Accelerometric Method for Cuffless Continuous Blood Pressure Measurement
aa r X i v : . [ ee ss . SP ] A ug Accelerometric Method for Cuffless ContinuousBlood Pressure Measurement
Mousumi Das, Tilendra Choudhary, L.N. Sharma, and M.K. Bhuyan
Abstract —Pulse transit time (PTT) has been widely used forcuffless blood pressure (BP) measurement. But, it requires morethan one cardiovascular signals involving more than one sensingdevice. In this paper, we propose a method for continuous cufflessblood pressure measurement with the help of left ventricularejection time (LVET). The LVET is estimated using a signal ob-tained through a micro-electromechanical system (MEMS)-basedaccelerometric sensor. The sensor acquires a seismocardiogram(SCG) signal at the chest surface, and the LVET informationis extracted. Both systolic blood pressure (SBP) and diastolicblood pressure (DBP) are estimated by calibrating the systemwith the original arterial blood pressure values of the subjects.The proposed method is evaluated using different quantitativemeasures on the signals collected from ten subjects under thesupine position. The performance of the proposed method is alsocompared with two earlier approaches, where PTT intervals areestimated from electrocardiogram (ECG)-photoplethysmogram(PPG) and SCG-PPG, respectively. The performance resultsclearly show that the proposed method is comparable with thestate-of-the-art methods. Also, the computed blood pressure iscompared with the original one, measured through a CNAPsystem. It gives the mean errors of the estimated systolic BP anddiastolic BP within the range of − ± − ± ± Index Terms —PTT, LVET, Pulse transit time, SCG, ECG,Cuffless blood pressure.
I. I
NTRODUCTION H IGH blood pressure is a major risk factor that leads toserious cardiovascular diseases (CVDs), such as hyper-tensive heart disease, stroke, and coronary artery disease [1].Therefore, effective and regular monitoring of blood pressure(BP) is necessary for early detection of hypertension. Thishelps clinical maintenance of the CVDs. To monitor thecontinuous BP in real time in an intensive care unit (ICU),the most common method is to insert a cannula needle ina suitable artery. For arterial blood pressure measurement, itis a gold standard method. Because of its invasive nature,it may introduce some other clinical complications, such asbleeding, infection, and ischemia. Also, it requires a standardclinical setting and expert. Among other modalities, mercurysphygmomanometer is considered as the most accurate non-invasive device. But, it requires an inflatable cuff that may
M. Das, T. Choudhary, L.N. Sharma, and M.K. Bhuyan are with theDepartment of Electronics and Electrical Engineering, Indian Institute ofTechnology Guwahati, India-781039 (E-mails: { mousumi18a, tilendra, lns,mkb } @iitg.ac.in). cause discomfort and pain to the user. Moreover, it onlyprovides the instantaneous blood pressure, and thus, cannotbe used for long term ambulatory BP monitoring. Arterialtonometry and arterial volume clamp are other primary tech-niques to obtain continuous BP in non-invasive way, but theyneed occlusive cuff [2]. All these devices have limitations forcontinuous and long term BP measurement. Therefore, thereis a need of advance technology that can provide continuousBP measurement.In recent years, research on cuffless BP measurement meth-ods for continuous BP monitoring has gained much attention.In literature, it is found that the pulse transit time (PTT) isemployed to estimate blood pressure. The PTT is defined asthe time taken for a cardiac pulse in artery to travel from aproximal point to a distal point. Usually, for the calculationof PTT interval, R-peak of electrocardiogram (ECG) andsystolic peak or onset of photoplethysmogram (PPG) are used[2]–[7]. Recently, some of the BP estimation methods haveintroduced the use of concurrent seismocardiogram (SCG) andPPG signals for the PTT estimation [8]–[10]. For this purpose,they suggested aortic valve opening (AO) peak of the SCG asa proximal point, and systolic peak or onset of the PPG asa distal point. All of these approaches require two correlatedcardiac pulse signals, either ECG-PPG or SCG-PPG. Acquir-ing more than one signal and estimating the location specificfeature points is computationally complex. It also reduces theflexibility of the system due to the involvement of differentsensing modalities. This type of complex system involvingtwo or more signals to measure the BP parameters may alsobe expensive. However, the SCG components of two differentorthogonal axes have also been investigated [11], where thecomponents are acquired from a single sensor. The methodrelies on time-interval from systole-profile onset in x-axisto diastole-profile onset in z-axis of SCG for systolic bloodpressure (SBP) estimation. However, determination of theseonset points is a crucial task. Additionally, achieving a goodaccuracy is still a problem. The research on BP estimationusing SCG signals is still far from maturity and has not yetadvanced to a stage where it can be successfully deployedas a medical user interface in consumer-grade applications.The proposed method is the first ever one of its kind whereBP is estimated using a single SCG component. Lookingat the above facts, the main aim of this proposed work isto estimate blood pressure using a single signal componentacquired through a micro-electromechanical system (MEMS)-based accelerometric device. A tri-axial accelerometer is usedto record the SCG signal. However, the proposed method doesnot rely on SCG components of x - and y -axis. Only the z - Fig. 1: Simultaneously recorded cardiological signals and their annotations: (a) Sensor placement on human body, (b) ECGsignal, (c) SCG signal, (d) PPG signal, and (e) ABP signal.axis component is used for processing. Seismocardigraphyis a non-invasive technique used to measure vibrations onthe chest wall induced by heartbeats and cardiac movements[12]. The SCG signal is used to determine the left ventricularejection time (LVET). Then, it computes the BP metrics usinga calibration procedure. Thus, we propose a low cost, lightweight, portable as well as user friendly device for continuouscuffless BP measurement instead of using multi-modal system.Also, recording process of SCG is simpler and providescomfortability to the user.II. SCG S
IGNAL AND ITS R ELATION TO
BPThe generation of chest-wall vibrations is mainly causeddue to change in volume, pressure and shape of the heartduring different stages of a cardiac cycle. The chest wallvibrational data can be non-invasively acquired by placinga MEMS-based tiny tri-axial accelerometer sensor on theprecordial area of the chest, generally at the lower end ofthe sternum on the xiphoid process [12]. The SCG capturesmechanical activities of the heart, such as opening and closingof valves, isovolumetric moment, isotonic contraction, rapidblood filling and its ejection. It is expected that an SCG signalcan give more clinically useful information as compared toother cardiac signals. In Fig. 1, the SCG signal is shown alongwith other concurrent ECG, PPG and arterial blood pressure(ABP) signals.During major left ventricular depolarization through purk-inje fibers, the mitral valve is closed, and in a standard ECG, the R-wave is seen for this phase (Fig. 1). After the occurrenceof R-peak in the electrocardiogram, further contraction ofmyocardium increases the pressure in the left ventricle. Thefirst heart sound, ‘ S1 ’, is produced due to closure of themitral and tricuspid valves. At this moment, aortic valve isnot opened. The ventricle works as a closed chamber. Thevolume remains the same, the pressure increases rapidly dueto rapid contraction. This phase is known as isovolumetriccontraction. The period until the opening of aortic valve fromthe depolarization of the left ventricle is known as pre-ejectionperiod (PEP). When the ventricular pressure reaches to theaortic pressure, the aortic valve opens (AO peak in SCG)and blood flows out of the aorta. At this instant, pressurepulse is initiated in aorta for systematic circulation througharteries. To calculate the PTT, the R-peak is considered as theproximal point in the literature. However, in reality, R-peakrepresents the depolarization of left ventricles, which does notimply the flow of the blood. If the R-peak is considered asproximal point, it includes the PEP interval and may resultan incorrect estimation of the PTT. Therefore, it is suggestedto consider aortic valve opening point as a proximal point atwhich the blood starts flowing. Once the ejection of bloodfrom ventricle is completed, the aortic valve is closed. Dueto closure of pulmonary and aortic valve, the second heartsound ‘ S2 ’ is produced. But, at this point, the mitral valve isnot opened and ventricle is in isovolumetric relaxation phase.Due to re-polarization of ventricles (T-wave in ECG), theventricular pressure falls down quickly. Because of this, the mitral valve is opened and pressure pulse in the arteries is diedout completely at AC point in SCG. During each heart cycle,the blood pressure pulse fluctuates between its maximum andminimum values. Thus, it is justified to compute the LVETusing AO-AC points of the SCG trace. This may resolve theexisting anomalies in various methods for the estimation ofPTT.Similar to the PTT, it is obvious that the LVET is a functionof pressure, and so, the pressure parameters can be estimatedfrom it. Within the time-period between AO and AC, theblood flows throughout the body, and thus, blood pressurepulse reaches peripheral sites. So, the occurrence of pulseat peripheral sites takes place only within the systole period.Thus, the end of the systolic period, i.e., AC can be consideredas the distal point. Once LVET is calculated, the BP can beeasily estimated through a calibration method.The paper is organized as follows: in Section III, we presentthe proposed cuffless and continuous BP measurement method.Section IV presents the experiments and the results, followedby conclusion in Section V.III. P
ROPOSED M ETHOD
During the systolic phase of the cardiac cycle, blood ispumped out from the heart. It exerts force on the wall ofarteries in systematic circulation and it is measured as bloodpressure (BP). In a normal subject, the systolic and diastolicpressure is clinically written as:
Systolic blood pressureDiastolic blood pressure =
SBPDBP = 120 mmHg80 mmHg (1)The computation of SBP and DBP is based on the principleof pulse propagation through major arteries. The systolicand diastolic blood pressure in a major artery are directlyproportional to the total cardiac output ( Q ) and total peripheralresistance ( Ω ). The Ω includes resistances of the arteries, arte-rioles, capillaries, venules, and veins in systematic circulation.The BP is given as: BP = Q × Ω (2)The cardiac output, Q , mainly depends on heart rate (HR) andstroke volume (SV). Therefore, it is expressed as: Q = HR × SV (3)The SV depends on end systolic volume ( SV l ) and enddiastolic volume ( SV m ) of a ventricle with a relation, SV = SV l − SV m (4)The LVET gives the systolic time intervals that correspondsthe interval between ejection and termination of the aortic flow.Thus, the LVET is directly influenced by the SV. The LVETis expressed as: LV ET = AC − AO (5)where, AO and AC represents aortic valve opening andclosing instants, respectively. At the AO instant, the bloodis ejected out of the heart, and therefore, can be considered asa proximal point for a pressure pulse. At the AC instant, the blood reaches the measurement sites and returns back fromthe whole body to the heart, and therefore can be consideredas a termination point of the pulse. This justifies the use ofLVET in place of PTT to calculate BP. Also, the LVET isinversely related to HR [13]. It is difficult to identify ACdue to large morphological variability of SCGs among thesubjects, and the SCG is susceptible to the distortions causeddue to body movements, respiration and other environmentalnoises [14]. The proposed model uses the timing-informationof pAC peak (peak just after AC point as shown in Fig. 1) forthe computation of LVET interval. With this modification, anapproximation of LVET, say LVET ′ , is used for modelling theBP in the proposed method. The LVET ′ is computed as: LV ET ′ = pAC − AO (6)By utilising the logarithmic LVET ′ and HR information, theBP is modelled as: BP = a · ln ( LV ET ′ ) + b · HR + c (7)where, a , b , and c are used to parametrize the proposed model.These parameters are subject-specific and are estimated viamodelling in a least square sense. The HR is calculated fromthe time interval between consecutive AO peaks. The proposedmodel can be represented in a vector-matrix form as: BP BP ...BP n = ln ( LV ET ′ ) HR ln ( LV ET ′ ) HR . . .. . .. . .ln ( LV ET ′ n ) HR n abc (8)where, n is the total number of beats, the BP vector withdimension n × , represents the original arterial blood pressureand it is used for training the model for both SBP and theDBP estimation. The LV ET ′ and HR are the parametersthat are estimated simultaneously for each beat along with theoriginal BP and represented by a n × matrix . The unknowncoefficients a , b and c are represented by a × vector.The unknown vector is estimated using linear regression byapplying pseudo inverse as: X = ( A T .A ) − .A T .BP (9)where, X is the unknown coefficient vector and A is the matrixthat carries the information related to the estimated LV ET ′ and HR values. The coefficients are estimated for each indi-vidual for both SBP and DBP, separately. Fig. 2 shows a blockdiagram of the proposed cuffless BP estimation algorithm. Ithas training and testing phases. The training phase is requiredonly during initialization where the coefficients a , b and c arecomputed using the regression method. Then, the testing phasebecomes independent of BP parameters for a personalizedmeasurement. The steps involved in the proposed method arediscussed in the following subsections. A. AO Peak Detection
Initially, the AO peaks in the SCG are detected using ourearlier proposed VMD-based scheme [15]. The pseudocode is
Fig. 2: Block diagram of the proposed BP estimation method.
Algorithm 1
AO instant detection in an SCG signal
Input:
SCG signal x [ n ]; n = 1 , , ....., N Low-Frequency artifacts removal
1: Decompose signal x [ n ] using MVMD [ m Ii , ω Ii ] = VMD ( x, α , K ) ; i = 1 , , ...K // where m Ii := decomposed i th mode of x [ n ] , ω Ii := central frequency of m Ii , α := allowing mode-bandwidth, K := number of decomposed modes2: Extract the detrended SCG signal x d [ n ] = x [ n ] − P i m Ii [ n ] Systolic Profile Enhancement
3: Resolve useful signal components using MVMD [ m IIj , ω
IIj ] = VMD ( x d , α , K ) ; j = 1 , , ...K
4: Extract Gaussian derivavtive filtered modes (GDFMs) g j [ n ] = d [ m ] ⊛ m IIj [ n ] ; m = 1 , , ...L − // where Gaussian derivative kernel d [ m ] := g [ m + 1] − g [ m ] ,Gaussian window g [ l ] := exp (cid:18) − σ h l ( L − / i (cid:19) ; 0 ≤| l |≤ ( L −
5: Compute relative GDFM energy (RGE) for mode selectionRGE: e j = E g j P j E g j ; j = 1 , ....K Reconstructed signal: s [ n ] = e j ∗ g j ∗ + P e j ∗ − g j ∗ − + Q e j ∗ +1 g j ∗ +1 // where e j ∗ denotes maximum RGE and j ∗ is corresponding mode indexP=1, Q=0 ← if | e j ∗ − e j ∗ − i | < ρ ; ρ ∈ [0 , P=0, Q=1 ← if | e j ∗ − e j ∗ + i | < ρ ; P=0, Q=0 ← otherwise AO peak approximation
6: Construct envelope on systolic profilesDifference envelope: D env = U env − L env // where U env , L env : upper and lower envelopes of s[n], respectivelyThresholding: T env [ n ] = ( D env [ n ] > τ ) × D env [ n ]; τ >
7: Find zero-crossing locationsApproximated AO peaks: a [ v ] ← Positive zero crossings of b T env [ n ] // where b T env corresponds Hilbert transform of T env ; and v = 1 , , ..., V
8: Estimate True AO peaks using cardiac cycle envelope (CCE)Construct CCE using absolute operator and triple integration on GDFMsTrue AO peaks: e a [ k ] ⇐ = a [ v ] elements lying within the range Q // where Q ∈ ( CCE peak k , CCE peak k + 350 ms ) ; k = 1 , , ....., K Output:
Detected AO peak locations e a [ k ] given in Algorithm 1. Although the method provides a robustdetection of the AO peaks, it produces misdetections and false-detections in some critical cases. To address this issue, a post-processing scheme is proposed. Systole profile Diastole profile (a) (b) (c)
True AO peak locations Detected AO peaks
Fig. 3: SCG envelopes showing false alarming conditions inAO peak detection. (a) Case 1: Missing AO peak, (b) case 2:false detection on systole profile, and (c) case 3: false detectionon diastole profile
B. Post-processing for AO Peak Detection
A post-processing technique is developed and employed inthis algorithm (Algorithm 1) to improve the performance of theAO peak detection by reducing the false alarming conditions.Let, pks = [ p , p , · · · , p n ] is a vector of detected AOpeaks in a SCG segment. At first, the difference of eachconsecutive AO peak instants is calculated as: d i = p i +1 − p i , (10)where, i ( i = 1 , , · · · , n − ) denotes the peak index. Then,median-difference of these AO instants is computed as areference, say M . The error between d i and M is comparedwith a threshold (say T H ). Usually, the following three falsealarming situations are encountered in the AO peak detectionprocess. These cases are shown through the SCG envelopes inFig. 3.1)
Missing AO peaks:
This false negative condition isidentified if , ( d i − M ) > × T H , and it can be avoidedby determining the maxima between p i + τ and p i +1 − τ in the SCG, where τ = 60 ms.2) False detection on systole profile:
An AO peak detectedin the systole profile is considered as false positive, if ( d i − M ) < − . × M , and the condition can be avoidedby discarding p i +1 peak.3) False detection on diastole profile:
After taking careof the first two conditions, the pks and d i vectors areupdated. Subsequently, the method looks for the falsepositive in the diastole profile by checking the condition, [( d i − M ) < − T H and ( d i +1 − M ) < − T H ] , and ifthe condition is satisfied, p i +1 peak is removed.The criterion mentioned above were designed with empiricalanalysis. In this way, an AO peak correction strategy is adopted Fig. 4: Experimental setup for signal acquisition.in our proposed method. After that, pAC peaks on the SCGsignal are detected to measure the LVET ′ interval. C. pAC Peak Detection
The pAC peaks are detected using a technique given in [14].At first, it divides each of the consecutive AO-AO intervals attheir midpoint ( M ), which is expressed as: M = p i + p i +1 (11)where, p i and p i +1 are consecutive AO peaks from the updated pks vector. In a similar manner, the interval between p i and M is again divided to obtain the pAC as: M = p i + M p i + p i +1 (12)Now, the SCG signal is segmented from M to M foreach of the cycles. Subsequently, the DC-offset is removed andamplitude normalization is performed. Subsequently, movingaverage filtering is applied to reduce irregularities. To detectall pAC points, the trends in the segments are removed byapplying a Butterworth high pass IIR filter with an empiricalcut-off frequency of Hz. Finally, the pAC peaks are detectedas maximas of the processed SCG segments. After annotatingAO and pAC peaks, HR and LVET ′ interval for each of thebeat are calculated. Using these LVET ′ and HR values for eachindividual subject, the system is calibrated with the originalSBPs and DBPs using a linear regression model. Hence, modelcoefficients are obtained for both SBPs and DBPs, separately.The coefficients obtained for a particular subject are finallyused to estimate SBP and DBP values.IV. E XPERIMENTS AND R ESULTS
Fig. 4 shows the experimental setup for simultaneousrecording of SCG, ECG, PPG and ABP signals. Data wasrecorded for minutes from 10 healthy subjects in supineposition. For all the subjects, recordings were done for ECGin Lead-II configuration, z-axis signal of SCG using tri-axial accelerometer, PPG at fingertip, and ABP using finger-cuff provided by CNAP R (cid:13) Monitor 500. All signals are syn-chronously obtained using MP150 DAQ system (BIOPAC TABLE I: BP and HR ranges in the database
Min Max STD MeanDBP (mmHg)
SBP (mmHg)
HR (bpm)
TABLE II: Comparison of estimated BPs using our proposedLVET ′ -based method with reference ABP measurements Subjects DBP (mmHg) SBP (mmHg)ME MAE STD ME MAE STD ≈ . Average -1.29 2.6 2.6 -0.19 3.2 3.3
Systems, Inc.) at a sampling rate of 1 kHz. Data acquisition isperformed at Electro Medical and Speech Technology (EMST)Laboratory, IIT Guwahati with prior consent of the subjectsafter clearance from institutional ethics review board. For eachsubject, of the data is used for calibration, while theremaining is used for the testing purpose. The ABP signal wasused for calibration of the SBP and DBP of an individual. It isused as a benchmark to test and validate the proposed system.Table I shows statistical distribution in terms of minimum(Min), maximum (Max), standard deviation (STD) and meanvalues of the BP and HR for in-house recorded database.The Min and Max SBP values are . mmHg and . mmHg, whereas for DBP, Min and Max values are . mmHg and . mmHg, respectively. Min recorded HR is . bpm, while max recorded HR in the in-house data is . bpm. For each individual subject, of the datais used for calibration and the algorithm is evaluated onthe remaining data. Mean error (ME), mean absolute error(MAE) and standard deviation (STD) are used to comparethe performance with the original values. BP is estimatedusing the proposed LVET ′ -based approach. The estimated SBPand DBP values are compared with ABP measurements. Theperformance of the proposed approach is evaluated in termsof ME, MAE and STD measures, and the results are shown inTable II. For the DBP measurement, the minimum ME andMAE values are − . and . , respectively for subject 5with a STD value of . . Similarly, for the SBP measurement,minimum ME is closed to zero for subject 4 with a STD value . . The minimum MAE value is . for subject 5 with aSTD value . . For all the subjects, the average ME with STDof the estimated DBP and SBP are − . ± . mmHg and − . ± . mmHg, respectively, whereas the average MAEare, . and . mmHg, respectively. The obtained resultssuggest that the accuracy of BP estimations using the proposedsystem satisfies the requirements of the IEEE standard ± mmHg [3], [16]. TABLE III: Statistical comparison of the proposed LVET ′ -based method with the PTT and PTT1 based methods for DBP andSBP measurements Subjects DBP (mmHg) SBP (mmHg)ME
PTT ME PTT1
MAE
PTT
MAE
PTT1
STD
PTT
STD
PTT1 r1 r2 ME
PTT ME PTT1
MAE
PTT
MAE
PTT1
STD
PTT
STD
PTT1 r1 r2
Average DB P ( mm H g ) PTT- based approachLVET- based approachPTT1- based approach0 20 40 60 80 100 120
Time (Sec) S B P ( mm H g ) PTT- based approachLVET- based approachPTT1- based approach (a) DBP estimation(b) SBP estimation
Fig. 5: Comparison of the proposed LVET ′ based method withPTT and PTT1 based methods for estimated DBP and SBPvalues.The estimated BP values are also compared with the BPvalues computed through earlier PTT- and PTT1-based ap-proaches. In PTT-based approach, the duration of PTT is cal-culated using the combination of ‘ECG-PPG’ [2]–[7], whereasin the PTT1-based approach, ‘SCG-PPG’ signal pair is used[8]–[10]. Table III shows the comparison of the proposedmethod with the existing methods. The average ME and MAEbetween the proposed LVET ′ -based and PTT-based methodsare found as . ± . mmHg and . mmHg for DBP,and . ± . mmHg and . mmHg for SBP measurement,respectively. While, for the PTT1-based method, these metricsare observed as − . ± . mmHg and . mmHg for DBPmeasurement. While for SBP, they are found as . ± . mmHg and . mmHg, respectively. Additionally, the averagecorrelation of the BPs estimated from the proposed methodwith that of the earlier methods is also computed. For the PTT-based method, the average correlation coefficients, denotedby r , are obtained as 0.6 and 0.5 for the DBP- and SBP-estimation, respectively. Similarly, for PTT1-based method,the correlation coefficient ( r ) values are observed as 0.7and 0.5 for the DBP and the SBP, respectively. The goodcorrelation exhibited by the proposed method clearly showsthat it is comparable with the existing methods. Therefore, asingle sensor based proposed method may replace the use ofmulti-modal system for the BP estimation.Further, to show the performance and its comparison withthe existing methods, the BP traces estimated from a singlesubject ( x -axis represents the average of the BP values obtained bythe proposed and the existing method, whereas the differencebetween both the BP values is shown in the y -axis. It isdesirable to have a low value of bias or mean difference.This result indicates that the proposed method can give similarmeasurement values as that of the method used for comparisonof performance. As shown by the bold line in Fig. 6, themean differences of the estimated DBPs and SBPs betweenthe proposed and the PTT-based methods are . ± . mmHg and . ± . mmHg, respectively. Whereas, theyare found to be − . ± . mmHg and − . ± . mmHg, respectively for the PTT1-based method. It can beobserved that the majority of the points lie within the limitsof agreement, i.e., m ± . s , where m and s denote the meanand standard deviation of the BP difference, respectively. Thelimits of agreement are shown by dashed lines in the figure.This shows the suitability of our method as a substitute of theexisting methods.In addition, the performance comparison is also shown withthe help of regression plots in Fig. 7. The correspondingcorrelation coefficients between the proposed LVET ′ -basedand the PTT based methods are approximately 0.7 and 0.5 forthe DBP and SBP estimation, respectively, as shown in Fig.7(a) and (b). Whereas, in case of the PTT1-based method, itis found 0.9 for both the DBPs and SBPs as shown in Fig.7(c) and (d). V. C ONCLUSION
In this work, a method is proposed for noninvasive andcuffless long-term BP measurement using an accelerometricSCG signal. Initially, the AO and pAC peaks are detectedfor the estimation of LVET ′ and HR parameters. With theseparameters, the BP modeling is performed for the SBPs andthe DBPs using individual calibration. The existing methodsrequire a PPG signal along with either ECG or SCG signalsto perform this task. However, the use of bio-electrodes for Fig. 6: Bland-Altman plot to compare the performance of the proposed method with PTT- and PTT1-based methods. Panels(a) and (b) represent comparison with PTT-based method for DBPs and SBPs, respectively, and panels (c) and (d) representcomparison with PTT1-based method for DBPs and SBPs, respectively.Fig. 7: Regression plots of estimated BP measurements for a single subject. (a) and (c) show the regression of DBPs estimatedby the proposed method with that of the PTT- and PTT1-based methods, respectively, (b) and (d) show the regression of SBPsestimated by the proposed method with that of the PTT- and PTT1-based methods, respectively. the ECG and the photo-detector for PPG acquisition has beenavoided in the proposed method by using only SCG data. Thismakes it more flexible and offers mobility to the user. Also,the light-weight of the sensing device and its simple operationmake it versatile to be operated independently without anyhelp of an expert. With in-house recordings, quite satisfactorycorrelation results between the proposed method and theexisting methods are obtained. The performance results clearlyshow that our method is able to achieve the performanceup to the level obtained by the state-of-the-art methods onlywith a single cardiac sensor. In this research, the proposedmethod is tested and validated on healthy individuals under anormal condition, and hence, the performance could be eval-uated on different physiological and pathological conditionsin the future works. However, the quantitative results showthe potential of the proposed method to replace the existingmethods for continuous long-term BP monitoring.R
EFERENCES[1] Prospective Studies Collaboration, “Age-specific relevance of usualblood pressure to vascular mortality: a meta-analysis of individual datafor one million adults in 61 prospective studies,”
The Lancet , vol. 360,no. 9349, pp. 1903–1913, 2002.[2] X. Ding and Y. T. Zhang, “Pulse transit time technique for cufflessunobtrusive blood pressure measurement: from theory to algorithm,”
Biomedical Engineering Letters , vol. 9, no. 1, pp. 37–52, 2019.[3] Z. Tang et al ., “A chair-based unobtrusive cuffless blood pressure moni-toring system based on pulse arrival time,”
IEEE Journal of Biomedicaland Health Informatics , vol. 21, no. 5, pp. 1194–1205, 2016.[4] A. Esmaili, M. Kachuee, and M. Shabany, “Nonlinear cuffless bloodpressure estimation of healthy subjects using pulse transit time andarrival time,”
IEEE Transactions on Instrumentation and Measurement ,vol 66, no. 12, pp. 3299–3308, 2017.[5] X. R. Ding et al ., “Continuous cuffless blood pressure estimationusing pulse transit time and photoplethysmogram intensity ratio,”
IEEETransactions on Biomedical Engineering , vol. 63, no. 5, pp. 964–972,2015.[6] Y. Y. Hsieh et al ., “A linear regression model with dynamic pulse transittime features for noninvasive blood pressure prediction,” in
Proc. IEEEIntr. Conf. on Biomedical Circuits and Systems (BioCAS) , Oct. 2017,pp. 604–607.[7] R. Mukkamala et al ., “Toward ubiquitous blood pressure monitoring viapulse transit time: theory and practice,”
IEEE Transactions on BiomedicalEngineering , vol. 62, no. 8, pp. 1879–1901, 2015.[8] K. Soerensen et al ., “Challenges in using seismocardiography for bloodpressure monitoring,” in
IEEE Intr. Conf. on Computing in Cardiology(CinC) , pp. 1–4, Sep. 2017.[9] C. Yang, and N. Tavassolian, “Pulse transit time measurement us-ing seismocardiogram, photoplethysmogram, and acoustic recordings:Evaluation and comparison,”
IEEE Journal of Biomedical and HealthInformatics , vol. 22, no. 3, pp.733–740, 2017.[10] C. S. Kim et al ., “Ballistocardiogram as proximal timing reference forpulse transit time measurement: Potential for cuffless blood pressuremonitoring,”
IEEE Transactions on Biomedical Engineering , vol. 62,no. 11, pp. 2657–2664, 2015.[11] M. S. Imtiaz et al ., “Correlation between seismocardiogram and systolicblood pressure,” in
Proc. IEEE Conf. Electrical and Computer Engineer-ing (CCECE) , May 2013, pp. 1–4.[12] 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, 2018.[13] A. M. Weissler , R. G. Peeler , W. H. Roehll Jr, “ Relationships betweenleft ventricular ejection time, stroke volume, and heart rate in normalindividuals and patients with cardiovascular disease,”
American heartjournal , vol. 62, no. 3, pp. 367–78, Sep. 1961.[14] 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, Elsevier , vol. 50, pp. 45–51,2019. [15] T. Choudhary, M. K. Bhuyan, and L. N. Sharma, “A Novel Methodfor Aortic Valve Opening Phase Detection Using SCG Signal,”
IEEESensors Journal , vol. 20, no. 2, pp. 899–908, 2019.[16] IEEE Standard Association, “IEEE standard for wearable cuffless bloodpressure measuring devices,”