Delineation and Analysis of Seismocardiographic Systole and Diastole Profiles
aa r X i v : . [ ee ss . SP ] F e b Delineation and Analysis of SeismocardiographicSystole and Diastole Profiles
Tilendra Choudhary,
Member, IEEE,
M.K. Bhuyan,
Senior Member, IEEE , and L.N. Sharma
Abstract —Precise estimation of fiducial points of a seismocar-diogram (SCG) signal is a challenging problem for its clinicalusage. Delineation techniques proposed in the existing literaturedo not estimate all the clinically significant points of a SCG signal,simultaneously. The aim of this research work is to propose adelineation framework to identify IM, AO, IC, AC, pAC andMO fiducial points with the help of a PPG signal. The proposeddelineation method processes a wavelet based scalographic PPGand an envelope construction scheme is proposed to estimatethe prominent peaks of the PPG signal. A set of amplitude-histogram based decision rules is developed for estimation ofSCG diastole phases, namely AC, pAC and MO. Subsequently,the systolic phases, IM, AO and IC are detected by applyingdiastole masking on SCG and decision rules. Experimental resultson real-time SCG signals acquired from our designed data-acquisition-circuitry and their analysis show effectiveness of theproposed scheme. Additionally, these estimated parameters areanalyzed to show the discrimination between normal breathingand breathlessness conditions.
Index Terms —Seismocardiogram; ECG; Photoplethysmogra-phy; Wavelet scalogram; Cardiorespiratory; Heart signals I. INTRODUCTION P HYSIOLOGICAL signal generated through heart me-chanical actions, such as seismocardiogram (SCG) can beemployed for assessment of cardiac performance [1], [2]. TheSCG signal reflects cardiac-generated mechanical vibrations,which are recorded by mounting an accelerometer on thechest surface [3], [4]. A heart cycle includes many importantcardiovascular events, such as aortic valve opening (AO)and its closure (AC), mitral valve opening (MO) and itsclosure (MC), rapid blood filling (RF) and rapid ejection (RE),isovolumic moment (IM), and isotonic contraction (IC). Peaksand troughs of systole and diastole profiles of a SCG cycleindicate these events, which are called fiducial points [5].Delineation of SCG waveform refers to the determination offiducial points and the estimation of time intervals between thefiducial points, which are clinically significant for pathologicalinterpretations of a cardiovascular system [4]. The currentresearches mainly focus on the definition and identificationof fiducial points of a SCG signal and relate them with theechocardiogram for identification of physiological events ofa cardiac cycle [4], [6]. Fig. 1 illustrates the simultaneousrecordings of ECG, fingertip-photoplethysmogram (PPG), andSCG signals. Rhythmic variations of blood-volume at differentparts of the body is generally observed due to the pumpingactions of the heart, which is indicated by a PPG signal [7].
T. Choudhary, M.K. Bhuyan, and L.N. Sharma are with the Departmentof Electronics and Electrical Engineering, Indian Institute of TechnologyGuwahati, Assam-781039, India (e-mail: { tilendra, mkb, lns } @iitg.ac.in). Fig. 1. Simultaneous recordings of ECG, fingertip-PPG and SCG signals. The delineation of a SCG signal helps in evaluation ofcardiovascular parameters, such as extraction and analysis ofheartbeats, estimation of heart rate variability (HRV) [8]–[16], hemodynamic parameter estimation [17], associationwith other cardiac signals [18], detection of valvular defectsand diagnosis of other cardiovascular diseases [19], [20]. Thedelineation can also improve the ability of wearable SCGsensor nodes in the development of pervasive wireless bodyarea networks [1]. An extensive investigation has been done forannotation of AO phase in the SCG signal [5], [8]–[13], [15],[16], [21]. Most of the studies employed ECG as a referencesignal for finding out cardiac information from a SCG signal[11], [13], [15]. The determination of IM, AO and AC pointsare proposed in [9] using a high pass filter and a triple integra-tion technique. To annotate AC and IM points, high frequencycardiac acceleration (HFACC) envelope was suggested in [12].Khavar et al. presented a method to estimate IM, AO and ACphases in SCG signal [8]. The method employed a high passfilter, a moving averaging filter, an extrema search block and aprobabilistic decision making technique on systolic, diastolicand heart-rate envelopes. In [22], the SCG systole points areannotated using calibration, heartbeat detection, and prototypematching scheme. By utilizing orthogonal subspace projectionscheme, the fiducial points AO and pAC (peak just after AC)were estimated in our previous reported work [15].Typically, the AO point creates a prominent peak under thesystolic profile of the SCG signal [21]. Thus, other informativesystolic points may be estimated with the help of the AOpeaks. However, delineation of a SCG diastole profile is stilla challenge, and the task is even more complicated when bothsystole and diastole profiles have similar characteristics. Itis observed in Fig. 1 that the diastole delineation would becomparatively easier with the help of a PPG signal. This isdue to the fact that projections of PPG peaks lie on AC-to-MO segment of a SCG signal. The PPG signal is less
Fig. 2. Difficulty in localization of a SCG profile in presence of interbeatand intrabeat variabilities in contrast to PPG signals. noisy, while a SCG signal changes its basic characteristics ifthe sensor placement is affected by the body movements. Asillustrated by a concurrently recorded SCG and PPG signalsfrom two subjects in Fig. 2, it is difficult to localize a SCGprofile due to inter and intra beat variabilities. Moreover, theexisting methods are limited to estimate only a few fiducialpoints, through which the characterization of both systolic anddiastolic profiles of the SCG would not be possible.To address these issues, a PPG-assisted automatic delin-eation framework for a SCG signal is proposed, which canestimate IM, AO, IC, AC, pAC and MO fiducial points insystole and diastole profiles. The framework can also estimatetwo important cardiac intervals, namely IVRT (AC-to-MO)and LVET (AO-to-AC). Multiscale scalogram is employed forspectro-temporal localization of nonstationary and nonlinearsignals [23]. To analyze the SCG waveform, a PPG signal isemployed and wavelet scalogram-based time-frequency anal-ysis is performed onto it. An envelope construction methodis proposed and decision rules are used for estimation offiducial points. In addition, one of its applications is presentedin cardiorespiratory analysis in order to show the abilityof the delineated points. Since a change in the respiratoryeffort-level affects the morphology of the SCG waveform [2],[24], the fiducial points of a SCG signal may be used inassessing the fitness of a respiratory system. As an extensionof the proposed delineation work, the breathlessness (shortnessof breath or dyspnea) and normal breathing conditions areperfectly discriminated by the delineated points. The proposedmethod is elaborately discussed in the following sections.II. P
ROPOSED M ETHOD
Wavelet transform (WT) represents a signal as linear combi-nations of a basis function-set and its variations. The variationis achieved by means of scaling ( l ) and shifting ( m ) parametersof a single mother wavelet function ψ ( t ) . So, signal x ( t ) canbe analyzed using WT as [25]: w l,m { x } = 1 √ a Z + ∞−∞ x ( t ) ψ ∗ (cid:18) t − ml (cid:19) dt, l > (1)where, w l,m represents m th wavelet coefficient at l th scaleand the asterisk ( ∗ ) is a complex conjugate operator. Since,the second derivative of a Gaussian function ( Gaus-2 ) has amorphological similarity with the pulsatile profile of a PPGsignal, it is chosen as the mother wavelet for continuouswavelet transform (CWT). The prototype wavelet ψ ( t ) canbe defined as [26]: ψ ( t ) = − dg ( t ) dt = C (1 − t ) e − t / (2) where, g ( t ) and C denote a Gaussian function N (0 , andnormalization factor, respectively. A. Processing of Scalogram
A scalogram shows 3-D multiscale distributions of signalcomponents with the help of wavelet transformation [23], [26].It indicates the relative wavelet energy for each coefficient w l,m , which is obtained as [26]: E l,m = | w l,m | (3) S l,m ←− E l,m P l P m E l,m (4)where, E l,m and S l,m denote wavelet energy density andscalogram, respectively. A PPG-scalogram is obtained byprocessing an input PPG signal upto 150 scales, which isillustrated in Fig. 3. The corresponding scalogram is shownin Fig. 3(b) and subsequently, its all possible two dimen-sional interpretations are shown in Fig. 3(c)–(f). From thisrepresentation, a time series of maximum relative waveletenergy (MRWE) is obtained irrespective to the scales as shownin Fig. 3(c). A larger MRWE value corresponds maximumresemblance between the wavelet function and the pulsatilewaveform, which in turn gives an indication of the presenceof PPG peaks. Thus, the MRWE signal is selected for furtheranalysis. The PPG signal is detrended and subsequently, addedup with the MRWE signal to emphasize the PPG-peaks andto suppress other spurious peaks. The detrending processremoves the baseline-drift (BD) from the signal, which uses ahigh pass filtering scheme. The signal produced by ensembled-sum is denoted by x ∗ p [ n ] . B. Proposed Envelope Construction Scheme
The Shannon entropy (SE) and Shannon energy (SEE) basedenvelopes produce minima point for higher normalized-signal-amplitude. The methods for estimating envelopes such as ab-solute, energy, analytic signal, SE and SEE were compared in[5]. The ensembled-sum signal x ∗ p [ n ] is employed for envelopeconstruction using SE and SEE as shown in Fig. 4(b)–(e).They provide bifurcated profiles for input peaks having largerelative amplitudes. Hence, a modified transfer characteristic isproposed as shown in Fig. 4(f), which can effectively estimatethe envelope profile of x ∗ p [ n ] by creating sharp impulses atthe locations of PPG-peaks (Fig. 4(g)). In contrast to othermethods, it emphasizes the PPG-peaks without bifurcation andsuppresses the other regions relatively better. The proposedprofile can be obtained via fitting an exponential curve in aleast square sense. Mathematically, it can be represented viatwo exponential expressions, (1 − e − qx ) and / (1 + e − qx ) ,where x corresponds to the input vector. Hence, the functionfor fitting of the desired characteristic can be expressed as: y = 1 − e − qx pe − qx , ∀ x ∈ [0 , , p, q ∈ N (5)To achieve the desired profile, the modelling parameters p and q need to be estimated for different input values. Fig. 5shows a variation of correlation between desired and esti-mated characteristic. The estimated transfer characteristic is Sample number -0.500.51 A m p li t ude M a x . r e l a t i v e w a v e l e t ene r g y × -3 MRWE × -3 S c a l e s S c a l e s × -4 Sample number S c a l e s Scales S a m p l e s (d) (a)(e)(f)(c)(b) Fig. 3. Scalogram of PPG signal. (a) Input PPG signal, (b) PPG scalogram, (c) change of maximum relative wavelet energy (MRWE) with time, (d) variationof scales with time, (e) predominating energy (MRWE) with the higher scales, and (f) overlapping area of time and scales.Fig. 4. Performance comparison of the proposed envelope construction for x ∗ p [ n ] . (a) PPG and ensemble sum signal; (b)-(g) transfer characteristics andenvelope profile estimation using using SE, SEE and the proposed method.Red dots denote local peaks of the envelope impulses.Fig. 5. Estimation of modeling parameters. (a) Desired transfer characteristicin a piecewise linear fashion, and (b) mesh plot representing variations ofcorrelation for obtained characteristic corresponding to curve fitting parame-ters ‘ p ’ and ‘ q ’ [Eq. (5)]. Estimated parameters corresponding to maximumcorrelation are highlighted with 95% confidence interval using gray colorbands. Transfer function produced by fitted parameters achieves maximumcorrelation. achieved by varying the parameters. The fitted parametersare also highlighted with 95% confidence intervals, whichare presented by gray rectangular patches in p - q plane. Bythe usage of fitted p and q parameters, the designed modelshows the highest correlation with the desired characteristicas shown in Fig. 5(b). Thus, in an average least square sense,parameters p and q are estimated as 39 and 16, respectively.In the proposed method, this envelope extraction scheme isused for the detection of true PPG-peaks and later, for theestimation of AO peaks. C. Detection of PPG-Peaks
This developed model is applied to the ensembled-sumsignal x ∗ p [ n ] to get the desired envelope profile. The envelopeprovides localization of PPG-peaks in the form of sharpimpulses. The peaks of these impulses can be easily identifiedusing simple amplitude-temporal based thresholding schemes. D. Delineation of Diastolic Regions
The diastole profiles of a SCG signal are estimated with thehelp of PPG signals at the initial stage. The detected PPG-peaks are found useful in estimating the pAC peaks in SCG-diastole profile. The pAC point corresponds to the maxima in aSCG diastole, and it can be located by searching the peaks of aPPG signal with the help of a narrow and symmetric window.Subsequently, a proposed set of amplitude-histogram baseddecision rules can be employed for estimation of adjacentfiducial points (AC and MO) as follows:
1) Amplitude-Histogram based Proposed Decision Rules:
The decision rules (Algorithm 1) block is employed on bothdiastole and systole profiles of a SCG signal. At first, each ofthe SCG diastolic profiles is segmented into two small blocksadjacent to pAC peaks, and all local peaks are detected for boththe blocks. It is observed that amplitude of the peaks gradually
Algorithm 1
Proposed decision rules
Inputs: x s [ n ] , pks i =1 , ,..., pks // x s [ n ] corresponds to the input SCG signal// pks corresponds to pAC peaks in diastole (or AO peaks in systoleanalysis)1: Segment the SCG profiles into two blocks:blockA i = x s [( pks i − ms ) : pks i ] blockB i = x s [ pks i : ( pks i + 200 ms )]
2: Detect all local peaks of blockA and blockB.3: Process the following steps for each of the blocks:4: for i = 1 to pks do (a) Compute amplitude histogram of local peaks: fr e qu e n c yb i n s F i ← peaks lying in amplitude range (80 − of pks i F i ← peaks lying in amplitude range (60 − of pks i ... ... ... ... ... F i ← peaks lying in amplitude range (0 − of pks i (b) Estimate first significant peak nearest to pks i for j = 1 to do if n ( F ji ) ≥ then // where n ( · ) denotes number of elements.9: P i ← Peak nearest to pks i in frequency bin F ji ;10: break for loop ;11: end if end for (c) Detect minimum points between P i and pks i from bothblocks A and B → M i and M i , respectively14: end forOutputs: Fiducial points: M , M and pks decreases from pAC locations. Hence, each of the blocks isemployed for the following set of operations. Five frequencybins are created to show the distribution of local peaks withrespect to the pAC point under a segmented block of diastoleprofile as listed in step 5 (Algorithm 1). Subsequently, all thefrequency bins are scanned sequentially ( F – F ) to find non-zero elements (local peaks). In the scanning process, the firstbin corresponding to have non-zero elements is observed andselected for further investigation. Among all these local peaksfrom the selected bin, the first significant peak closest to pACis estimated, say P . Then, the segment P to pAC is appliedfor minima search for both the blocks, which can estimate ACand MO fiducial points. E. Delineation of Systolic Regions
Once the fiducial parameters of the diastolic region aredetermined, delineation is carried out for the systole. If thefiducial points, AC, pAC and MO are perfectly determined,then it is possible to precisely localize diastolic profiles inthe SCG signal. Usually, both systolic and diastolic profilesshare the same spectral range and/or have similar ampli-tude strengths. Due to that, identifying systolic profile inthe presence of diastolic profile in a SCG cycle becomes agreat challenge. Thus, the diastole profiles are masked in theSCG signal by replacing their corresponding samples withzero amplitude levels. Typically, the AO peak possesses theprominent amplitude under a systole profile. But, its actuallocation cannot always be found by determining the maximaof the masked-SCG-cycle, which is created by a segmentbetween any two same diastole-fiducials. The reason behindis sometimes, the AO peak has a smaller amplitude than
Fig. 6. Block diagram of the proposed method.TABLE ID
EMOGRAPHICS OF SUBJECTS AND THEIR RECORDING INFORMATION
16 8 M 28.75 71.63 5’7.6” 79.18 3429 3425 ± ± ± ± MC or RE, which are the peaks adjacent to AO in SCGsignal. The SCG segment, IM-AO-IC shows frequency rangetypically higher than other wave-segments in the SCG-systole.So, the masked-SCG is bandpass filtered in a relatively higherfrequency range of 20–30 Hz. Due to that, the AO peaks areemphasized in the resulting signal. Then, its upper envelopeis extracted and subsequently, its instantaneous energy signalis employed to the proposed envelope construction technique(Eq. (5)). Finally, the locations of all the AO peaks, which areindicated through impulses in the resultant signal, are detectedusing simple thresholding and peak correction scheme. As likediastole profile, the proposed decision rules (Algorithm 1) aresubsequently applied to the masked-SCG signal with detectedAO peaks. It can localize two another systole fiducial points,IM and IC.A block diagram of the proposed method for estimation ofcardiac phases is shown in Fig. 6. Initially, MRWE signal isextracted from the PPG-scalogram, which is further summed-up synchronously with the BD-suppressed PPG signal. Subse-quently, an envelope is constructed on the output signal usingthe proposed scheme, and PPG-peaks are detected. The PPG-peaks can detect the pAC peaks in the SCG signal. Otheressential diastole phase instants, AC and MO are estimatedusing the proposed decision rules. After this process, thediastole profiles are localized and masked in the SCG signal,and subsequently, bandpass filtering is employed to estimatethe AO instants. Similarly, other important systole phaseinstants, IM and IC are estimated using the proposed decisionrules. Two important cardiac time intervals, namely LVET andIVRT can also be determined using the estimated AO, AC andMO parameters. III. E
XPERIMENTAL S ETUP
A. Measurement Protocol and Data Acquisition
The SCG signals were acquired from eight male subjects bymounting the sensor node near lower end of the sternum onthe chest wall. The demographics of involved subjects and therecording information are tabulated in Table I. Additionally,
Fig. 7. Peak detection process of a PPG signal. (a) Detrended PPG d [ n ] , (b)MRWE signal r [ n ] , (c) d [ n ] + r [ n ] , (d) envelope, (e) peaks of the PPG.Fig. 8. Determination of parameters of SCG diastoles with the help of PPGpeaks. (a) PPG signal with its peaks, (b) SCG signal with detected fiducialpoints (AC, pAC and MO) of diastoles using the proposed decision rules. the ECG (Lead-II) and the PPG (at fingertip) signals werealso acquired in a standard setting procedure. All these threesignals are digitized and synchronized using Biopac MP150DAQ system at a sampling rate of 1 kHz. The signals wererecorded in two different sessions: normal breathing for 5 min-utes followed by holding breath for 50 s. So, breathlessnesscondition is artificially achieved for every subject by holdingor stopping his respiratory activity. With an ethical approvaland proper consent of subjects, the SCG signals were recordedusing a small sized signal recording system. The designed PCBfor SCG signal acquisition is shown in Fig. 10. The systemconsists of a miniaturized MEMS accelerometer (ADXL335, ± B. Experiment
The proposed method is tested and validated with 16different signals. According to breathing pattern, entire datacan be divided into normal breathing (NBDB) and stoppedbreathing (SBDB) datasets. The proposed method employssimultaneously recorded SCG and PPG signals, while ECG is
Fig. 9. Delineation process for systolic profiles of SCG. (a) Input SCG signal,(b) masking of diastole boundaries with the help of estimated diastolic fiducialpoints, (c) band pass filtered signal and its upper envelope construction, (d)final envelope created by using the proposed envelope scheme, (e) detectedAO instants, and (f) estimated IM and IC instants using the decision rules.Fig. 10. Back and front views of our designed PCB for SCG signal recording.Fig. 11. Experimental setup for recording of signals. used as a reference signal for manual annotation of the SCGsignal. The proposed delineation framework is employed oneach of the signals with a 10 s time window. The PPG-peaksare determined initially with the help of wavelet scalogram, asshown in Fig. 7, which further helps in locating the pAC-peaksunder diastole profiles. Subsequently, the adjacent fiducialpoints AC and MO are estimated using the proposed histogrambased decision rules. It is illustrated with an example in Fig. 8.The localization of diastolic region helps in precise detectionof AO peaks under systole profiles, and the neighborhoodfiducial points (IM and IC) can also be located with the help
Fig. 12. Error-bar graph provides distribution of extracted features for boththe classes. The distribution is shown via statistical parameters mean (red/bluecolor dot) and standard deviation (symmetric vertical lines around mean). of same decision rules. The entire process of estimation ofsystolic parameters is shown in Fig. 9. For both the profilesin the SCG waveform, the proposed envelope constructionscheme helps in indicating the prominent peak locations.IV. R
ESULTS AND D ISCUSSION
A. Performance Evaluation
The performance of the proposed delineation framework isanalyzed for all the estimated parameters of both systolic anddiastolic profiles. The proposed method is tested with SCGsignals of NBDB and SBDB databases. The performance isevaluated using three quantitative measures namely, the sen-sitivity (cid:0) Se = TPTP+FN (cid:1) , the positive predictivity (cid:0) +P = TPTP+FP (cid:1) and the accuracy (cid:0)
Acc = TPTP+FP+FN (cid:1) , where TP, FP and FNhave their usual meaning [27]. The overall performance inthe estimation of systolic fiducial points AO, IM and ICis summarized in Table II. The proposed method achievesaverage accuracies (in %) of 95.2, 88.18 and 85.32 for AO, IMand IC points, respectively, on NBDB database, while 90.08,86.09 and 89.13 are achieved for AO, IM and IC, respectively,on SBDB database. The maximum sensitivity and positivepredictivity are found in the detection of prominent AO peaksas compared to other systolic points.In a similar manner, the performance in the estimation ofdiastolic fiducial points pAC, AC and MO is summarizedin Table III. The proposed method achieves mean accuracies(in %) of 98.96, 97.55 and 94.54 for pAC, AC and MO points,respectively, on NBDB database, while 98.27, 95.06 and 95.06are achieved for pAC, AC and MO, respectively, on SBDBdatabase. The maximum sensitivity and positive predictivityare found in the detection of pAC instants as compared toother diastolic points, which are 99.04 and 99.93 for NBDB,respectively, and 99.58 and 98.69 for SBDB, respectively. Theobtained results clearly show the excellent performance of ourdelineation method.
B. Applicability in Respiratory-Effort Level Identification
The delineated parameters can be used for detection andclassification of many cardiovascular events. To show theapplicability of the proposed framework, a study of SCG-basedcardiorespiratory analysis is performed. For this, the effectof respiratory effort levels reflected in SCG-morphology isobserved. Based on morphological variations, the respiratory
TABLE IIP
ERFORMANCE OF THE PROPOSED METHOD IN THE DETECTION OFSYSTOLIC FIDUCIAL POINTS
Record AO Instant Detection IM Instant Detection IC Instant DetectionSe (%) +P (%) Acc (%) Se (%) +P (%) Acc (%) Se (%) +P (%) Acc (%) N B D B s01a 99.11 99.7 98.81 96.73 97.31 94.2 99.1 99.4 98.52s02a 90.98 91.48 83.88 90.98 91.48 83.88 90.98 91.48 83.88s03a 98.28 99.31 97.61 79.11 79.93 66 85.27 86.16 75s04a 97.65 98.31 96.04 96.31 96.96 93.49 88.26 88.85 79.46s05a 95.51 95.73 91.61 88.42 89.26 79.91 94.56 94.79 89.89s06a 99.54 98.87 98.42 97.27 96.61 94.05 99.77 99.09 98.87s07a 98.91 99.56 98.47 98.47 99.12 97.61 75.49 75.99 60.95s08a 97.31 99.45 96.79 97.04 99.18 96.27 96.51 99.45 95.99 Overall 97.16 97.8 95.2 93.04 93.73 88.18 91.24 91.9 85.32 ± ± ± ± ± ± ± ± ± s01b 100 100 100 100 100 100 100 100 100s02b 75.41 86.79 67.65 75.41 86.79 67.65 75.41 86.79 67.65s03b 95.74 97.83 93.75 91.49 93.48 86 95.74 97.83 93.75s04b 88.24 88.24 78.95 86.54 88.24 77.59 86.54 88.24 77.59s05b 98.33 98.33 96.72 98.33 98.33 96.72 95 95 90.48s06b 100 100 100 100 100 100 100 100 100s07b 96.55 94.92 91.8 94.83 93.22 88.71 96.55 94.92 91.8s08b 94.92 96.55 91.8 83.05 84.48 72.06 94.92 96.55 91.8 S B D B Overall ± ± ± ± ± ± ± ± ± Overall performances are given in terms of mean ± standard deviation. TABLE IIIP
ERFORMANCE OF THE PROPOSED METHOD IN THE DETECTION OFDIASTOLIC FIDUCIAL POINTS
Record pAC Instant Detection AC Instant Detection MO Instant DetectionSe (%) +P (%) Acc (%) Se (%) +P (%) Acc (%) Se (%) +P (%) Acc (%) N B D B s01a 98.81 100 98.81 98.21 99.7 97.92 98.21 99.4 97.63s02a 98.9 100 98.9 98.9 100 98.9 98.36 99.45 97.82s03a 99.32 99.66 98.98 98.97 99.31 98.3 99.32 99.66 98.98s04a 99.33 100 99.33 97.64 98.31 96.03 97.98 98.64 96.68s05a 98.82 100 98.82 98.58 99.76 98.35 96.22 97.37 93.78s06a 98.86 99.77 98.64 98.41 99.77 98.18 97.72 99.08 96.84s07a 98.25 100 98.25 95.84 97.77 93.79 95.4 97.1 92.77s08a 100 100 100 99.46 99.46 98.93 90.03 90.03 81.86 Overall 99.04 99.93 98.96 98.25 99.26 97.55 96.65 97.59 94.54 ± ± ± ± ± ± ± ± ± s01b 100 100 100 100 100 100 98.15 98.15 96.36s02b 100 89.55 89.55 100 89.55 89.55 95 87.69 83.82s03b 100 100 100 100 100 100 97.83 97.83 95.74s04b 100 100 100 100 100 100 98.08 98.08 96.23s05b 100 100 100 100 100 100 100 100 100s06b 98.28 100 98.28 87.72 89.29 79.37 94.74 94.74 90s07b 100 100 100 96.49 96.49 93.22 100 100 100s08b 98.33 100 98.33 98.33 100 98.33 98.33 100 98.33 S B D B Overall ± ± ± ± ± ± ± ± ± Overall performances are given in terms of mean ± standard deviation. conditions of a subject can be categorized as normal breath-ing and breathlessness conditions. A set of twelve extractedfeatures (f k ∈ R , k = 1 , , .. ) is studied for this purpose,and they are AO-AO interval derived heart-rate (HR), relativetiming information of AO, IC, AC, pAC and MO with respectto IM, normalized signal amplitudes corresponding to IM,AO, IC, AC, pAC and MO points. The features are extractedfrom each of the cardiac cycles and normalized. In this study,a 40 s SCG-segment from each of the signals is used forfeature extraction. The distributions of normalized features aredisplayed in Fig. 12 in terms of mean ± standard-deviation.More discriminative and independent features are selected TABLE IVP
ERFORMANCE RESULTS OF DIFFERENT CLASSIFIERS FOR STOPPEDBREATHING DETECTION WITH
FOLD CROSS VALIDATION F e a t u re s M e t r i c s SVM kNN NB Discrim. Ana.RBF Linear Fine Med. Coarse LDA QDA
ACC ∗ TPR
Abbreviations – SVM: support vector machine, RBF: radial basis function, kNN:K-nearest neighbour, Fine: kNN with K = 5, Medium: kNN with K = 11, Coarse: kNNwith K = 101, NB: naive Bayes, LDA: linear discriminant analysis based classifier,QDA: quadratic discriminant analysis based classifier. ∗ SF: selected feature-set. Note that all the metrics are presented here in terms of %.
Fig. 13. ROC of different classifier-models on selected features. SVM-RBFbased classifier outperforms others by achieving the largest AUC of 0.9986. after a statistical analysis of feature distributions and the useof standard T-test procedure. The selected features are f k ∗ ∈ R , k ∗ = 1 , , , , , , , for this event classificationtask. Finally, support vector machine (SVM) with RBF kernelis used to identify the breathlessness condition. The choiceof training and test cycles is done using the standard 10-foldcross-validation technique [28].The classification performance is evaluated and comparedwith various other classifiers. The performance evaluation isdone with three measurement parameters, recognition accu-racy (cid:0) ACC = TP+TNTP+TN+FP+FN (cid:1) , true positive rate (cid:0)
TPR = TPTP+FN (cid:1) and false positive rate (cid:0)
FPR = FPTN+FP (cid:1) [29]. The averageperformance metrics are tabulated in Table IV. The SVM-RBF based classifier outperforms others by producing thesemetrics as 97.25, 96.72 and 2.29, respectively for consider-ing all features, and 98.35, 97.32 and 0.76, respectively forconsidering selected feature-set. Fig. 13 shows the estimatedreceiver operating characteristic (ROC) curves plotted betweenTPR and FPR, and area under the ROC curve (AUC) for theidentification of SB class. The best performance is achievedwith the SVM-RBF classifier. This experiment validates thatthe delineated parameters from the proposed method could beemployed not only for the cardiac analysis, but also for theassessment of respiratory system. V. C
ONCLUSION
In this work, a novel framework is proposed for delineationof systole and diastole parameters of a SCG signal withthe help of a temporally concurrent fingertip-PPG signal.The proposed SCG delineation framework is mainly basedon wavelet scalogram, the proposed envelope constructionscheme, and the proposed decision rules. It consists of threemajor blocks– estimation of PPG-peaks, estimation of diastole-phases (AC, pAC and MO) in the SCG and estimation ofsystole-phases (IM, AO and IC) in SCG. Along with thedelineated parameters, LVET and IVRT time-intervals can alsobe estimated from the parameters. The experimental results onNBDB and SBDB databases show very promising results, andso, the proposed framework may be deployed for cardiac-cycleevent detection for pathological interpretations. The salientpoint of the proposed method is that it can efficiently estimatesix important cardiac phases simultaneously, which character-ize events associated with both systolic and diastolic regions.Additionally, as an application for cardiorespiratory analysis,the information extracted from the estimated fiducial pointscan be used to identify normal breathing and breathlessnessconditions. The SVM-RBF classifier gives good results onour selected feature-set. However, the framework needs tobe tested with cardio- and respiratory related pathologicalscenarios for its clinical applications.R
EFERENCES[1] M. Di. Rienzo et al. , “Wearable seismocardiography: Towards a beat-by-beat assessment of cardiac mechanics in ambulant subjects,”
AutonomicNeuroscience , vol. 178, no. 1–2, pp. 50 – 59, 2013.[2] A. Taebi et al. , “Recent advances in seismocardiography,”
Vibration ,vol. 2, no. 1, pp. 64–86, 2019.[3] O. T. Inan et al. , “Ballistocardiography and seismocardiography: Areview of recent advances,”
IEEE J. Biomed. Health Inform. , vol. 19,no. 4, pp. 1414–1427, 2015.[4] K. Sørensen et al. , “Definition of fiducial points in the normalseismocardiogram,”
Scientific reports Nature , vol. 8, no. 1, pp. 15455,2018.[5] T. Choudhary, L. N. Sharma, and M. K. Bhuyan, “Automatic Detectionof Aortic Valve Opening Using Seismocardiography in Healthy Individ-uals,”
IEEE Journal of Biomedical and Health Informatics , vol. 23, no.3, pp. 1032–1040, 2019.[6] R. S. Crow et al. , “Relationship between seismocardiogram and echocar-diogram for events in the cardiac cycle,”
American Journal of Noninva-sive Cardiology , vol. 8, no. 1, pp. 39–46, 1994.[7] A. B. Hertzman, “The blood supply of various skin areas as estimatedby the photoelectric plethysmograph,”
American Journal of Physiology-Legacy Content , vol. 124, no. 2, pp. 328–340, 1938.[8] F. K.-Khavar et al. , “Automatic and robust delineation of the fiducialpoints of the seismocardiogram signal for non-invasive estimation ofcardiac time intervals,”
IEEE Trans. Biomed. Eng. , vol. 64, no. 8, pp.1701–1710, 2017.[9] F. K.-Khavar, K. Tavakolian, and C. Menon, “Moving toward automaticand standalone delineation of seismocardiogram signal,” in
IEEE Conf.EMBC , Aug 2015, pp. 7163–7166.[10] H. Nguyen, J. Zhang, and Y. H. Nam, “Timing detection and seismo-cardiography waveform extraction,” in
IEEE Conf. on Engineering inMedicine and Biology Society (EMBC) , Aug 2012, pp. 3553–3556.[11] M. J. Tadi et al. , “Seismocardiography: Toward heart rate variability(hrv) estimation,” in
IEEE Intr. Symposium on Medical Measurementsand Applications (MeMeA) , May 2015, pp. 261–266.[12] F. K.-Khavar et al. , “Automatic annotation of seismocardiogram withhigh-frequency precordial accelerations,”
IEEE J. Biomed. Health In-form. , vol. 19, no. 4, pp. 1428–1434, July 2015.[13] G. Shafiq, S. Tatinati, and K. C. Veluvolu, “Automatic annotation ofpeaks in seismocardiogram for systolic time intervals,”
IEEE Conf.EMBC , Aug 2016, pp. 2672–2675. [14] W. Xu et al. , “Detection of the seismocardiogram w complex based onmultiscale edges,” in
IEEE EMBC , vol. 3, 1996, pp. 1023–1024.[15] T. Choudhary, M.K. Bhuyan, and L.N. Sharma, “Orthogonal subspaceprojection based framework to extract heart cycles from SCG signal,”
Biomedical Signal Processing and Control , vol. 50, pp. 45 – 51, 2019.[16] A. Laurin, F. K.-Khavar, A. P. Blaber, and K. Tavakolian, “Accurate andconsistent automatic seismocardiogram annotation without concurrentECG,”
Physiol. Meas., Inst. Phys. , vol. 37, pp. 1588–1604, 2016.[17] K. Tavakolian et al. , “Estimation of hemodynamic parameters fromseismocardiogram,” in
IEEE Conf. CinC , Sept 2010, pp. 1055–1058.[18] T. Choudhary, L.N. Sharma, and M.K. Bhuyan, “Heart sound extrac-tion from sternal seismocardiographic signal,”
IEEE Signal ProcessingLetters , vol. 25, no. 4, pp. 482–486, 2018.[19] M. Paukkunen et al. , “Beat-by-beat quantification of cardiac cycle eventsdetected from three-dimensional precordial acceleration signals,”
IEEEJ. Biomed. Health Inform. , vol. 20, no. 2, pp. 435–439, 2016.[20] D. M. Salerno et al. , “Seismocardiographic changes associated withobstruction of coronary blood flow during balloon angioplasty,”
TheAmerican Journal of Cardiology , vol. 68, no. 2, pp. 201 – 207, 1991.[21] T. Choudhary, M. K. Bhuyan, and L. N. Sharma, “A novel method foraortic valve opening phase detection using SCG signal,”
IEEE SensorsJournal , 2019. doi: 10.1109/JSEN.2019.2944235[22] N. Mora, F. Cocconcelli, G. Matrella, and P. Ciampolini, “Fullyautomated annotation of seismocardiogram for noninvasive vital signmeasurements,”
IEEE Transactions on Instrumentation and Measure-ment , 2019. doi: 10.1109/TIM.2019.2908511[23] J. T. Bialasiewicz, D. Gonzlez, J. Balcells, and J. Gago, “Wavelet-based approach to evaluation of signal integrity,”
IEEE Transactions onIndustrial Electronics , vol. 60, no. 10, pp. 4590–4598, 2013.[24] T. Choudhary, M. K. Bhuyan, and L. N. Sharma, “Effect of respiratoryeffort levels on SCG signals,” in
IEEE Region 10 Symposium (TEN-SYMP) , 2019.[25] R. Hussein, K. B. Shaban, and A. H. El-Hag, “Wavelet transform withhistogram-based threshold estimation for online partial discharge signaldenoising,”
IEEE Transactions on Instrumentation and Measurement ,vol. 64, no. 12, pp. 3601–3614, 2015.[26] P. Addison, “Wavelet transforms and the ECG: a review,”
Physiologicalmeasurement , vol. 26, no. 5, pp. R155, 2005.[27] Y. Wang, C. J. Deepu, and Y. Lian, “A computationally efficient QRSdetection algorithm for wearable ECG sensors,” in
Proc. IEEE Intr. Conf.Engineering in Medicine and Biology Society , 2011, pp. 5641–5644.[28] C. Ye, B. V. K. Vijaya Kumar, and M. T. Coimbra, “Heartbeat clas-sification using morphological and dynamic features of ECG signals,”
IEEE Trans. Biomed. Eng. , vol. 59, no. 10, pp. 2930–2941, 2012.[29] A. Tharwat, “Classification assessment methods,”