Towards Wearable Estimation of Tidal Volume via Electrocardiogram and Seismocardiogram Signals

Abstract

The current COVID-19 pandemic highlights the critical importance of ubiquitous respiratory health monitoring. The two fundamental elements of monitoring respiration are respiration rate (the frequency of breathing) and tidal volume (TV, the volume of air breathed by the lungs in each breath). Wearable sensing systems have been demonstrated to provide accurate measurement of respiration rate, but TV remains challenging to measure accurately with wearable and unobtrusive technology. In this work, we leveraged electrocardiogram (ECG) and seismocardiogram (SCG) measurements obtained with a custom wearable sensing patch to derive an estimate of TV from healthy human participants. Specifically, we fused both ECG-derived and SCG-derived respiratory signals (EDR and SDR) and trained a machine learning model with gas rebreathing as the ground truth to estimate TV. The respiration cycle modulates ECG and SCG signals in multiple different ways that are synergistic. Thus, here we extract EDRs and SDRs using a multitude of different demodulation techniques. The extracted features are used to train a subject independent machine learning model to accurately estimate TV. By fusing the extracted EDRs and SDRs, we were able to estimate the TV with a root-mean-square error (RMSE) of 181.45 mL and Pearson correlation coefficient (r) of 0.61, with a global subject-independent model. We further show that SDRs are better TV estimators than EDRs. Among SDRs, amplitude modulated (AM) SCG features are the most correlated to TV. We demonstrated that fusing EDRs and SDRs can result in moderately accurate estimation of TV using a subject-independent model. Additionally, we highlight the most informative features for estimating TV. This work presents a significant step towards achieving continuous, calibration free, and unobtrusive TV estimation, which could advance the state of the art in wearable respiratory monitoring.

I. Introduction

Monitoring the respiratory system is important for providing valuable health markers such as tidal volume (TV), the volume of inhaled and exhaled air by the lungs during a single breath, and respiratory rate (RR), the number of breathing cycles per minute. These key respiratory health indicators, among others, deteriorate in a diseased state, such as respiratory failure, which is one of the major causes of admission to intensive care units (ICU) [1–3]. Clinical deterioration due to respiratory illnesses can either be acute or chronic, which motivates continuous monitoring of respiratory markers [4]. With the global spread of the coronavirus disease (COVID-19), continuous and convenient monitoring of respiratory health has become critical [5–8].

Unfortunately, current techniques to measure respiratory parameters are performed either by manual observation or using hospital-based devices. The manual observation method is conducted by healthcare providers [4], [9] and may lack precision compared to quantified assessments based on continuously measured parameters. Hospital-based methods require breathing through some measuring apparatus such as a spirometer [10], nasal thermocouples, pneumatic respiration transducers, fiber-optic sensors [11], or Doppler radar [12]. However, these techniques are obtrusive, expensive, and / or not practical for continuous monitoring outside hospital settings. An ideal respiratory monitoring system would provide continuous information on RR and TV in an unobtrusive, real-time, and portable form factor with acceptable accuracy and cost [4].

Recent advances in sensor technologies enable continuous and portable daily monitoring of respiratory health through indirect estimation of respiratory parameters from bio signals, such as the electrocardiogram (ECG) [13–18], photoplethysmogram (PPG) [19–21], seismocardiogram (SCG) [22–24] or a combination of them [25–28]. The aforementioned biosignals are easily and widely acquired by noninvasive sensors in both and consumer devices, making them suitable candidates for monitoring respiratory health in a range of settings. It is widely reported that the ECG and SCG both exhibit three respiratory modulations as illustrated in Figure 1, namely: amplitude modulation (AM), frequency modulation (FM), and baseline wander (BW) [26]. AM and BW of the ECG are caused by changes in thoracic impedance and changes in the orientation of the heart’s electrical axis relative to the electrodes [29]. For SCG, BW is associated with the movement of the chest wall during breathing and AM with the reduction in stroke volume during breathing [30]. FM is the result of respiratory sinus arrhythmia (RSA), which is the increase in heart rate during inhalation and decrease during exhalation [31].

Fig. 1.

An illustration of idealized modulations for a) ECG: AM (top), FM (center), and BW (bottom), and b) SCG: AM (top), FM (center), and BW (bottom).

Indirect respiratory health monitoring has been investigated widely in literature. A multitude of these studies use ECG or PPG or both [26–28] to estimate respiratory parameters. Other studies use accelerometer signals to assess respiration as in [22]. The authors in [32] explored the estimation of TV using 3 and 12 lead ECG signals. They compared the correlation between different ECG extracted features with the TV for different activity levels. Although their measurement device is not portable, their best reported subject independent linear model during recovery was able to estimate the TV with median error of 490 mL and an interquartile range (IQR) of 360 mL.

Fekar et al. [33] introduced the use of a single accelerometer to estimate RR and TV variability. Their sensor is attached to the sternum, which makes the device portable and suitable for continuous monitoring in different settings. The study included 8 subjects and used subject specific calibrated models that can estimate the TV variability using extracted peaks and valleys of the accelerometer signal which is equivalent to the BW peaks and valleys of the SCG signal shown in Figure 1. Their reported Pearson correlation coefficient (r) between estimated TV and gold standard spirometer signal is 0.87 while their subject specific average mean squared error (MSE) derived from normalized TV estimate and normalized respirometer signals for the four different protocols in their study is 8.48 × 10⁻⁴.

In this work, we introduce a novel approach to estimate TV using a combination of single lead ECG and tri-axial SCG features. These physiological signals are captured using our custom patch introduced in [34–37]. This hardware is portable, relatively unobtrusive, and suitable for continuous monitoring. Our proposed approach is subject-independent which allows for its usage without any further calibration. We also compare the feature importance scores of different ECG and SCG extracted respiratory surrogates and highlight the most salient features to capture TV information.

II. Methods

A. Wearable Patch Dataset

In this work, we use a dataset collected in a previously published study [34], [35], which operated under a protocol (H18452) approved by the Georgia Institute of Technology Institutional Review Board. The aforementioned dataset was de-identified and is only used here for post hoc analysis. Twenty-one healthy subjects were recruited in this study from which eighteen were selected for our analysis. Two of the excluded subjects have corrupted ECG data while a third one has outlier TV values. The participants’ average age was 26.8±4.1 years, and the weight and height were 67.5±14.1 kg and 170.5±9.9 cm respectively.

Different physiological signals were recorded for participants using our wearable patch, including single lead ECG sampled at 1 kHz, tri-axial SCG sampled at 1 kHz, and skin temperature and atmospheric pressure sampled at 65 Hz. The wearable patch was placed at the mid sternum. Simultaneously, gold standard TV and other metabolic variables were recorded using the COSMED K5 wearable metabolic system (COSMED, Rome, Italy), at a rate of one sample per breath.

The study protocol started with a baseline measurement for the subjects followed by indoor treadmill walking and running at different speeds; then, each subject performed an outdoor walking task. Finally, the protocol ended with a recovery stage during which each subject stood still for 5 minutes. The data collected from the patch is synchronized with the COSMED K5 TV recordings up to a millisecond precision. We focus our work presented here on the recovery period for two reasons: 1) Subjects are at complete rest during recovery which minimizes motion artifacts in the recorded signals. 2) Recovery involves the widest dynamic range of TV (180 mL – 1600 mL) for different subjects.

B. Preprocessing Pipeline

We started by demodulating physiological signals to obtain multiple respiratory surrogates. The demodulation was performed in three different ways: amplitude, frequency, and baseline wander demodulation. Once we obtained the surrogates, peak-to-peak amplitudes of these signals were extracted and used as features to estimate the TV using machine learning regression model. We used random forests regressor to fuse multiple features from different physiological signals and demodulation schemes. The detailed pipeline schematic is shown in Figure 2.

Fig. 2.

Schematic diagram of the demodulation block. ECG and SCG signals are first filtered to eliminate out of band noise. Then ECG and SCG peaks and valleys (PK & VLY) are detected, then the peak-to-peak (PK-PK) and valley-to-valley (VLY-VLY) intervals are quantified. To extract BW, the signals are filtered using a band pass filter at the respiration frequency range which is the branch shown at the bottom. At the following stage, respiratory surrogates are examined for outliers in the outlier removal stage. Outliers are deemed to be values that are ±3 standard deviations away from the mean. Then a final peak and valley detection is applied to extract final TV features (F₁,.., F_M).

Using finite impulse response Kaiser window band-pass filters, the raw ECG and triaxial SCG signals were filtered at cut-off frequencies 10–40 Hz and 3–40 Hz respectively. The cut-off frequencies for ECG were chosen to suppress the T-wave and emphasize the R-peaks to improve the accuracy of R-peak detection at the subsequent stage [38], [39]. For triaxial SCG, the cut-off frequencies were chosen to eliminate out of band noise without distorting the features of interest [40]. Since our subjects were standing still during the study, we refer to head to foot, lateral and dorsoventral SCG directions as SCG_y, SCG_x and SCG_z respectively throughout the rest of the paper.

Using these filtered SCG components, we computed two additional SCG components that we assumed to contain useful information about the chest movement during respiration [41]. These components are the magnitude of the SCG signal along the head to foot and dorsoventral planes referred to as SCG_yz and the magnitude of the SCG signal in the three-dimensional coordinates system referred to as SCG_m. These two derived SCG components are computed using the following equations.

$${\text{SCG}}_{yz}=\sqrt{{\text{SCG}}_{\text{y}}^2+{\text{SCG}}_{\text{z}}^2}$$ (1)

$${\text{SCG}}_{\text{m}}=\sqrt{{\text{SCG}}_{\text{x}}^2+{\text{SCG}}_{\text{z}}^2}$$ (2)

After the filtering stage, we used the Pan Tompkins algorithm [42], [43] to detect the R-peaks of the ECG signal. The detected peaks were used to segment the wearable signals, ECG and five waveforms of SCG, into individual heartbeats. Once signals are segmented, we computed the signal quality index (SQI) [44] to quantify the quality of individual beats of the SCG signal using a previously validated algorithm. The SQI algorithm matches every SCG segmented beat to a template. The matching is made using dynamic time feature matching, which is an improved version of dynamic time warping introduced in [45]. This matching results in a number that represents the distance between the beat and the chosen template. A similarity score between the beat and the template is chosen to be the inverse of that distance. For every beat, the chosen template is a 30-beat average around that beat. The generated similarity score is in the range [0,1] with 0 and 1 representing the lowest and highest similarity scores respectively. The algorithm generated a similarity score for every beat for every SCG axis. For a specific beat, the SQI score was chosen to be the minimum score among different SCG axes for that beat. We did not use a universal hard threshold for the SQI cutoff score for two reasons: 1) The SQI score for every beat is a relative score indicating the quality of that beat compared to its neighborhood. Thus, we need a relative cutoff to discard beats based on neighborhood statistics to be compatible with our SQI algorithm which is a neighborhood quality indicator. 2) The quality of beats changes from one subject to another and from one protocol stage to another. Thus, using a universal cutoff threshold would result in highly unbalanced dataset, where subjects with relatively high-quality data dominating the dataset. This could negatively impact the generalizability of our algorithm. Based on manual inspection, we found that discarding beats that fall in the bottom quartile of similarity score for each subject eliminated most of the low-quality beats during the recovery period and reduced our dataset biases towards subjects with relatively high-quality data.

C. Tidal Volume Feature Extraction

Using an automated algorithm, we extracted 30 respiratory signal surrogates. Figure 3 a) shows ECG and SCG_z extracted fiducial points that are used to create ECG and SCG_z surrogates. For ECG, we extracted R-wave peaks as described above and the valleys are extracted by applying the same algorithm to the voltage inverted version of the ECG signal. For SCG_z, the peaks and valleys were extracted using algorithm in [46] and likewise for SCG_x and SCG_y. For SCG_yz and SCG_m we only extracted the peaks since these signals have only non-negative values.

Fig. 3.

a) Description of the extracted respiration features from ECG (top) and SCG (bottom). Amplitude features are numbered 1, 2 for ECG and 5, 6 for SCG. While frequency features are numbered 3, 4 for ECG and 7,8 for SCG. b) The corresponding TV is extracted from the features in a) by detecting peaks and valleys.

The consecutive peaks, valleys, and peak minus valley values capture the respiration amplitude modulation while the consecutive peak-to-peak and valley-to-valley intervals capture the frequency modulation. We extract three amplitudes modulated and two frequency modulated respiration surrogates from each of ECG, SCG_x, SCG_y and SCG_z. For SCG_yz and SCG_m we extract one amplitude modulated and one frequency modulated surrogate for each. For the rest of the work the nomenclature for the respiratory surrogate signals is given by the signal they are derived from followed by their demodulation technique and extrema. For example, SCG_am,min−max is the respiration surrogate obtained by detecting the values of peaks minus valleys of consecutive heart beats. The obtained surrogates are then checked for outliers that fall outside 3 standard deviations away from the mean on a 30 second window basis. The surrogate signals are divided into 30 second windows. For each window, we computed the upper envelope, lower envelope and envelop difference. We then calculated the mean and standard deviation of the envelope difference. We excluded surrogate points that are 3 standard deviations away from the mean [47]. After extracting the respiratory surrogates, we extracted their peaks to peak amplitudes. We hypothesize that the peak-to-peak amplitudes of these surrogates are correlated to the TV of the participant [41]. The full list of features and their names are shown in Table 1.

TABLE I.

A list of all extracted features acronyms is presented on the left column and their description is shown on the right column.

Feature Name	Description
ECGbw	Baseline wander surrogate from ECG
ECGam,max	Amplitude demodulated surrogate from ECG peaks
ECGam,min	Amplitude demodulated surrogates from ECG valleys
ECGam,max-min	Amplitude demodulated surrogates from ECG peak to peak amplitudes
ECGfm,max	Frequency demodulated surrogates from ECG peaks
ECGfm,min	Frequency demodulated surrogates from ECG valleys
	Baseline wander surrogate from ECG
SCWbw	Amplitude demodulated surrogate from SCG peaks
SCGam,max	Amplitude demodulated surrogates from SCG valleys
SCGam,min	Baseline wander surrogate from SCG
SCGam,max-min	Amplitude demodulated surrogates from SCG peak to peak amplitudes
SCGfm,max	Frequency demodulated surrogates from SCG peaks
SCGfm,min	Frequency demodulated surrogates from SCG valleys

D. Tidal Volume Estimation (ML)

The respiratory peak-to-peak amplitudes from different physiological signals are used as input features to the machine learning model. We used random forests [48] as our regression model to estimate TV due to its relatively lower generalization error compared to other classical machine learning regression models such as linear regression. We perform leave-one-subject-out (LOSO) [49] subject scheme during validation and testing as well. Before starting the training, we hold one subject as a test subject. Then we perform LOSO cross validation (CV) on the remaining subjects. The maximum tree depth is the only model hyperparameter that is tuned during training and validation. The model with the lowest RMSE during validation is picked for testing. We repeat the process of holding one test subject and performing LOSO CV for the rest of the subjects. The random forest regressor was responsible of fusing different features from different signals to robustly estimate the TV.

We permute over different combinations of features to train and test our model. The root-mean-square error (RMSE) and Pearson’s correlation coefficient (r) evaluated on estimated and true TV values are our evaluation metrics which are calculated using the following formulas

$$\text{RMSE}=\sqrt{\frac{1}{N}\sum _{i=1}^N{\left(\widehat{\text{TV}}1-TVi\right)}^2}$$ (3)

$$r=\frac{cov\left(\widehat{TV},TV\right)}{{\sigma }_{\widehat{\text{TV}}}{\sigma }_{\text{TV}}}$$ (4)

respectively. Where cov(x,y) is the covariance of random variables x and y, and σ_x is the standard deviation of random variable x. We combine all the features in Table 1 into different sets based on the physiological signal they are extracted from, and the demodulation technique used to extract them.

III. Results

We report the testing results for different permutations of feature sets in Table 2. The ECG extracted features are used as our baseline since they are the commonly investigated feature in literature. We then add SCG features to ECG ones and report different combinations. In Figure 5, we show more insights about the relative importance of different feature sets as well as individual features. Figure 5 a) presents the relative feature importance for different features that is computed as the mean decrease in impurity [50].

TABLE II.

Different models are trained and tested on feature groups on the left most column and their corresponding testing rmse and r are reported on the two right columns respectively.

Features	RMSE (mL)	r
ECG	235.6	0.26
SCG	203.6	0.49
ECG+SCGbw	214.9	0.45
ECG+SCG am	181.5	0.61
ECG+SCGfm	236.3	0.28
ECG+SCGbw+SCGam	184.5	0.60
ECG+SCGbw+SCGfm	215.4	0.46
ECG+SCGam+SCGfm	182.8	0.60
ECG+SCGbw+SCGam+SCGfm	188.1	0.59

Fig. 5.

Extracted features are ranked in descending order based on their importance. The black lines represent ± 1 standard deviation of the importance value. a) The importance values of all extracted features are individually presented b) The aggregated importance per demodulation technique is shown. It is important to note that SCG_am demodulated features are the most salient among other extracted features.

The ECG features alone achieve a RMSE of 235.6 mL and r of 0.26 while SCG features achieve 203.6 mL and 0.49, respectively. When adding different SCG demodulation schemes to ECG features, we find improved performance for SCG_am, with RMSE of 181.45 mL and r of 0.61, and for SCG_bw the RMSE was 214.9 mL and r was 0.45. In contrast, SCG_fm achieves RMSE of 236.3 mL and r of 0.28 which did not improve the performance. For other permutations of SCG demodulation schemes, the regressor’s accuracy improves compared to ECG alone as shown in Table 2. Upon combining all the features from ECG and SCG, the RMSE was 188.1 mL and r was 0.59.

The best performing regression model achieves an RMSE of 181.45 mL, r of 0.61 when using all ECG features plus the amplitude demodulated SCG features. The regression plot for this combination of features is shown in Figure 4 a) with different subjects plotted using different colors. The Bland-Altman analysis for the best performing feature set is provided in Figure 4 b). Our model exhibits a bias of −1.47 mL compared to the dynamic range of TV values in the dataset that range from 180 mL to 1600 mL. To further evaluate the information gained by the regressor when trained on different feature, we plot the feature importance for all the features in Figure 5 a) and the feature importance of different feature sets in Figure 5 b). In Figure 5 a), the SCG AM features show the most correlation to the TV as well as the SCG BW then on the third place comes the ECG features. In Figure 5 b) we also show the feature importance of different feature sets, the same pattern is noticed that SCG AM followed by SCG baseline wander and then ECG AM features are the most important and informative features when estimating the TV. Among SCG features, AM features are the most important.

Fig. 4.

a) Correlation plot between actual and estimated TV. The black line is of unity slope and zero intercept. b) Bland-Altman analysis for actual and estimated TV. The middle line is the mean while the two other lines indicate mean ± 1.96 × standard deviation.

We also report the regional performance of our model when estimating TV values at different ranges as shown in table 3. The region with lowest error is for TV values from 767.5 mL to 889.0 mL and has an RMSE of 136.7 mL. While TV values below 631.4 mL shows the highest error with RMSE of 318.8 mL.

TABLE III.

Regional accuracy of estimated TV for the best performing feature set using ECG + SCGam.

Percentile	Range (mL)	RMSE (mL)
<20%	<631.4	318.8
20% – 40%	631.4 – 767.5	152.3
40% – 60%	767.5 – 889.0	136.7
60% – 80%	889.0 – 1084.2	151.3
>80%	> 1084.2	245.4

IV. Discussion

A. Multimodal Fusion

Our improved performance compared to [32] is mainly due to combining ECG and SCG features to estimate TV as compared to their model which only uses ECG features. The authors in [32] used ECG derived features to estimate TV before, during and after varying treadmill exercise. They reported their recovery TV estimation median of absolute error (MAE) was 490 mL, absolute error IQR was 390 mL, median percentage error (MPE) was 19.75% and percentage error IQR is 12.33%. In contrast, our corresponding model results are 112.4 mL for MAE and 152.9 mL for absolute IQR. Our MPE is 13.3% and our percentage error IQR is 18.3%.

Our model achieves a slightly higher percentage IQR compared to [32]. Their study recruited only male participants who are active, practicing aerobic training at least three times a week. Their reported dataset demographic characteristics for weight and height are 74.8±7.0 kg and 178.5±5.5 cm respectively, as compared to our dataset’s corresponding values of 67.5±14.1 kg and 170.5±9.9 cm. Our study compared to [32] recruited both genders, with no specific fitness constraints, and more diverse weight and height, all of which are shown to be relevant for baseline TV values [51]. Accordingly, our dataset has more inter-subject variability, which may have contributed to higher IQR. The subject colored in blue appears to be an outlier since that subject has relatively lower weight, height and TV dynamic range compared to the rest of the group so our decision tree regressor de-emphasizes that subject. It is apparent that regression models can generally estimate real-time TV accurately.

The accuracy of estimated TV is dependent on the TV value. As shown in table 2, the regional accuracy of the model for TV values below the 20% and above 80% is relatively lower than other regions. We believe that this is due to that the number of data points that belong to these regions are underrepresented in the dataset compared to other TV values. So, the model had a smaller number of examples to learn from.

While [32] didn’t report r as one of their performance metrics, the authors in [33] reported an r of 0.87 compared to our best r of 0.61. The authors in [33] only used a subject specific model which needs to be calibrated for every subject before use. Such subject specific calibration eliminates the intersubject variability, and hence results in higher correlation, but it also limits the generalizability and deployment of their model. On the other hand, our model is subject independent and doesn’t require any subject specific calibration. This exposes our model to higher intersubject variability and thus results in relatively lower values of r. We also ran our algorithm on per subject normalized TV values, which makes the model subject dependent but does result in high correlation values like in [33]. That said, we did not include this subject dependent model approach and result in the manuscript to avoid confusion, as we believe that a subject independent model is preferred since calibration is impractical in most real-world scenarios.

Table 4 compares this work to other TV estimation works in the existing literature based on performance, activities performed by subjects during TV measurement and estimation in the study, demographics of the subjects, calibration of the estimation algorithm, portability and placement of the sensor and the modality (technology) used for measurement. Different accuracy measures are adopted by different studies to evaluate TV estimation accuracy. We report these measures as they were presented in the original work.

TABLE IV.

Comparison between this work and previous work in terms of modality used to estimate TV, placement of sensors, portability of the device, the need for subject specific calibration, demographics of the subjects used for the study, activities the study implemented during TV estimation, and the performance of different studies.

Ref.	Modality	Placement	Portable	Calibration	Demographics	Activity	Performance
[32]	ECG	Electrodes on the chest	No	Subject Specific	25 males, age 33.4 ± 5.2 years	Rest, treadmill exercise and recovery	MAE = 490 mL
[33]	Accelerometer	Accelerometer on the sternum	yes	Subject Specific	8 subjects, age range 18–46 years	Controlled breathing at rest	r = 0.87
[54]	RIP belt	Belt around the chest	yes	Subject Specific	5 Healthy subjects, 12 patients with lung diseases	Supine, sitting and upright/walking	m = 0, σ = 0.04
[55]	Strain sensor	Sensor on the chest and abdomen	yes	Subject Specific	8 participants (5 men, 3 women)	Controlled breathing at rest	m = −0.077, σ = 0.15
[56]	IP	Electrodes on the chest	No	Subject/Posture Specific	15 healthy subjects (Male: 9, Female: 6; Age: 25.80 ± 3.30)	Controlled breathing at rest	<3%
[57]	RF	Belt around the chest and abdomen	No	Subject/Posture Specific	20 healthy participants (14 females)	Controlled breathing at rest	r = 0.84
[58]	Doppler radar	Antenna placed near subjects	No	Subject Specific	6 subjects (age 25 ± 5 years)	Controlled breathing at rest	MAE = 0.07 – 0.19
[59]	Camera	Camera placed near subjects	No	Subject Specific	15 healthy subjects (14 males and 1 female; age 28.73 ± 9.27 years)	Controlled breathing at rest	RMSE = 182 mL ± 107 mL
This Work	ECG + Accelerometer	Sensors on the sternum	Yes	Subject Independent	18 subjects (10 males, 8 females; age 26.8±4.1 years)	Exercise recovery	RMSE = 181.5 mL, r = 0.61

The performance of different models is affected by the type of activities. Estimating the TV during rest might impact the estimation accuracy since less motion artifacts influence the sensors. Studies including more subjects of different genders with wider age range introduces more intersubjective variability for the estimation algorithm. All the studies reported in the table, except ours, use subject specific calibration models to handle the intersubjective variability which requires the model to be calibrated before use for each subject. However, our algorithm does not require that extra step. Our used sensor is also one of the few portable options which may allow continuous monitoring of TV.

B. SCG AM is the Most Salient Feature

The relationship between the aortic opening vibration intensity, namely SCG_am,max in this study, and volume of air inhaled during breathing, TV in this study, is shown to exhibit the highest correlation among all other features as shown in Figure 5 a) and b). In Figure 5 a) the SCG_y,am, dorsoventral (DV) amplitude modulated component of SCG, feature is the most salient compared to other SCG directions and demodulation techniques. The driving force behind the decrease of the aortic opening sound intensity is the decrease of the intrapleural pressure during inspiration which is correlated to the volume of inhaled air during inspiration.

As the intrapleural pressure decreases the amount of inhaled air increases but the aortic opening sound intensity decreases according to [51, 52]. During inspiration, the decrease in intrapleural pressure results in increased venous return, and thereby increased preload of the right ventricle. This leads to a decrease in the left ventricle compliance resulting from ventricular interdependence. The decrease in compliance of the left ventricle results in a preload reduction of the left ventricle and hence a reduction in the force of contraction of the left ventricle. This ultimately results in a decreased intensity of the aortic opening heart sound during inspiration.

C. SCG BW is the Second Most Salient Feature

The BW SCG features are shown to be highly correlated with TV. Estimating the TV using a combination of SCG_bw and ECG features, results in more accurate TV estimation as opposed to using ECG features alone. As shown in Table 2, using ECG features only results in RMSE of 235.6 mL and r of 0.26. However, when adding SCG_bw features to ECG ones the RMSE drops to 214.9 mL and the r rises to 0.45. This makes SCG_bw the second most important feature in our feature set for TV estimation which is consistent with SCG_bw being the second most important feature, among those considered, as shown in Figure 5 b). The fact that SCG_bw is highly correlated to TV was also reported in [33]. The authors used SCG_bw features only to estimate TV variability using a subject specific model and reported r of 0.87. Although our model is subject independent and estimates the actual TV, a strong correlation between actual TV values and SCG_bw was also found.

D. SCG Correlates Better to TV than ECG

Another key finding of this work is that SCG is better correlated to TV than ECG as suggested by Table 2. The first row in Table 2 shows TV estimation RMSE of 235.6 mL and r of 0.26 when only using all ECG features for training and testing. However, when using SCG features only, the RMSE is 203.6 mL and r is 0.49. Such performance improvement suggests that SCG features are more suited for indirect TV estimation applications than ECG ones. This is further supported when considering the high correlation between SCG_am, SCG_bw and TV as discussed earlier. Another piece of evidence can be found in Figure 5 b) where different feature sets of ECG and SCG are compared against each other. The SCG features show higher importance to TV estimation compared to ECG features for a specific modulation scheme and hence in total.

E. Limitations and Future Directions

In this work, only the recovery part of the dataset with minimal motion artifacts was analyzed. However, in home monitoring scenarios or when collecting data in the wild, external sources of noise including motion artifacts are expected to impact the quality of the data. Such artifacts may affect the accuracy of TV estimation. Extending our pipeline to handle larger motion artifacts and noise is necessary to achieve better generalization. This can be done through analyzing other parts of the dataset that include treadmill exercises as well as outdoor walks. Optimizing the pipeline to handle extreme noise and motion artifacts is beyond the scope of this work which focuses on features importance and selection but will be an important aspect of future work.

Another important current limitation is that our target population for this study is under-represented in the dataset. This study is focused on measuring the TV for patients with pulmonary disease. Such participants are not existent in the dataset where data is only collected on young and healthy individuals. While it is important to first start with establishing an understanding of the fundamental relationships between ECG and SCG signals and TV in healthy participants, future work should expand upon our initial studies to include individuals with pulmonary diseases and participants with wider age range. We anticipate that the findings that we present in this work about the effective feature selection and fusion will still apply on the target population data in future studies.

V. Conclusion

In this paper we have introduced a wearable, unobtrusive, chest patch with associated algorithms that can accurately estimate TV in healthy young participants. Our proposed model uses multimodal demodulation and fusion of different physiological signals to acquire a robust estimation of TV. The model is subject independent and can be used with no further subject specific calibration required. Although the estimation error might be higher than the range required for clinical use ±3% [10], the performance is better than other approaches with similarly unobtrusive hardware that can be used in the wild. We also explored different physiological features and showed that the AM SCG feature correlate most closely to TV among other ECG and SCG extracted features. Such findings shed light on how to design a more accurate, calibration free TV estimation algorithms that can be deployable for ubiquitous respiratory health monitoring.

VI. Disclosure

OTI is a co-founder and board member for Cardiosense, Inc., a company focusing on commercializing wearable patch and machine learning technologies for cardiovascular monitoring.

Biographies

Moamen M. Soliman received the B.S. degree in electronics and communications engineering from Alexandria University, Egypt, in 2012, and the M.S. degree in electrical and computer engineering from the Georgia Institute of Technology, Atlanta, GA, USA, in 2019. He is currently pursuing the Ph.D. degree in Machine Learning program with the Georgia Institute of Technology, Atlanta, GA, USA. In 2021, he joined the Inan Research Laboratory as a Graduate Research Assistant. His research focuses on using machine learning to develop health monitoring algorithms.

Venu G. Ganti received the B.S. degree in electrical engineering from the Georgia Institute of Technology, Atlanta, GA, USA, in 2017, and the Ph.D. degree in bioengineering from the Georgia Institute of Technology, Atlanta, GA, USA, in 2021. His research interests include the development of non-invasive multimodal physiological monitoring systems.

Omer T. Inan (S’06, M’09, SM’15) received the B.S., M.S., and Ph.D. degrees in electrical engineering from Stanford University, Stanford, CA, in 2004, 2005, and 2009, respectively. He joined ALZA Corporation (A Johnson and Johnson Company) in 2006, where he designed micropower circuits for iontophoretic drug delivery. In 2007, he joined Countryman Associates, Inc., Menlo Park, CA where he was Chief Engineer, involved in designing and developing high-end professional audio circuits and systems. From 2009–2013, he was also a Visiting Scholar in the Department of Electrical Engineering, Stanford University. From 2013–2018, Dr. Inan was Assistant Professor then Associate Professor of Electrical and Computer Engineering at the Georgia Institute of Technology, where he is currently Linda J. and Mark C. Smith Chair in Bioscience and Bioengineering, Professor of Electrical and Computer Engineering, and Adjunct Professor of Biomedical Engineering. His research focuses on non-invasive physiologic sensing and modulation for human health and performance, including for chronic disease management, acute musculoskeletal injuries and disorders, and pediatrics.

Dr. Inan is an Associate Editor of the IEEE Transactions on Biomedical Engineering and the IEEE Journal of Biomedical and Health Informatics, Theme 10 Editor for the IEEE Engineering in Medicine and Biology Conference, Associate Editor for the IEEE Biomedical and Health Informatics Conference, Vice Chair of the IEEE Technical Committee (TC) on Wearable Biomedical Sensing Systems, and Invited Member of the IEEE TC on Translational Engineering for Healthcare Innovation and the TC on Cardiopulmonary Systems, and Technical Program Committee Member or Track Chair for several other major international biomedical engineering conferences. He has published more than 280 technical articles in peer-reviewed international journals and conferences and has eleven issued patents. Dr. Inan received the Gerald J. Lieberman Fellowship in 2009, the Lockheed Dean’s Excellence in Teaching Award in 2016, the Sigma Xi Young Faculty Award in 2017, the IEEE Sensors Early Career Award in 2018, the Office of Naval Research Young Investigator Award in 2018, and the National Science Foundation CAREER Award in 2018. He received an Academy Award for Technical Achievement from the Academy of Motion Picture Arts and Sciences (The Oscars) in 2021. He is a Fellow of the American Institute for Medical and Biological Engineering (AIMBE). He was a National Collegiate Athletic Association (NCAA) All-American in the discus throw for three consecutive years (2001–2003).