amp Second Max Loc Second Min Amp Second Min Loc RMS Kurtosis

Third Max Amp Third Max Loc Third Min Amp Third Min Loc Mean

Figure 2: (a) Receiver Operating Characteristic (ROC) Curves for different blood volume loss stages. (b) Normalized feature

importance values for the ﬁrst 15 features out of 17 (max: maximum, min: minimum, amp: amplitude, rms: root mean square).

Mapping Seismocardiogram Characteristics to Hemorrhage Status and Vascular Pressure: A Novel Approach for Triage Assessment

651

Table 2: Performance metrics for different machine learning models.

Classiﬁcation (Accuracy)

Random Forest XGBoost MLP SVM

Blood-Level Classiﬁcation 0.92 0.96 0.55 0.45

Regression (R

)

Random Forest XGBoost MLP SVM

Right Atrium Pressure 0.86 0.87 0.56 0.24

Aortic Root Pressure 0.90 0.94 0.56 0.25

Femoral Artery Pressure 0.60 0.58 0.32 0.14

Pulmonary Capillary Wedge Pressure 0.85 0.90 0.34 0.21

Table 3: XGBoost classiﬁcation and regression results.

XGBoost Classiﬁcation

Accuracy Precision Recall F1-Score

Blood-Level Classiﬁcation

0.96 0.96 0.96 0.96

XGBoost Regression

Mean Absolute Error

Right Atrium Pressure

0.87 1.29

Aortic Root Pressure

0.94 1.54

Femoral Artery Pressure

0.58 2.76

Pulmonary Capillary Wedge Pressure

0.90 0.95

3 RESULTS

3.1 Feature Selection with Variance

Thresholding

In this project, both temporal and spectral domain

analyses were leveraged as previous studies have

shown the success of each in various applications.

In total there were 154 features and this number de-

creased to 17 following variance thresholding. The

remaining features are presented in Table 1. As seen,

only the temporal ones, i.e. time domain and statis-

tical features, appeared in the remaining 17 features

and none of the spectral features were included in the

resulting set.

3.2 Hemorrhage Severity Assessment

As detailed in Section 2.3.1, four classiﬁcation mod-

els were trained using the remaining 17 features for

hemorrhage severity assessment. The models’ perfor-

mance was ﬁrst evaluated with the accuracy metric.

As reported in Table 2, RF, XGBoost, MLP and SVM

models resulted in an accuracy of 0.92, 0.96, 0.55 and

0.45, respectively. Out of the four models, XGBoost

was found to be the best performing one. For the XG-

Boost classiﬁer, each of the accuracy, recall, preci-

sion and f1-score was also calculated to be 0.96, as

reported in Table 3.

To investigate how the model behaves for differ-

ent threshold values, receiver operating characteristic

(ROC) curve for each blood level was plotted. For

7%, 14%, 21% and 28% blood loss levels, the area un-

der curve (AUC) values were calculated to be 0.991,

0.986, 0.992 and 1.000, respectively (Fig. 2(a)).

3.3 Vascular Pressure Estimation

Under the vascular pressure estimation task, regres-

sion models were trained for each of the aortic root,

femoral artery, right atrium and pulmonary capillary

wedge pressure values. For each location, four dif-

ferent models, RF, XGBoost, MLP and SVM were

leveraged similar to the classiﬁcation task. For the

right atrium, aortic root and pulmonary capillary lo-

cations, XGBoost resulted in the best accuracy val-

ues (0.87, 0.94 and 0.90, respectively). On the other

hand, for the femoral artery pressure estimation, RF

slightly outperformed XGBoost (0.60 vs. 0.58). Still,

XGBoost was selected as the ﬁnal model for all four

locations. As presented in Table 3, MAE values were

calculated to be 1.29, 1.54, 2.76 and 0.95 for right

atrium, aortic root, femoral artery and pulmonary cap-

illary models, respectively. The scatter plots repre-

senting the predicted and actual vascular pressure val-

ues with the corresponding R

results are presented in

Fig. 3.

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

652

Figure 3: Regression plots for each pressure location. Black lines correspond to the best ﬁt lines.

3.4 Feature Importance Ranking

As explained in Section 2.4, feature importance rank-

ing was applied using the XGBoost model on the re-

maining 17 temporal features. The normalized impor-

tance scores of the ﬁrst ﬁfteen out of seventeen fea-

tures are shown in Fig. 2(b). As seen, energy, ﬁrst

maximum amplitude (Max 1) and ﬁrst minimum am-

plitude (Min 1) appeared as the top three most im-

portant features. On the other hand, the statistical

features appeared in the lower end on the importance

ranking.

4 DISCUSSION

4.1 Model Training and Interpretation

Both for the hemorrhage severity assessment and

vascular pressure estimation tasks, 5-fold cross-

validation was leveraged instead of leave-one-subject-

out cross-validation (LOSO-CV). Since the pigs

reached cardiovascular collapse at different blood-

loss levels, some pigs did not have any data relating to

21% or 28% blood loss, as previously shown in Fig.

1(a). Due to this imbalance, the models would not be

able to learn properly if LOSO-CV was used; hence,

5-fold cross-validation was employed in all tasks.

Overall, for both classiﬁcation and regression

tasks, high performance metrics were obtained.

Among different vascular pressure values, the worst

performing model was the one estimating the femoral

Mapping Seismocardiogram Characteristics to Hemorrhage Status and Vascular Pressure: A Novel Approach for Triage Assessment

653

artery pressure. Considering that the accelerometer

acquiring the SCG signal was placed on the mid-

sternum of the pig, femoral artery was at a rela-

tively distal location compared to remaining three

pressure locations. Indeed, when the pressure mea-

surements are taken further from the aorta, a rela-

tively higher systolic pressure, lower end diastolic

pressure and later arrival of pulse have been observed

(Chambers et al., 2019). Hence, collecting the SCG

signals from the mid-sternum and trying to estimate

femoral artery pressure would naturally be prone to

more intra-subject variability compared to other three

locations.

4.2 Feature Interpretation

As previously explained, the temporal features out-

performed spectral features in terms of relative impor-

tance in hemorrhage assessment. Having amplitude-

related features as the most important ones was indeed

consistent with the underlying physiological events.

It has previously been shown that there is salient in-

formation regarding stroke volume in SCG amplitude

characteristics as the SCG signal represents the local

vibrations originating from ejection of blood in each

contraction (Semiz et al., 2020). Similarly, in another

study, it has been shown that the percent change in

post-hemorrhage cardiac output and percent reduc-

tion in blood volume had a linear relationship (Chien

and Billig, 1961). Based on these studies, having am-

plitude features as the most important ones in distin-

guishing between different blood-loss levels was in-

deed consistent with the literature.

4.3 Limitations and Future Work

There were several limitations in the proposed work.

First, the dataset size was relatively small (including

6 pigs). In addition, there was an imbalance in the

number of samples available for each blood-loss level

as the experimental protocol was terminated at differ-

ent blood levels for each pig. Hence, future work will

focus on testing the proposed features and pipelines

in larger datasets for assessing the generalizability of

the models.

5 CONCLUSION

The main objectives of this study were to explore

how blood loss levels relate to the SCG signals, and

to establish a correlation between the characteristics

of SCG signals and the vascular pressure values ob-

tained from various parts of the body. It was hypoth-

esized that such a system could be useful in prioritiz-

ing triage processes and monitoring critical parame-

ters. For hemorrhage severity classiﬁcation, SCG data

taken during different blood loss levels were classi-

ﬁed using one vs. all approach. Each of the precision,

recall and f1-score was calculated to be 0.96, and tem-

poral features outperformed spectral ones in terms of

added information. On the other hand, for vascular

pressure assessment, MAE values were calculated to

be 1.29, 1.54, 2.76 and 0.95 for right atrium, aortic

root, femoral artery and pulmonary capillary models,

respectively. Out of these pressure locations, femoral

artery pressure estimation resulted in the worst per-

formance with an R

value of 0.58 due to its distal lo-

cation relative to mid-sternum. Overall, this study in-

troduced new use areas for the SCG signal, which can

be utilized in the development of continuous and non-

invasive monitoring systems to prioritize the triage

processes and track vital parameters.

REFERENCES

Bishop, C. M. et al. (1995). Neural networks for pattern

recognition. Oxford university press.

Bommert, A., Sun, X., Bischl, B., Rahnenf

uhrer, J., and

Lang, M. (2020). Benchmark for ﬁlter methods

for feature selection in high-dimensional classiﬁca-

tion data. Computational Statistics & Data Analysis,

143:106839.

Breiman, L. (2001). Random forests. Machine learning,

45:5–32.

Chambers, D., Huang, C., and Matthews, G. (2019). Arte-

rial Pressure Waveforms, page 155–157. Cambridge

University Press, 2 edition.

Chen, T. and Guestrin, C. (2016). Xgboost: A scalable

tree boosting system. In Proceedings of the 22nd acm

sigkdd international conference on knowledge discov-

ery and data mining, pages 785–794.

cheol Jeong, I., Bychkov, D., and Searson, P. C. (2018).

Wearable devices for precision medicine and health

state monitoring. IEEE Transactions on Biomedical

Engineering, 66(5):1242–1258.

Chien, S. and Billig, S. (1961). Effect of hemorrhage on

cardiac output of sympathectomized dogs. American

Journal of Physiology-Legacy Content, 201(3):475–

479.

Erin, E. and Semiz, B. (2023). Spectral analysis of cardio-

genic vibrations to distinguish between valvular heart

diseases.

Friedman, J. H. (2001). Greedy function approximation: a

gradient boosting machine. Annals of statistics, pages

1189–1232.

unerbel, L., Heinrich, F., Zett, O., Axelsson, K., and

Schumann, M. (2023). Towards an intelligent triage

bracelet: A conceptual study of a semi-automated pre-

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

654

hospital triage algorithm and the integration of blood

pressure measurement.

Hersek, S., Semiz, B., Shandhi, M. M. H., Orlandic, L.,

and Inan, O. T. (2019). A globalized model for map-

ping wearable seismocardiogram signals to whole-

body ballistocardiogram signals based on deep learn-

ing. IEEE journal of biomedical and health informat-

ics, 24(5):1296–1309.

Imirzalioglu, M. and Semiz, B. (2022). Quantifying respira-

tion effects on cardiac vibrations using teager energy

operator and gradient boosted trees. In 2022 44th An-

nual International Conference of the IEEE Engineer-

ing in Medicine & Biology Society (EMBC), pages

1935–1938. IEEE.

Inan, O. T., Baran Pouyan, M., Javaid, A. Q., Dowling, S.,

Etemadi, M., Dorier, A., Heller, J. A., Bicen, A. O.,

Roy, S., De Marco, T., et al. (2018). Novel wearable

seismocardiography and machine learning algorithms

can assess clinical status of heart failure patients. Cir-

culation: Heart Failure, 11(1):e004313.

Inan, O. T., Migeotte, P.-F., Park, K.-S., Etemadi, M.,

Tavakolian, K., Casanella, R., Zanetti, J., Tank, J.,

Funtova, I., Prisk, G. K., et al. (2014). Ballistocardio-

graphy and seismocardiography: A review of recent

advances. IEEE journal of biomedical and health in-

formatics, 19(4):1414–1427.

Kimball, J. P., Zia, J. S., An, S., Rolfes, C., Hahn, J.-O.,

Sawka, M. N., and Inan, O. T. (2021). Unifying the

estimation of blood volume decompensation status in

a porcine model of relative and absolute hypovolemia

via wearable sensing. IEEE Journal of Biomedical

and Health Informatics, 25(9):3351–3360.

Noble, W. S. (2006). What is a support vector machine?

Nature biotechnology, 24(12):1565–1567.

Pandia, K., Inan, O. T., Kovacs, G. T., and Giovangrandi, L.

(2012). Extracting respiratory information from seis-

mocardiogram signals acquired on the chest using a

miniature accelerometer. Physiological measurement ,

33(10):1643.

Park, J. Y. (2021). Real-time monitoring electronic triage

tag system for improving survival rate in disaster-

induced mass casualty incidents. In Healthcare, vol-

ume 9, page 877. MDPI.

Rodriguez, D., Heuer, S., Guerra, A., Stork, W., Weber, B.,

and Eichler, M. (2014). Towards automatic sensor-

based triage for individual remote monitoring dur-

ing mass casualty incidents. In 2014 IEEE interna-

tional conference on bioinformatics and biomedicine

(BIBM), pages 544–551. IEEE.

Sakanushi, K., Hieda, T., Shiraishi, T., Ode, Y., Takeuchi,

Y., Imai, M., Higashino, T., and Tanaka, H. (2013).

Electronic triage system for continuously monitoring

casualties at disaster scenes. Journal of Ambient Intel-

ligence and Humanized Computing, 4:547–558.

Semiz, B., Carek, A. M., Johnson, J. C., Ahmad, S.,

Heller, J. A., Vicente, F. G., Caron, S., Hogue, C. W.,

Etemadi, M., and Inan, O. T. (2020). Non-invasive

wearable patch utilizing seismocardiography for peri-

operative use in surgical patients. IEEE journal of

biomedical and health informatics, 25(5):1572–1582.

Shandhi, M. M. H., Semiz, B., Hersek, S., Goller, N.,

Ayazi, F., and Inan, O. T. (2019). Performance

analysis of gyroscope and accelerometer sensors for

seismocardiography-based wearable pre-ejection pe-

riod estimation. IEEE journal of biomedical and

health informatics, 23(6):2365–2374.

Tavakolian, K., Portacio, G., Tamddondoust, N. R., Jahns,

G., Ngai, B., Dumont, G. A., and Blaber, A. P. (2012).

Myocardial contractility: A seismocardiography ap-

proach. In 2012 Annual International Conference of

the IEEE Engineering in Medicine and Biology Soci-

ety, pages 3801–3804. IEEE.

Tokmak, F. and Semiz, B. (2022). Unveiling the relation-

ships between seismocardiogram signals, physical ac-

tivity types and metabolic equivalent of task scores.

IEEE Transactions on Biomedical Engineering.

Zia, J., Kimball, J., Hahn, J.-O., Rolfes, C., and Inan, O.

(2020a). Wearable- and catheter-based cardiovascular

signals during progressive exsanguination in a porcine

model of hemorrhage.

Zia, J., Kimball, J., Rolfes, C., Hahn, J.-O., and Inan, O. T.

(2020b). Enabling the assessment of trauma-induced

hemorrhage via smart wearable systems. Science ad-

vances, 6(30):eabb1708.

Mapping Seismocardiogram Characteristics to Hemorrhage Status and Vascular Pressure: A Novel Approach for Triage Assessment

655