Smart Seismocardiography: A Machine Learning Approach for Automatic Data Processing ^†

Abstract

Seismocardiography (SCG) is a non-invasive method that measures local vibrations created by the mechanical cardiovascular exercises on the chest wall. Thereby, mechanical movements of the heart are recorded in real-time from vibration sensors positioned on the chest of the subject, to further compute the heart rate and retrieve the SCG waveform. Although such events have been widely studied, robust signal processing methods remain a challenging task. On the other hand, the use of piezoelectric sensors has been favored in recent years due to its features and low cost. However, robust data processing techniques should be developed to increase their performance and reliability. In this work, we propose an attractive method for SCG data processing based on the K-Means clustering algorithm to automatically label waveform events. Interestingly, the SCG signals are recovered from a custom-made device built around an ultra-low-cost piezoelectric sensor. Once the signals are measured, they are pre-processed by spectral filtering. Afterwards, the signal spectrum is used to compute the heart rate (HR). Thereby, the filtered signal is sequentially segmented, and every frame is processed by a light-weight K-Means algorithm. Finally, we show the performance of the smart seismocardiography by analyzing SCG waveforms at different physiological conditions.

1. Introduction

Cardiovascular disease (CVD) is a major cause of death worldwide [1]. Continuous cardiac monitoring is necessary for diagnosis and to follow-up the CVD. However, common cardiac health monitoring systems are highly expensive and require specialized medical personnel for testing and diagnostic, which implies that patients should go to hospitals whenever they need a check-up. To circumvent such difficulties, several techniques have been proposed besides of the electrocardiography (ECG), among which the seismocardiography (SCG) stands out.

SCG is a non invasive technique to measure vibrations on the chest wall, caused by cardiac mechanical processes, for example, heart valves closure and opening, blood momentum changes and myocardial movements [2,3]. The change in volume, pressure and shape of the heart, during these mechanical processes, produces vibrations on the tissues near the heart generating pulsations in the chest wall [4]. Then, the pulsations are recorded from vibration sensors to retrieve the SCG waveform, to further compute the heart rate and assess the cardiovascular phenomena.

To perform SCG measurements, accelerometers are widely accepted to recover the seismographic signal due to its performance [2]. Nevertheless, it requires a correct sensor placement in the chest wall of the test subjects, which is complicated due to its rigid structure. Recently, piezoelectric sensors have demonstrated its effectiveness to measure SCG signals, enabling flexibility and reliable results [5,6,7,8,9,10]. Piezoelectricity is a phenomenon that occurs in certain crystals that, when subject to a mechanical stress, produce a potential difference at their surface [11]. Piezoelectric materials are small, flexible, and relatively inexpensive. However, the use of piezoelectric brass diaphragms, for SCG, has not been studied as an alternative for developing wearable devices (WD), despite their low cost and small size.

In general, an SCG waveform is labeled using ten cardiac mechanical processes as shown in

Table 1. Cardiac mechanical processes associated with the SCG signal events.

SCG Event	Cardiac Mechanical Process	SCG Event	Cardiac Mechanical Process
AS	Atrial systole	RE	Rapid ventricular ejection
MC	Mitral valve closure	PE	Peak ventricular ejection
IM	Isovolumetric movement	AC	Aortic valve closure
AO	Aortic valve opening	MO	Mitral valve opening
IC	Isotonic contraction	RF	Rapid ventricular filling

, and three cardiac time intervals, delimited by the temporary appearance of signal peaks [12,13,14,15,16]: Isovolumetric Contraction Time (IVCT) (MC to AO), Left Ventricular Ejection Time (LVET) (AO to AC) and Isovolumetric Relaxation Time (IVRT) (AC to MO). Inherently, the SCG wave morphology is complex due to the large inter-subject and intra-subject variability. Depending on age, weight, gender and posture, heart rate, sensor type or position a SCG recording could contain many cardiac cycles with low-quality peaks [17]. Thereby, identifying peaks and events within SCG signals is not straightforward using semi-empirical-based methods, and robust methods should be introduced.

In [18], a model that incorporates different methods to create an accurate ML training model and achieve effective prediction of heart disease from data collection and processing is proposed. The use of ML approaches for data-driven problems has also increased in recent years [4,12,16,17,19,20,21,22,23]. Particularly, ML has been applied to heart beat segmentation and cardiac events. For instance, in [12,17], an automatic annotation of peaks is proposed. Otherwise, unsupervised ML algorithms are used for clustering SCG signals in [4,16] using K-Means to detect patterns.

In this work, we introduce smart seismocardiography as an attractive tool for measurement and data processing to assess cardiovascular events. The SCG signal is measured from a custom-built WD built around an ultra-low-cost brass piezoelectric diaphragm. Once the signals are recorded, they are cleaned-up by spectral filtering. Thus, the filtered signal is sequentially segmented, and each frame is processed by a lightweight K-Means algorithm for clustering and automatic annotation of SCG events.

2. Materials and Methods

Figure 1 shows the block diagram of the smart seismocardiography system. First, the brass piezoelectric sensor is positioned on the lower sternum body, from where the raw signal is measured. Then, the signal was conditioned using a voltage-mode amplifier and digitized with a digital-to-analog converter (ADC)-based acquisition system. Subsequently, the acquired signal was pre-processed with a spectral filtering technique to remove high-frequency components. Afterwards, a peak detection algorithm (PDA) was used to segment the SCG signal using the IC cardiovascular event as the reference (see Table 1). Finally, the SCG segments were examined to find the possible SCG event, which were then processed by the K-Means algorithm to provide an automatic labeling method.

2.1. Sensor and Signal Conditioning

For SCG measurements, we used the CEB-27D44 device, an ultra-low-cost brass piezoelectric diaphragm sensor with a 27 mm diameter [24]. The sensor was placed into the chest wall to measure the pulsations caused by the heartbeat. These induced small deformations in the sensor material, thus producing voltages with an amplitude of around 10 mVpp. To measure the SCG signal, the output of the piezoelectric sensor was amplified using a voltage-mode amplifier, which was built around the TLV2771 operational amplifier [25]. Thereby, the output voltage $V_o$ can be expressed as

$$V_o=V_p\times G+V_{\mathrm{offset}},$$

(1)

where $V_p=q_p/(C_p+C_c)$ is the voltage produced by the piezoelectric sensor, with $q_p$ the electric charge produced by the sensor, $C_p$ the capacitance determined by the area, the width, and the dielectric constant of the material, and $C_c$ the lead capacitance. On the other hand, $G=1+(R_f/R_g)$ is the amplifier gain, which is user-selected given the resistors $R_f$ and $R_g$, thus producing an output voltage swinging around the voltage level $V_{\mathrm{offset}}$.

2.2. Measurement Protocol

Two test subjects participated in our study: S1 (male, 24 years old, 70 Kg) and S2 (female, 25 years old, 60 Kg). The subjects provided their consent and verbally reported no history of cardiovascular disease. The subjects were comfortably seated in a chair with a back. The sensing device was placed on the low sternum fixed with a medical grade transparent film adhesive. The test subjects were asked to relax and hold their breath during data recording. The signal was acquired by means of an ADC with 16-bit resolution and a frequency sampling of 11 kHz, thus leading to a record of approximately $N=$ 300,000 samples per measurement.

2.3. SCG Signal Pre-Processing

Once the SCG signal was recorded, it was pre-processed using a spectral filtering technique with a low-pass frequency of 25 Hz. Since the SCG spectrum covers the infrasonic range [7], we limited the high-pass frequency from the highest peak, in the spectral range of 0.8 to 2.0 Hz, which corresponds to the heart rate (HR) under regular conditions. Owing that each SCG cycle exhibits a minimum peak that corresponds to an IC event [14], the PDA detects the IC events as a reference to indicate the start of the SCG segment in the AS event, this allowing us to perform an automatic segmentation procedure.

2.4. K-Means Algorithm for SCG Clustering

The algorithm uses an iterative process, which aims to cluster an input data set into K groups [26,27,28]. To execute the algorithm, we pass as input the data set and a value of K. The dataset will be the characteristics or features for each point, in this case, the amplitude and sample number of the presumed cardiovascular events. The initial positions of the K centroids will be randomly assigned from any point in the input dataset. Then it iterates in two steps: Data assignment and update centroids.

In the first step, each row in our data set was assigned to the closest centroid based on the Euclidean distance (d) of data vectors X and Y as follows:

$$d(X,Y)=\sqrt{\sum _{i=1}^{i=n}{(x_i-y_i)}^2},$$

(2)

where $x_i$ represent the $i$-th value of horizontal axis in the coordinate plane, $y_i$ stand for the value of the vertical axis in the coordinate plane, and n is the number of observations. Subsequently, the centroids of each group are recalculated. This is done by taking an average of all the points assigned in the previous step. The algorithm iterates between these steps until it meets the following stop criterion: if there are no changes in the points assigned to the groups, or if the sum of the distances is minimized.

3. Results

In Figure 2a, we show the low-pass filtered SCG signal (continuous line) and the IC peaks retrieved by the PDA (marks). As can be seen in Figure 2b, the SCG signal was segmented using the detected peaks and then the average of the segments was computed to account for the variability of the measurement. Therefore, once the segmentation succeeded, the presumed SCG events, from each cycle, were used as the input dataset for the clustering algorithm.

As the input of the K-Means algorithm, we considered the signal amplitude and sample number of each presumed SCG event. Moreover, each event was discriminated depending on whether it is an event associated with a minimum peak or a maximum peak. Subsequently, the algorithm clusters first the presumed events associated with maximum peaks (AS, MC, AO, RE, PE, AC, RF), and then, those associated with minimum peaks (IM, IC, MO). Results of the K-Means algorithm are depicted in Figure 3, alongside the SCG average, as a reference. Each label was then assigned in the order of appearance of the clusters.

4. Discussion

4.1. Discussion

In Figure 2b, we show the SCG signal segments founded by the FPA and then the average of the segments. It is worth noticing that the SCG signal showed negligible differences in the morphology of each segment and, instead, it was systematic and uniform throughout the segments. The performance of this segmentation technique obtained results similar to the DTW used in [16], finding traces of similar features and lengths for each SCG cycle. In addition, the use of peak detection techniques has a lower computational cost compared to DTW techniques. However, the FPA depends on good quality SCG signal morphology, so it is limited to acquisitions, where the test subject is holding their breath.

The clustering procedure showed an excellent performance by grouping each of the cardiovascular events with enough accuracy and sensitivity, as shown in Figure 3. Interestingly, the IC cluster does not show temporal variability because the IC peaks were used as a reference for segmentation, so the temporal variability in the rest of the clusters is relative to the IC cluster.

To assess the variability of the proposed method, in Figure 4 we show box plots for the statistical analysis of the SCG clusters of subjects S1 and S2.

The analysis was performed by considering the variation of the amplitude in each cardiovascular event (Figure 4a,d). Therein, the analysis shows low variability between and within subjects (tens of mV), despite physiological conditions. The amplitude of the AO peak in S1 is the highest peak, while in S2 it is the RF peak. In both cases, the lowest peak corresponds to the IC peak. On the other hand, Figure 4b,e shows the variability in the time differences for each SCG cluster. The IC cluster shows temporal variability approximately equal to zero due to the IC peaks were used as a reference for segmentation. It is worth noticing that PE and AC clusters are those with the large temporal variability (20 to 50 ms), whereas the other clusters do not exceed 20 ms. Finally, as shown in Figure 4c,f, it was possible to estimate the cardiac time intervals, thus retrieving three clusters with similar mean values each for both test subjects. This makes sense since S1 and S2 were subjects with normal cardiovascular conditions, which guarantees reproducible results for the smart seismocardiography.

4.2. Future Work

In the future, we aim to compare the performance of the acquisition system. Since the ECG signals possess a more simple morphology compared to the SCG signals, make simultaneous acquisitions of SCG and ECG to evaluate the segmentation performance from FPA and other techniques are proposed. Hence, and similar to [18], from the results of this segmentation and clustering of SCG events, it will be possible to predict and characterize these events using supervised ML algorithms. In [19], the identification of CVD risk factors was approached with ML techniques and data analysis. This is the reason the relationship between variability in SCG groups and CVD needs further investigation. In the future, a more detailed multivariable analysis will be applied, to know how the relative amplitude and time difference of each SCG cluster are related to each other and their relationship with CVD.

5. Conclusions

In this work, we introduced the so-called smart seismocardiography (SCG) as an attractive method for clustering cardiovascular events. The proposal worked around a wearable device (WD) based on an ultra-low-cost brass piezoelectric sensor that captures the mechanical vibrations of the heart. We showed how the K-Means algorithm can automatically cluster SCG events using unsupervised ML techniques, which remains a scientific challenge. Preliminary results indicated that WD coupled with ML leads to a powerful tool for retrieving information on cardiac mechanical processes and cardiac time intervals. We showed how the smart seismocardiography could serve as proof-of-concept to design novel home-made and cost-effective and smart devices, exhibiting enough sensitivity and accuracy to automatically assess physiological signals.