A Hierarchical Framework for Apnea Detection and Respiration Pace

Assessment Using Seismocardiogram Signals

Berke Kizir and Beren Semiz

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

Keywords:

Seismocardiogram, Respiration Rate, Apnea, Health Monitoring.

Abstract:

Sleep constitutes one-third of human life and plays a critical role in physical repair, mental functioning, and

memory consolidation. Although polysomnography (PSG) has been used to assess sleep performance; this

test requires participants to visit a sleep clinic and have multiple sensors attached to their bodies. Hence,

there is a need for alternative methods which can provide sleep monitoring outside clinical settings, but with

clinical standards. In this work, a novel hierarchical framework was built to leverage the seismocardiogram

(SCG) signals in apnea detection and respiration pace assessment using a simulated data collection protocol.

In the ﬁrst step of the framework, a binary Light Gradient-Boosting Machine (LGBM) model was trained to

detect the breath-holding (apnea) episodes. If the prediction was not a breath-holding state, the data was fed

into a multi-class LGBM model to distinguish between normal, slow and fast breathing episodes. Overall, the

binary LGBM resulted in an accuracy, recall, precision and f1-score of 0.99, 0.95, 0.87 and 0.91, respectively;

whereas for the multi-class case all metrics were 0.96. Additionally, the optimum window length to achieve

real-time detection was determined as 5 seconds. The results show that the SCG signals hold substantial

information regarding the changes in breathing patterns, thus could potentially be leveraged in the design of

wearable systems as an alternative to the PSG test.

1 INTRODUCTION

Sleep constitutes one-third of human life and plays

a critical role in physical repair, mental functioning,

and memory consolidation (Kwon et al., 2021). Re-

cent surveys have shown that the 44% of adults have

experienced a decline in the quality of their sleep

over the past ﬁve years, and eight out of every ten

adults expressed a desire to improve their sleep qual-

ity (Philips, 2019). The most commonly observed

sleep problems include insomnia, sleep apnea, and

narcolepsy. Moreover, deteriorations in sleep efﬁ-

ciency due to these problems are associated with sec-

ondary conditions such as depression, obesity, dia-

betes, heart diseases, and neurocognitive disorders

(Altevogt et al., 2006).

Traditionally, polysomnography (PSG) has been

used to assess sleep performance; however, this test

requires participants to visit a sleep clinic and have

multiple sensors attached to their bodies. Although

this system provides a thorough assessment of sleep

quality and related problems, assessing sleep in natu-

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

ral settings (at home and without multiple sensor at-

tachment) could potentially provide a more realistic

evaluation. On the other hand, there have emerged

watch-like systems (Apple Watch, etc.) providing

sleep monitoring and staging, however these systems

only use customized algorithms based on movement

and heart rate, and they cannot provide detailed infor-

mation regarding other vital parameters, such as res-

piration rate. Hence, there is a need for alternative

methods which can provide sleep monitoring outside

clinical settings, but with clinical standards.

Recent studies have shown that the seismocardio-

gram (SCG) signal collected from the thoracic re-

gion can provide information about various hemody-

namic parameters (Inan et al., 2014). SCG signal

corresponds to the chest micro-vibrations occurring

due to the ejection of blood and contraction of the

heart in each cardiac cycle. The part of the SCG be-

low 1 Hz corresponds to the chest movements associ-

ated with respiration, the frequencies between 1 - 20

Hz includes the cardiac vibrations, and the compo-

nent above 20 Hz represents the heart sounds (Pandia

et al., 2012). As SCG signals can assess the thoracic

region from different perspectives, they are widely

Kizir, B. and Semiz, B.

A Hierarchical Framework for Apnea Detection and Respiration Pace Assessment Using Seismocardiogram Signals.

DOI: 10.5220/0012446400003657

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 793-798

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

793

used in wearable systems aiming to monitor cardio-

vascular and cardiopulmonary diseases (Hayirlioglu

and Semiz, 2023). Indeed, previous studies have

shown that the SCG signals could potentially be used

in the detection of heart failure (Inan et al., 2018),

aortic stenosis (Yang et al., 2019), and atrial ﬁbril-

lation (Hurnanen et al., 2016), as well as predicting

stroke volume values (Semiz et al., 2020), estimating

systolic time intervals (Shandhi et al., 2019), classify-

ing valvular heart disease locations (Erin and Semiz,

2023) and assessing respiration phases (Imirzalioglu

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

In this study, we developed a novel hierarchi-

cal framework to leverage the SCG signals in ap-

nea detection and respiration pace assessment using

a simulated data collection protocol including breath-

holding, slow-breathing, normal-breathing and fast-

breathing episodes. Instead of focusing on the com-

ponent below 1 Hz, we leveraged the vibration and

acoustic components of the signal (1-40 Hz) as all ap-

nea types are not necessarily associated with halted

chest movements. In the ﬁrst step of the framework,

as the most crucial aspect of the study was to de-

termine the breath-holding states, a binary classiﬁca-

tion model was trained to detect the breath-holding

(apnea) episodes. If the prediction was not breath-

holding state, the data was fed into a second model,

which was designed as a multi-class classiﬁcation

model to distinguish between normal, slow and fast

breathing episodes. In addition, the optimum win-

dow length for achieving real-time apnea detection

and breathing rate assessment was studied. Overall,

the results show that the SCG signals hold substantial

information regarding the changes in breathing pat-

terns, thus could potentially be leveraged in the de-

sign of wearable systems as an alternative to the PSG

test.

2 METHODS

2.1 Data Collection Protocol

This study was conducted under a protocol approved

by the Koc University Institutional Review Board and

all subjects provided written consent. A total of 8

healthy subjects (6 females and 2 males) participated

in the study (Age: 21.9 ± 3.0, height: 171.6 ±

7.7, and weight: 65.4 ± 12.2). The signals were

simultaneously collected using the BIOPAC system

(BIOPAC Systems, Inc. Goleta, CA, USA) at 2 kHz.

The electrocardiogram (ECG) and respiration signals

were acquired using three gel electrodes and a respira-

tion transducer, respectively. The signals were trans-

Figure 1: The locations of the sensors.

ferred to the BIOPAC system using wireless Biono-

madix RSPEC-R module. To record the SCG signal, a

tri-axial low noise analog accelerometer (ADXL354,

Analog Devices, Inc., Norwood, MA) was used. It

was placed on the mid-sternum of the subject using

hypoallergenic transparent medical tape. The sensor

locations are detailed in Fig. 1. The X, Y, Z axes

of the accelerometer were corresponding to the vibra-

tions in lateral, head-to-foot and dorso-ventral direc-

tions, respectively.

Data collection protocol is detailed in Fig. 2.

The subject was ﬁrst asked to breathe normally for

1 minute, followed by Valsalva maneuver where the

subject held breath for 15 seconds. It was followed by

another 1 minute-long normal breathing. The subject

was then asked to perform slow breathing at 40 beats-

per-minute (bpm) controlled by an online metronome.

After 1 minute-long normal breathing, the subject

performed fast breathing at 120 bpm, again controlled

by a metronome. The last steps were including an ad-

ditional 1 minute-long normal breathing, 15-seconds-

long Valsalva maneuver and another normal breathing

phase for 1 minute. The signals were recorded con-

tinuously during the protocol and the timestamps for

each transition were recorded thoroughly.

2.2 Preprocessing

SCG signals were ﬁltered with a Kaiser window ﬁnite

impulse response ﬁlter to remove out-of-band noise.

The cut-off frequencies were selected as 1-40 Hz for

all three axes. No other preprocessing steps were ap-

plied not to lose any information (spikes, oscillations,

etc.), which might be useful in determining different

breathing states.

In this work, one of the fundamental aims was

to assess the effect of window length in detection

performance. To that end, the analysis pipeline was

repeated using different window lengths (1, 2, 3, 4

and 5 seconds). Between the consecutive windows,

500 milliseconds-long (0.5 seconds) overlap was em-

ployed. By this way, the number of instances to be

used during training was increased, additionally the

subject’s state could be updated every 0.5 seconds.

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

794

Figure 2: Experimental protocol.

Figure 3: Study pipeline.

It should be noted that the ECG was not used as the

reference since the real-time apnea detection scenario

will necessitate the use of continuous data streaming,

which can be achieved through a sliding window.

2.3 Feature Extraction

Statistical, temporal and spectral features were ex-

tracted from each SCG window in all three axes (Ta-

ble 1). As the statistical features, mean, variance,

skewness (asymmetry) and kurtosis (tailedness) were

calculated. As the temporal feature, signal energy, i.e.

the squared sum of the samples, were computed. On

the other hand, spectral domain features were includ-

ing centroid, spread, rolloff and bandpowers. Rolloff

indicates the frequency at which a speciﬁc percent-

age of the signal energy is accumulated. On the other

hand, centroid and spread relate to the center of mass

and the distribution of frequencies in the spectrum, re-

spectively (Giannakopoulos and Pikrakis, 2014). Fi-

nally, bandpower frequency intervals were selected

as logarithmically spaced frequency bands between 1

and 40 Hz. After extracting all features, a dataframe

was generated where columns including the feature

values and rows indicating the corresponding SCG

frames. The whole feature extraction and data frame

generation steps were repeated for different window

lengths.

Table 1: Feature groups.

Statistical

Features

Temporal

Features

Spectral

Features

Mean

Signal Energy

Centroid

Variance Spread

Skewness Rolloff

Kurtosis

Bandpowers

(logarithmically spaced

between 1 - 40 Hz)

2.4 Model Training and Feature

Importance Analysis

2.4.1 Hierarchical Model Framework

In this study a hierarchical model framework was built

(Fig. 3). Since the most crucial aspect of the study

was to determine the breath-holding states, a binary

classiﬁcation model was trained ﬁrst. In this model,

breath-holding states were labeled as 1, while the oth-

ers (normal, slow, fast) were all labeled as 0. Test data

was ﬁrst fed into the binary classiﬁer. If the prediction

was not breath-holding state, the data was fed into a

second model, which was designed as a multi-class

classiﬁcation model. In this secondary step, the aim

was to distinguish between the normal, slow and fast

breathing episodes. This hierarchical pipeline was

built so that the breath-holding states (i.e. apnea peri-

ods) could be determined as fast as possible regardless

of the changes in breathing pace in-between.

2.4.2 Model Selection and Validation

As the superior performance of the tree-based meth-

ods are well known in SCG-related applications,

four different tree-based models were trained and

compared: decision tree (DT), random forest (RF),

extreme gradient boosting trees (XGB) and light

gradient-boosting machine (LGBM). For all classiﬁ-

cation models, 5-fold cross validation was applied.

• Decision Tree (DT): Decision tree is a tree-like

ﬂowchart which utilizes a divide and conquer

strategy, employing a greedy search to ﬁnd the

best split points. This splitting process is per-

formed in a top-down fashion until the data-of-

interest has been assigned class labels (Song and

Ying, 2015).

A Hierarchical Framework for Apnea Detection and Respiration Pace Assessment Using Seismocardiogram Signals

795

• Random Forest (RF): Instead of depending on a

single tree, RF involves bootstrapping multiple

trees by utilizing randomized subsets drawn from

the dataset. These trees are trained independently

and in parallel. Majority voting is then applied on

the outputs obtained from these trees to yield one

single class estimation (Breiman, 2001).

• Extreme Gradient Boosting Trees (XGB): XGB is

one of the popular boosting algorithms. As RF re-

lies of bagging, XGB operates sequentially, i.e.,

each subsequent tree relies on the the outcome of

the previous one. Overall, in the training process,

multiple decision trees are trained iteratively, al-

lowing the prediction and adjustment of residual

errors from the previous iteration as the training

advances (Chen and Guestrin, 2016).

• Light Gradient-Boosting Machine (LGBM):

LGBM is another popular type of boosting algo-

rithms, however unlike the horizontal, level-wise

growth seen in XGB, LGBM follows a vertical,

leaf-wise growth pipeline. This approach leads

to increased loss reduction, resulting in higher

accuracy and faster processing (Ke et al., 2017).

To assess the performance of the models, accuracy

and weighted precision, recall and f1-score were used.

These equations are presented in Equations 1, 2, 3 and

4, respectively (TP: true positives, FP: false positives,

TN: true negatives and FN: false negatives). In addi-

tion, the area under the receiver operating character-

istics curve (ROC AUC) was computed for the binary

classiﬁcation task.

Accuracy =

T P + T N

T P + T N + FP + FN

(1)

Precision =

T P

T P + FP

(2)

Recall =

T P

T P + FN

(3)

score = 2 ∗

precision ∗ recall

precision + recall

(4)

2.4.3 Feature Importance Ranking

Feature importance scores were computed from the

best performing model (LGBM) to ﬁnd out the most

important features. The procedure was repeated for

both the binary and multi-class tasks. After all folds

were completed, average of the normalized LGBM

scores was calculated and determined as the ﬁnal im-

portance score. The corresponding scores were then

ranked in descending order to determine the feature

importance ranking.

Table 2: Performance comparison (binary).

Model Accuracy AUC Recall Precision F1

LGBM 0,99 0,99 0,95 0,87 0,91

RF 0,94 0,95 0,95 0,30 0,46

XGB 0,99 0,99 0,95 0,85 0,90

DT 0,92 0,73 0,56 0,49 0,52

Table 3: Performance comparison (multi-class).

Model Accuracy Recall Precision F1

LGBM 0,96 0,96 0,96 0,96

RF 0,88 0,88 0,94 0,90

XGB 0,95 0,95 0,96 0,96

DT 0,85 0,85 0,85 0,85

3 RESULTS AND DISCUSSION

3.1 Apnea Detection Results

The ﬁrst step of the hierarchical classiﬁcation frame-

work was to distinguish between the breath-holding

episodes and breathing (normal, fast, slow) periods.

To that end, a binary classiﬁcation model was trained.

The results obtained from different models with a

5-sec window were presented in Table 2. Overall,

LGBM and XGB outperformed DT and RF in all met-

rics. Indeed, the precision values obtained from the

RF and DT were signiﬁcantly lower than the other

models, which led to again low f1-scores. On the

other hand, XGB and LGBM indeed had compara-

ble performance, however LGBM performed slightly

better in terms of precision and f1-score. Having an

accuracy value of 0.99 revealed that there is almost no

carried-error from the binary model to the multi-class

task.

3.2 Breathing Rate Assessment

Using the primary model, the SCG windows were

labeled as breath-holding or not. If the prediction

was not breath-holding state, the data was fed into a

second model, which was designed as a multi-class

classiﬁcation model. In this secondary step, the aim

was to distinguish between the normal, slow and fast

breathing episodes. The performances of different

models are presented in Table 3. Similar to the bi-

nary case, LGBM and XGB outperformed RF and DT

in terms of all metrics, however the precision and f1-

score obtained from RF and DT in multi-class task

were signiﬁcantly higher than the ones obtained in

the binary task. Overall, the LGBM model performed

slightly better than the XGB, similar to the binary

case.

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

796

Table 4: LGBM results for different window lengths (binary).

Window Length (sec.) Accuracy AUC Recall Precision F1-Score

1 0,95 0,91 0,72 0,51 0,59

2 0,97 0,96 0,86 0,71 0,78

3 0,98 0,97 0,91 0,77 0,83

4 0,98 0,99 0,93 0,83 0,87

5 0,99 0,99 0,95 0,87 0,91

Table 5: LGBM results for different window lengths (multi-class).

Window Length (sec.) Accuracy Recall Precision F1-Score

1 0,85 0,85 0,88 0,86

2 0,90 0,90 0,92 0,91

3 0,93 0,93 0,94 0,93

4 0,95 0,95 0,96 0,95

5 0,96 0,96 0,96 0,96

3.3 Effect of Window Length

As previously discussed, the ECG was not used as

the reference since the real-time apnea detection sce-

nario necessitates the use of continuous data stream-

ing, which can be achieved through a sliding window.

However, determining the size of the optimum analy-

sis window is an important research question. Hence,

different window lengths (1, 2, 3, 4, 5 seconds) were

tested to ﬁnd the optimum length for the current ap-

plication. Throughout the experiments, the model was

set to LGBM.

For the binary and multi-class tasks, the perfor-

mance values for varying window lengths are pre-

sented in Tables 4 and 5. When window length was

selected as 1 second, the results got worse signiﬁ-

cantly for both tasks. For the binary case, the re-

call value was calculated as 0.72 when 1-sec windows

were used, i.e., there was a %24 decrease compared

to the case where 5-sec windows were used. Win-

dow lengths longer than 5-sec resulted in similar re-

sults, and when the length exceeded 10-sec, the per-

formance started to decrease. A similar pattern was

also valid for the multi-class task. Based on these

observations, the optimum window length was deter-

mined as 5-sec.

3.4 Feature Importance Ranking

Feature importance scores were computed from the

LGBM classiﬁer. After all folds were completed, av-

erage of normalized scores were calculated as the ﬁ-

nal score. The feature importance results for both

task are presented in Figure 4. Overall, bandpowers,

kurtosis and skewness were dominating in both bi-

nary and multi-class classiﬁcation tasks. Considering

that the kurtosis represents the tailedness and skew-

Figure 4: Feature importance ranking for the binary and

multi-class tasks.

ness represents asymmetry in the data, they are impor-

tant indicators of how the data is distributed. Having

kurtosis and skewness of multiple axes as the most

important features thus reveals that the distributions

of different breathing episodes were indeed different

than each other.

A Hierarchical Framework for Apnea Detection and Respiration Pace Assessment Using Seismocardiogram Signals

797

4 CONCLUSION

In this work, a novel hierarchical framework was built

using a simulated data collection protocol for evaluat-

ing the potential use of SCG signals in apnea detec-

tion and respiration pace assessment. In the ﬁrst step

of the framework, a binary Light Gradient-Boosting

Machine (LGBM) model was trained to detect the

breath-holding (apnea) episodes. If the prediction was

not a breath-holding state, the data was fed into a

multi-class LGBM model to distinguish between nor-

mal, slow and fast breathing episodes.

Overall, the binary LGBM model resulted in an

accuracy, recall, precision and f1-score of 0.99, 0.95,

0.87 and 0.91, respectively; whereas for the multi-

class case all metrics were 0.96. Additionally, differ-

ent window lengths (1, 2, 3, 4, 5 seconds) were tested

and the optimum window length was determined as 5

seconds.

The results show that the SCG signals hold sub-

stantial information regarding the changes in breath-

ing patterns, thus could potentially be leveraged in the

design of wearable systems as an alternative to the

PSG test. Future work will focus on validating these

results in larger datasets including real data from pa-

tients having sleep apnea.

REFERENCES

Altevogt, B. M., Colten, H. R., et al. (2006). Sleep dis-

orders and sleep deprivation: an unmet public health

problem.

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

45:5–32.

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.

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

genic vibrations to distinguish between valvular heart

diseases.

Giannakopoulos, T. and Pikrakis, A. (2014). Introduction

to audio analysis: a MATLAB® approach. Academic

Press.

Hayirlioglu, Y. Z. and Semiz, B. (2023). A novel multi-

modal sensing system prototype for cardiovascular

and cardiopulmonary monitoring.

Hurnanen, T., Lehtonen, E., Tadi, M. J., Kuusela, T.,

Kiviniemi, T., Saraste, A., Vasankari, T., Airaksinen,

J., Koivisto, T., and P

ank

a, M. (2016). Auto-

mated detection of atrial ﬁbrillation based on time–

frequency analysis of seismocardiograms. IEEE jour-

nal of biomedical and health informatics, 21(5):1233–

1241.

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.

Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W.,

Ye, Q., and Liu, T.-Y. (2017). Lightgbm: A highly

efﬁcient gradient boosting decision tree. Advances in

neural information processing systems, 30.

Kwon, S., Kim, H., and Yeo, W.-H. (2021). Recent ad-

vances in wearable sensors and portable electronics

for sleep monitoring. Iscience, 24(5).

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.

Philips (2019). The global pursuit of better sleep health.

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.

Song, Y.-Y. and Ying, L. (2015). Decision tree methods: ap-

plications for classiﬁcation and prediction. Shanghai

archives of psychiatry, 27(2):130.

Yang, C., Aranoff, N. D., Green, P., and Tavassolian, N.

(2019). Classiﬁcation of aortic stenosis using time–

frequency features from chest cardio-mechanical sig-

nals. IEEE Transactions on Biomedical Engineering,

67(6):1672–1683.

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

798