Abstract

Estimating central aortic blood pressure (BP) is important for cardiovascular (CV) health and risk prediction purposes. CV system is a multichannel dynamical system that yields multiple BPs at various body sites in response to central aortic BP. This paper concerns the development and analysis of an observer-based approach to deconvolution of unknown input in a class of coprime multichannel systems applicable to noninvasive estimation of central aortic BP. A multichannel system yields multiple outputs in response to a common input. Hence, the relationship between any pair of two outputs constitutes a hypothetical input–output system with unknown input embedded as a state. The central idea underlying our approach is to derive the unknown input by designing an observer for the hypothetical input–output system. In this paper, we developed an unknown input observer (UIO) for input deconvolution in coprime multichannel systems. We provided a universal design algorithm as well as meaningful physical insights and inherent performance limitations associated with the algorithm. The validity and potential of our approach were illustrated using a case study of estimating central aortic BP waveform from two noninvasively acquired peripheral arterial pulse waveforms. The UIO could reduce the root-mean-squared error (RMSE) associated with the central aortic BP by up to 27.5% and 28.8% against conventional inverse filtering (IF) and peripheral arterial pulse scaling techniques.

1 Introduction

Cardiovascular (CV) disease is a leading cause of mortality and morbidity in the U.S. and globally [1]. Brachial blood pressure (BP) is widely used in today's clinical practice to probe CV health and disease. But, central aortic BP measured near the heart has been suggested as superior signature of CV health and disease to conventional brachial BP [25]. However, direct measurement of central aortic BP involves invasive and/or inconvenient procedures and trained operators (e.g., cardiac catheterization [2,68] and carotid artery tonometry [9,10]), hampering its widespread use. Early efforts to overcome this obstacle was to construct a population-based mathematical transformation, called the generalized transfer function (GTF) [11], which converts a peripheral (e.g., brachial [6,12,13] or radial [1416]) arterial pulse waveform into central aortic BP waveform. The GTF is not patient-specific by nature, while the characteristics associated with the propagation of BP waves in the arteries exhibit a large degree of inter- and intra-individual variability [17]. Hence, the efficacy of the GTF technique has been controversial [18].

More recently, attempts have been made to develop patient-specific techniques for estimating central aortic BP waveform from peripheral BP waveform measurements. The vast majority of these techniques are built upon the so-called blind system identification methodology [19]. The methodology, when applied to coprime multichannel systems with unknown input signal, determines the channel dynamics and then deconvolves the input signal by exploiting the correlation relationship between the channels. In particular, the input-deconvolution step of the methodology has employed various techniques, such as direct inverse filtering (IF) [7,20], least-squares and maximum-likelihood-type deconvolution [21,22], and the design of dedicated deconvolution filters [23,24]. Direct inverse filtering has been a straightforward choice due to the nonminimum phase nature of the channel dynamics associated with the BP wave propagation in the arteries [25]. The least-squares and maximum likelihood-type deconvolution techniques were developed primarily for finite impulse response channel dynamics [26,27]. Hence, central aortic BP deconvolution based on these techniques involved the finite impulse response filter approximation of the BP propagation channel dynamics [21,22]. To relax such restrictions, design methodologies for the input deconvolution filters applicable to coprime multichannel systems with infinite impulse response channel dynamics have been developed [23,24,28]. However, all these techniques have a common critical weakness: the integrity of input deconvolution hinges upon the accuracy of the channel dynamics due to the open-loop nature of these techniques. Since the channel dynamics themselves are either population-based or estimated by the blind system identification methodology, the quality of the deconvolved input signal is impacted by the errors associated with the channel dynamics. In the context of estimating central aortic BP from peripheral arterial pulses, these errors ultimately boil down to errors in the clinically important characteristics in the central aortic BP waveform (e.g., systolic (SP) and pulse (PP) pressures), which can subsequently deteriorate its credibility for CV health and risk assessment.

In our endeavor to develop a novel input deconvolution technique equipped with robustness against the channel dynamics inaccuracy, we investigated the application of established state estimation techniques (i.e., the observers [29]) to input deconvolution in coprime multichannel systems. The central idea underlying this technique is to derive the unknown input signal by designing an observer for a hypothetical input–output system constructed using a pair of output signals in the multichannel system. This “closed-loop” input deconvolution idea was inspired by two key observations: (i) that the relationship between any pair of two outputs constitutes a hypothetical input–output system with unknown input embedded as an estimable state; and (ii) that the observer can compensate for the adverse impact of the channel dynamics inaccuracy on the integrity of the deconvolved input signal by way of its corrective error feedback. In this paper, we developed and analyzed an unknown input observer (UIO) for input deconvolution in coprime multichannel systems. In particular, we provided a universal design algorithm as well as meaningful physical insights and inherent performance limitations associated with the algorithm. Then, we illustrated its validity and potential using the clinically significant case study of estimating central aortic BP waveform from two noninvasively acquired peripheral arterial pulse waveforms.

This paper is organized as follows. In Sec. 2, a UIO for closed-loop input deconvolution in coprime multichannel systems is developed and analyzed. In Sec. 3, the application of the UIO to the estimation of central aortic BP waveform from noninvasive peripheral arterial pulse waveforms is described in detail. In Sec. 4, the results are presented and discussed. In Sec. 5, the paper is concluded with future work.

2 Unknown Input Observer Design for Coprime Multichannel Linear Dynamical Systems

2.1 Problem Formulation.

Consider a multichannel linear dynamical system in which a common yet unknown input signal generates multiple output signals (Fig. 1). It is assumed that the channels are coprime, i.e., they do not share common poles and/or zeros. The goal is to reconstruct the unknown input based on the output signals at the multiple channels. The basic idea of our approach is (i) to transform the multichannel dynamics into an equivalent input–output dynamics in which the unknown input is cast into an internal state variable and then (ii) to design an observer that can estimate the unknown input.

Fig. 1
A coprime multichannel linear dynamical system in which a common yet unknown input signal u(z) generates multiple output signals y1(z)…yN(z)
Fig. 1
A coprime multichannel linear dynamical system in which a common yet unknown input signal u(z) generates multiple output signals y1(z)…yN(z)
Close modal

2.2 Unknown Input Observer Design.

Without any loss of generality, the UIO design problem can be solved for a two-channel system case. The results can be readily generalized and expanded to systems with >2 channels. Indeed, in systems with >2 channels, any pair of two channels in the system can be selected to design an UIO. Consider the following transfer functions associated with a two-channel system:
(1)

where u is the common input, y1 and y2 are outputs, G1(z) and G2(z) are the channel transfer functions associated with y1 and y2, and N1(z), D1(z), N2(z), and D2(z) are polynomials in the discrete-time shift operator z of degrees m1, n1, m2, and n2, respectively. It is assumed that D1(z) and D2(z) are monic. It is further assumed that m1=n1, i.e., G1(z) is a proper transfer function, and that G2(z) does not have any zeros at the origin. Note that the properness of G1(z) is not a restriction, because any strictly proper G1(z) can be made proper by padding z's to its numerator polynomial N1(z): N¯1(z)=zr1N1(z), where r1=n1m1 is the relative degree of G1(z). In this way, an equivalent input–output relationship y¯1(z)=G¯1(z)u(z)=N¯1(z)D1(z)u(z) with the forward-shifted output y¯1(z)=y1(z)zr1 can be constructed. The zeros at the origin in G2(z) can be likewise be removed by backward-shifting y2(z). Then, the unknown input u(z) may be reconstructed from y1(z) and y2(z) as follows:

Theorem 1. Consider the two-channel system in Eq. (1) withG1(z)andG2(z)given by:
(2)

which satisfy the following properties:

  1. The transfer functions G1(z) and G2(z) are coprime

  2. The polynomials D1(z) and D2(z) are monic

  3. The transfer function G1(z) is proper and minimum phase

  4. The transfer function G2(z) does not have any zero at the origin

Let {A1,B1,C1,D1} with D1=bn1(1) and {A2,B2,C2,D2} with D2=0 be the controllable canonical form realizations of G1(z) and G2(z), respectively. Then, the unknown input u of this coprime multichannel system can be reconstructed by the following UIO:
(3)

where x(k)=[x1(k)x2(k)] with x1 and x2 the state vectors of dimension n1 and n2, A¯=[A1B1C1/bn1(1)0n1×n2B2C1/bn1(1)A2], B¯=1bn1(1)[B1B2], C¯=[01×n1C2], Cˇ=[C101×n2], and L¯ is a UIO gain matrix of dimension (n1+n2)×1 with which all the eigenvalues of A¯L¯C¯ are placed in the unit circle.

Proof and Analysis. The state space realization of the multichannel system in Eq. (1) is given by:
(4)
Given that u is unknown but can be represented as a function of x1 and y1: u(k)=1bn1(1)[y1(k)C1x1(k)], Eq. (4) can be rewritten as follows:
(5)

Hence, the UIO in Eq. (3) can reconstruct u from y1 and y2 if the pair (A¯,C¯) is observable. Noting that both {A1,B1,C1,D1} and {A2,B2,C2,D2} are associated with controllable canonical form, we have:

 
(6b)

Then, it can be easily shown that {A3,B3,C3,D3}={A1B1C1/bn1(1),B1/bn1(1),C1/bn1(1),1/D1} is the controllable canonical form realization of G11(z). Indeed, using Eq. (6a):

 
(7b)

Equation (7) yields C3(zIA3)1B3+D3=zn1+an11(1)zn11++a0(1)bn1(1)zn1+bn11(1)zn11++b0(1)=G11(z). Thus, A¯ can be rewritten as A¯=[A30n1×n2B2C3A2]. Computing the observability matrix of the pair (A¯,C¯) yields:
(8)
where r2=n2m2 is the relative degree of G2(z) and r2+1kn1+n2. Since the pair (A2,C2) is observable, O2 has full column rank, i.e., rank(O2)=n2. Further, O1 reduces to the following after Gaussian elimination:
(9)

which, since the pair (A3,C3) is observable and n1+n2r21n1, guarantees that O1 has full column rank, i.e., rank(O1)=n1. Thus, O¯ is a full rank matrix, meaning that the pair (A¯,C¯) is observable and the existence of the UIO in Eq. (3) is guaranteed.

Note that the rationale for the minimum phase requirement imposed on G1(z) is now obvious: according to Schur's determinant identity, the eigenvalues of A¯ are given by those of A2 and A3:
(10)

In addition, the eigenvalues of A3 correspond to the poles of G11(z), i.e., the zeros of G1(z).▪

2.3 Intuitive Interpretation and Limiting Behavior.

Intuitively, the UIO in Eq. (3) is equivalent to a Luenberger observer designed for the system with y1 and y2 as input and output and u as an internal state, i.e., y2(z)=G2(z)G11(z)y1(z) (Fig. 2). Indeed, since G1(z) and G2(z) are coprime, a state space realization of the system is given by:
(11a)
(11b)

where {A2,B2,C2,D2} and {A3,B3,C3,D3} are the controllable canonical form realization of G2(z) and G11(z) as defined in Eqs. (6) and (7). Serially concatenating Eqs. (11a) and (11b) yields:

12
(12a)
(12b)
which is identical to Eq. (5). Thus, the UIO in Eq. (3) for the plant in Eq. (5) is identical to the Luenberger observer for the plant in Eq. (12).

Fig. 2
A hypothetical input–output system derived from a coprime multichannel system by designating its one output signal as input to the hypothetical input–output system and its another output signal as output of the same system
Fig. 2
A hypothetical input–output system derived from a coprime multichannel system by designating its one output signal as input to the hypothetical input–output system and its another output signal as output of the same system
Close modal
The above intuitive insight streamlines the analysis of the limiting behavior of the UIO with respect to the magnitude of its gain L¯. On the one hand, if L¯0, Eq. (3) dictates that û(z)=G11(z)y1(z); indeed, when L¯0
(13)
which is identical to Eq. (11a). Hence, û is essentially given by inverse filtering of y1(k) with G1(z). On the other hand, if L¯, Eq. (3) dictates that û(z)=G21(z)y2(z); indeed, when L¯ is very large:
(14)
In addition, very large L¯ enforces ŷ2(k)=C2x̂2(k)=y2(k), which, in conjunction with Eq. (14), yields:
(15)

which reduces to ŷ2(z)G2(z)û(z), i.e., û is essentially given by inverse filtering of y2(k) with G2(z).

The above limiting behavior of the UIO provides an important insight into the fundamental limitation in its performance when applied to coprime multichannel systems: the UIO does not provide any improvement in the accuracy of the estimated unknown input beyond open-loop inverse filtering when L¯0 and L¯. Indeed, the accuracy of the estimated unknown input approaches to that associated with the open-loop inverse filtering of y1 with G1(z) if the gain is chosen to be too small (see Eq. (13)), whereas it approaches to that associated with the open-loop inverse filtering of y2 with G2(z) if the gain is chosen to be too large (see Eq. (15)). This unique limiting behavior can be attributed to the coprimeness of the channel dynamics. The plant in Eq. (5) (or equivalently, Eq. (12); also see Fig. 2) computes u from y1 (u(z)=G11(z)y1(z)), and then computes y2 from u thus computed (y2(z)=G2(z)u(z)). Since the UIO reduces to the open-loop plant dynamics (i.e., Eq. (5)) when L¯=0, and u can be estimated solely from y1, û(z)=G11(z)y1(z). On the other hand, when L¯=, the corrective error feedback acts to achieve ỹ2y2ŷ2=0. In other words, u computed by the plant dynamics as u(z)=G11(z)y1(z) is corrected to yield û to achieve y2(z)=G2(z)û(z). Thus, in this case u can be estimated solely from y2, û(z)=G21(z)y2(z). Given that the notion of how small or large L¯ is depends on the channel dynamics, the appropriate range of L¯ (i.e., the range of L¯ that is not too small and not too large) also depends on the problem at hand.

An important practical implication of the above limiting behavior associated with the UIO is that iterative trial and error process may be required to design a UIO whose performance is superior to simple open-loop inverse filtering: in contrast to conventional observer design problems (in which corrective error feedback action can be strengthened by simply increasing the observer gain to achieve a larger improvement in closed-loop estimation relative to its open-loop counterpart), the gain of the UIO designed for coprime multichannel systems must be carefully chosen (i.e., it must not be too small or too large) to maximize the benefit of the corrective error feedback action.

3 Observer-Based Deconvolution of Central Aortic Blood Pressure Waveform From Noninvasive Peripheral Arterial Pulse Waveforms

The validity, strengths, and limitations of the input deconvolution technique based on the UIO developed in this work was examined in a clinically significant real-world problem of estimating central aortic BP waveform from two noninvasively acquired peripheral arterial pulse waveform measurements (Fig. 3). Central aortic BP and pulse volume recordings (PVRs) at arm and leg sites were measured simultaneously using BP cuffs loaded at a subdiastolic pressure level. The multichannel artery dynamics associated with the central aortic BP-brachial PVR channel and the central aortic BP-leg PVR channel were estimated using the PVR signals [20]. Then, central aortic BP was estimated using the UIO designed based on the channel dynamics. Details follow.

Fig. 3
Observer-based deconvolution of central aortic BP waveform from noninvasive peripheral arterial pulse waveform measurements. The PVRs are made at the upper arm and leg sites using the BP cuffs loaded at a subdiastolic pressure level. These diametric PVR signals are applied to a system identification procedure [20] to derive the channel dynamics associated with (i) the propagation of the BP wave from the aorta to the peripheral arteries and (ii) the propagation and distortion of the peripheral BP waves into the PVR signals at the respective peripheral measurement sites. The UIO designed using the estimated channel dynamics estimates central aortic BP from peripheral PVR signals.
Fig. 3
Observer-based deconvolution of central aortic BP waveform from noninvasive peripheral arterial pulse waveform measurements. The PVRs are made at the upper arm and leg sites using the BP cuffs loaded at a subdiastolic pressure level. These diametric PVR signals are applied to a system identification procedure [20] to derive the channel dynamics associated with (i) the propagation of the BP wave from the aorta to the peripheral arteries and (ii) the propagation and distortion of the peripheral BP waves into the PVR signals at the respective peripheral measurement sites. The UIO designed using the estimated channel dynamics estimates central aortic BP from peripheral PVR signals.
Close modal

3.1 Experimental Data.

The experimental central aortic BP and peripheral PVR signals were simultaneously recorded from ten cardiac catheterization patients at the University of Maryland Medical Center under its IRB approval and written informed consent [20]. In each patient, the central aortic BP waveform was invasively measured with a catheter inserted through a femoral artery per routine standard of care, while the PVR waveforms were noninvasively measured at an upper arm and an upper leg, with occlusive BP cuffs loaded at a subdiastolic pressure level. All the signals were stored in a laptop computer through a data acquisition system at a sampling frequency of 1 kHz for approximately 1 min. The sensors and the data acquisition system were calibrated before collecting data from each patient to ensure the accuracy of the measurements. The two PVR signals were used to estimate the plant dynamics as well as to design the UIO, while the central aortic BP was used to evaluate the performance of the UIO. Table 1 shows the range of central aortic SP, PP, and mean pressures (MPs) associated with the ten patients considered in this work. Figure 4 presents representative measurements of (a) PVR signals at an arm and a leg and (b) central aortic BP signal.

Fig. 4
Representative measurements of (a) PVR signals at an arm and a leg and (b) central aortic BP signal
Fig. 4
Representative measurements of (a) PVR signals at an arm and a leg and (b) central aortic BP signal
Close modal
Table 1

The range of central aortic systolic (SP), pulse (PP), and mean (MP) pressures associated with ten cardiac catheterization patients (mean ± SD)

SP (mmHg)PP (mmHg)MP (mmHg)
126 ± 2654 ± 2095 ± 20
SP (mmHg)PP (mmHg)MP (mmHg)
126 ± 2654 ± 2095 ± 20

3.2 Plant Dynamics: Blood Pressure Wave Propagation in Multichannel Arteries.

The plant dynamics considered in this work involves two channels: one associated with the relationship between central aortic BP and arm PVR (C1; Fig. 3), and the other associated with the relationship between central aortic BP and leg PVR (C2; Fig. 3). These channels are likely coprime given that they are associated primarily with arm arteries and leg arteries, respectively, whose mechanical and wave propagation properties are largely distinct [30]. Each channel dynamics is represented based on (i) a tube-load model to represent the BP wave propagation in the artery, (ii) a viscoelastic model to represent the characteristics of the arterial wall and the tissues, and (iii) a physics-based model of the BP cuff [20] (Fig. 5). In an array of our prior work, we have shown that such a linear time-invariant representation of BP wave propagation in arteries and arterial pressure–volume relationship is valid and effective at least within a short time window (i.e., in the order of minutes) [17,20,23,3133]. In particular, the tube-load model associated with each channel dynamics was expressed by the following discrete-time rational transfer function:
(16)

where P0(t) and Pi(t) are the central aortic BP and the arterial BP at the peripheral site i (i=1,2), Z[·] is the Z transform, and ni, θ1,i, and θ2,i are the patient-specific tube-load model parameters. In each patient, these model parameters were determined from the experimental data collected from the patient using a blind system identification procedure developed in our prior work [20]. Then, we confirmed that the rational transfer functions associated with central aortic-arm channel and central aortic-leg channel were coprime by examining the poles and zeros therein. Table 2 presents the range of the parameter values thus determined.

Fig. 5
Multichannel wave propagation dynamics in the arteries represented by (i) a tube-load model to represent the BP wave propagation in the artery, (ii) a viscoelastic model to represent the characteristics of the arterial wall and the tissues, and (iii) a physics-based model of the BP cuff
Fig. 5
Multichannel wave propagation dynamics in the arteries represented by (i) a tube-load model to represent the BP wave propagation in the artery, (ii) a viscoelastic model to represent the characteristics of the arterial wall and the tissues, and (iii) a physics-based model of the BP cuff
Close modal
Table 2

The range of tube-load model parameters associated with ten cardiac catheterization patients

Channel 1 (upper arm)Channel 2 (upper leg)
n1θ1,1θ2,1n2θ1,2θ2,2
2–1856.1–294.00.01–18.67–243.4–500.20.01–249.1
Channel 1 (upper arm)Channel 2 (upper leg)
n1θ1,1θ2,1n2θ1,2θ2,2
2–1856.1–294.00.01–18.67–243.4–500.20.01–249.1

In the channel dynamics, the viscoelastic artery wall-tissue model and the physics-based BP cuff model are interconnected. In addition, the physics-based BP cuff model involves nonlinearity, which is not suited to the UIO design developed in this work. However, these models can be inverted, i.e., the models allow for the computation of the peripheral BP waveforms from the corresponding PVR signals (Fig. 5). Therefore, the UIO design was applied to the tube-load model component of the channel dynamics by considering the arm and leg BP waveforms estimated from the corresponding PVR signals as the outputs of the channel dynamics expressed by Eq. (16). Note that Eq. (16) meets, or can be manipulated to meet (as described in Sec. 2), all the properties listed in Theorem 1.

3.3 Unknown Input Observer Design.

For each patient, the UIO was designed using the patient-specific plant dynamics derived above and guided by Theorem 1. First, the transfer function Gi(z) shown in Eq. (16) associated with the channel Ci (Fig. 3) was made proper by padding 2ni forward shift operators z in its numerator polynomial (both i=1 and i=2 were considered). Second, the controllable canonical form realizations associated with the patient-specific channel dynamics G1(z) ({A1,B1,C1,D1}) and G2(z) ({A2,B2,C2,D2}) were obtained. Third, the matrices A¯, B¯, C¯, and Cˇ in Eq. (3) required for the UIO design were computed using the patient-specific controllable canonical form realizations thus obtained. Finally, the UIO in Eq. (3) was constructed by designing the gain L¯.

Considering that the primary interest of this work was to investigate the validity and potential of the UIO-based input deconvolution technique, and that the performance of the UIO approaches to open-loop inverse filtering if L¯=0 and L¯=, the UIO gain L¯ was designed by trial and error to investigate if any choice of L¯ can lead to the UIO whose input deconvolution performance is superior to open-loop inverse filtering (in terms of the discrepancy between the true versus UIO-estimated central aortic BP). Two alternative approaches were specifically employed: pole placement and linear matrix inequality (LMI).

In the pole placement approach, the Butterworth pole layout [34] was considered as the desired pole locations for the closed-loop error dynamics in the continuous-time domain. For each patient, the continuous-time Butterworth poles with unspecified cut-off frequency were calculated according to the system order (i.e., the order of Eq. (3)). Then, a large number of candidate UIO pole locations were created by varying the cut-off frequency within its maximal value dictated by the sampling frequency Fs (i.e., πFs). Subsequently, these candidate pole locations were transformed into the discrete-time domain. For each candidate pole location, a candidate UIO was designed with the pole placement technique [29]. The performance of the UIO was measured in terms of the root-mean-squared error (RMSE) between the true versus estimated central aortic BP waveforms. The UIO associated with the patient (called the UIO-PP) was then determined as the one exhibiting the smallest RMSE among all the candidate UIOs.

In the LMI approach, the UIO design problem was cast into an LMI problem so that the poles associated with the closed-loop error dynamics are clustered in the region specified by |z|<σ. In this way, the error convergence rate may be specified explicitly in the UIO design (i.e., the settling time associated with the UIO error convergence has a settling time smaller than 4Fslnσ). Based on the UIO in Eq. (3), the error dynamics given by:
(17)
According to a prior work [35], the poles of the matrix A¯L¯C¯ associated with the error dynamics in Eq. (17) can be clustered in a region if and only if there exists a symmetric positive definite matrix P>0 such that:
(18)
where is the Kronecker product of matrices, while Ω1 and Ω2 are square matrices that specify the region in terms of an LMI:
(19)
The region specified by |z|<σ can be expressed as the following LMI:
(20)
which implies that Ω1=[σ00σ] and Ω2=[0100]. Hence, Eq. (18) reduces to the following:
(21)
Since the eigenvalues are invariant against matrix transpose, Eq. (21) can be rewritten as follows:
(22)
Using the change of variable K=L¯TP, Eq. (22) becomes the following LMI:
(23)

Once the matrix K and P satisfying Eq. (23) are found, the UIO gain L¯ can be determined by L¯=P1KT.

For each patient, the UIO was designed by solving the LMI feasibility problem in Eq. (23) while varying σ values (0<σ<1) to yield a large number of candidate UIOs. Similar to the pole placement technique above, the performance of the UIO was measured in terms of the RMSE between the true versus estimated central aortic BP waveforms. The UIO associated with the patient (called the UIO-LMI) was then determined as the one exhibiting the smallest RMSE among all the candidate UIOs.

3.4 Unknown Input Observer Performance Analysis.

The performance of UIO-PP and UIO-LMI was measured in terms of the waveform RMSE as well as the absolute errors associated with SP (systolic pressure error (SPE)) and PP (=SP-DP (diastolic BP); pulse pressure error (PPE)) between the true versus estimated central aortic BP waveforms. These errors were computed for each patient. Then, the mean and standard deviation (SD) were computed.

To assess the performance of the UIO in estimating central aortic BP waveform relative to the conventional as well as primitive techniques, the same error metrics (RMSE, SPE, and PPE) were likewise computed for the conventional open-loop inverse filtering technique as well as the arm and leg PVR signals scaled to central aortic diastolic and mean pressures. For each patient, open-loop inverse filtering technique was performed by filtering (i) the arm PVR signal by the inverse of G1(z) as well as (ii) the leg PVR signal by the inverse of G2(z) to yield the estimated central aortic BP waveforms. For each patient, PVR scaling was performed by linearly calibrating the arm and leg PVR signals so that the diastolic and mean pressure levels associated with the PVR signals become identical to the central aortic diastolic and mean pressures measured from the patient as follows:
(24)

where u¯ and uˇ are central aortic mean and diastolic pressures, while w¯i and wˇi are mean and diastolic values of the PVR signal wi(t) (i=1,2). Then, the aforementioned error metrics between the true central aortic BP waveform versus these waveforms were computed for each patient, whose mean and SD were subsequently computed.

Statistical significance in the difference between the error metrics associated with UIO, open-loop inverse filtering, and scaled PVR signals was determined with the Wilcoxon signed-rank test, where the Bonferroni correction factor of 2 was used to account for the multiple comparisons (i.e., UIO-PP versus inverse filtering and scaled PVR signals, as well as UIO-LMI versus inverse filtering and scaled PVR signals).

4 Results and Discussion

Table 3 summarizes RMSE, SPE, and PPE associated with central aortic BP waveforms derived from UIO-PP, UIO-LMI, open-loop inverse filtering, and scaled PVR signals, while Fig. 6 illustrates a few representative examples of true versus estimated central BP waveforms: (a) an example where UIO shows performance marginally superior to inverse filtering and (b) an example where UIO shows performance largely superior to inverse filtering.

Fig. 6
Representative examples of true versus estimated central BP waveforms: (a) an example where UIO-PP shows performance marginally superior to inverse filtering and (b) an example where UIO-PP shows performance largely superior to inverse filtering
Fig. 6
Representative examples of true versus estimated central BP waveforms: (a) an example where UIO-PP shows performance marginally superior to inverse filtering and (b) an example where UIO-PP shows performance largely superior to inverse filtering
Close modal
Table 3

The RMSEs, SPEs, and PPEs associated with central aortic blood pressure waveforms derived from UIO-PP, UIO-LMI, open-loop IF, and scaled PVR signals (N = 10). UIO-PP: UIO designed with pole placement. UIO-LMI: UIO designed with LMI

UIO-PP1 (y1=P1)UIO-PP2 (y1=P2)Average
UIO-PP
RMSE (mmHg)3.7 ± 1.73.6 ± 1.03.7 ± 1.4a,b
SPE (mmHg)2.7 ± 2.02.4 ± 1.62.6 ± 1.7b
PPE (mmHg)3.2 ± 3.42.9 ± 1.93.0 ± 2.7a,b
UIO-PP1 (y1=P1)UIO-PP2 (y1=P2)Average
UIO-PP
RMSE (mmHg)3.7 ± 1.73.6 ± 1.03.7 ± 1.4a,b
SPE (mmHg)2.7 ± 2.02.4 ± 1.62.6 ± 1.7b
PPE (mmHg)3.2 ± 3.42.9 ± 1.93.0 ± 2.7a,b
UIO-LMI1 (y1=P1)UIO-LMI2 (y1=P2)Average
UIO-LMI
RMSE (mmHg)4.2 ± 2.04.4 ± 1.74.3 ± 1.8a
SPE (mmHg)3.4 ± 3.13.3 ± 2.03.3 ± 2.6b
PPE (mmHg)4.8 ± 4.14.9 ± 3.44.8 ± 3.7
UIO-LMI1 (y1=P1)UIO-LMI2 (y1=P2)Average
UIO-LMI
RMSE (mmHg)4.2 ± 2.04.4 ± 1.74.3 ± 1.8a
SPE (mmHg)3.4 ± 3.13.3 ± 2.03.3 ± 2.6b
PPE (mmHg)4.8 ± 4.14.9 ± 3.44.8 ± 3.7
IF1 (y1=P1)IF2 (y1=P2)Average
Inverse filtering
RMSE (mmHg)5.3 ± 2.44.9 ± 1.95.1 ± 2.1
SPE (mmHg)5.5 ± 4.72.8 ± 2.24.2 ± 3.9
PPE (mmHg)7.2 ± 5.13.9 ± 4.35.5 ± 4.9
IF1 (y1=P1)IF2 (y1=P2)Average
Inverse filtering
RMSE (mmHg)5.3 ± 2.44.9 ± 1.95.1 ± 2.1
SPE (mmHg)5.5 ± 4.72.8 ± 2.24.2 ± 3.9
PPE (mmHg)7.2 ± 5.13.9 ± 4.35.5 ± 4.9
Arm PVRLeg PVRAverage
Scaled PVR signals
RMSE (mmHg)4.6 ± 1.85.8 ± 2.65.2 ± 2.3
SPE (mmHg)8.8 ± 6.26.2 ± 5.87.5 ± 6.0
PPE (mmHg)7.7 ± 6.38.8 ± 9.28.2 ± 7.7
Arm PVRLeg PVRAverage
Scaled PVR signals
RMSE (mmHg)4.6 ± 1.85.8 ± 2.65.2 ± 2.3
SPE (mmHg)8.8 ± 6.26.2 ± 5.87.5 ± 6.0
PPE (mmHg)7.7 ± 6.38.8 ± 9.28.2 ± 7.7
a

p < 0.025 with respect to inverse filtering (Wilcoxon signed-rank test).

b

p < 0.025 with respect to scaled PVR signals (Wilcoxon signed-rank test).

The input deconvolution based on the UIO was in general superior to open-loop inverse filtering (which represents the limiting performance of the UIO at L¯=0 and L¯=) in the sense that central aortic BP waveform derived by the UIO was closer in shape to the true central aortic BP waveform than the one derived by the inverse filtering. This suggests that the corrective error feedback action provided by the UIO is practically meaningful. On the average, the RMSE associated with the UIO-PP was 27.5% smaller than the open-loop inverse filtering and 28.8% smaller than the scaled arm and leg PVR signals (Table 3). Likewise, the RMSE associated with the UIO-LMI was 15.7% smaller than the open-loop inverse filtering and 17.3% smaller than the scaled arm and leg PVR signals (Table 3). In case of UIO-PP, the UIO designed with P1 and P2 designated as its input and output (i.e., y1=P1 and y2=P2 in Eq. (3); UIO-PP1 in Table 3) was superior to the open-loop inverse filtering based on both P1 (i.e., L¯=0; IF1 in Table 3) and P2 (i.e., L¯=; IF2 in Table 3) in all ten patients, while the UIO designed with P2 and P1 designated as its input and output (i.e., y1=P2 and y2=P1 in Eq. (3); UIO-PP2 in Table 3) was superior to the same open-loop inverse filtering (i.e., IF1 and IF2) in most (i.e., 8) patients. In case of UIO-LMI, the UIO designed with (i) P1 and P2 designated as its input and output (UIO-LMI1 in Table 3) and (ii) P2 and P1 designated as its input and output (UIO-LMI2 in Table 3) was superior to both IF1 and IF2 in six and five patients, respectively. In the remaining patients whose UIO did not excel both IF1 and IF2, the UIO exhibited a RMSE value between the RMSE values associated with IF1 and IF2, as predicted by the mathematical analysis presented in Intuitive Interpretation and Limiting Behavior. Both UIO-PP and UIO-LMI exhibited RMSE values smaller than those associated with the scaled arm and leg PVR signals in seven and all ten patients, respectively.

The input deconvolution based on UIO-PP also estimated central aortic SP and PP more accurately than open-loop inverse filtering and scaled PVR signals. The SPE and PPE associated with the UIO-PP were smaller than those associated with both inverse filtering and scaled PVR signals. On the average, the SPE and PPE associated with the UIO-PP were 38.1% and 45.5% smaller than open-loop inverse filtering, and 56.3% and 63.4% smaller than the scaled arm and leg PVR signals (Table 3). The SPE and PPE associated with the UIO-LMI were likewise 21.4% and 12.7% smaller than inverse filtering, and 56.0% and 41.5% smaller than the scaled arm and leg PVR signals (Table 3). This is notable in that SPE and PPE were not explicitly optimized during the UIO design process (although they may improve as the waveform accuracy is optimized). However, in contrast to UIO-PP (which yielded average SPE and PPE values smaller than the same values associated with both IF1 and IF2), UIO-LMI yielded average SPE and PPE values between the same values associated with IF1 and IF2.

Comparing UIO-PP and UIO-LMI, the UIO-PP yielded superior accuracy to the UIO-LMI. However, its design process was more ad hoc and resource-intensive. In fact, the UIO-PP was obtained by rigorously optimizing the RMSE metric associated with the central aortic BP waveform, whereas the UIO-LMI was obtained simply by solving a LMI feasibility problem with a rudimentary constraint. It is plausible that the efficacy of the UIO-LMI may further be improved by augmenting the currently used LMI problem formulation with more stringent constraints that specify the transient and steady-state behaviors of the error convergence dynamics. Similarly, the efficacy of the UIO-PP may likewise be improved by considering a wide range of desired closed-loop pole locations. But in any case, given that the performance of the UIO-based input deconvolution technique is bounded by the open-loop inverse filtering, both design procedures may require careful selection and tuning of the observer gain.

5 Conclusions

In our attempt to enable more accurate patient-specific estimation of central aortic BP waveform from noninvasively acquired peripheral arterial pulse waveforms, we developed and validated an observer-based closed-loop approach to input deconvolution in coprime multichannel linear dynamical systems. A universal UIO design algorithm and insightful mathematical analysis were presented to elucidate key properties and inherent performance limitations of the UIO. The results obtained from the experimental data showed promise and consistency with the mathematical analysis. The UIO presented in our work has applicability to a broad spectrum of coprime multichannel linear systems beyond central aortic BP estimation problem. Future work includes the extension of the proposed input deconvolution algorithm to incorporate more advanced state estimation techniques with rigorous robustness guarantee against errors and uncertainties pertaining to the channel dynamics and measurement noise, as well as intensive evaluation of the algorithm in a larger datasets as well as other real-world applications.

Acknowledgment

Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the NSF and NIH.

Funding Data

  • National Science Foundation (NSF) (Grant No. CMMI-1431672; Funder ID: 10.13039/100000001).

  • National Institutes of Health (NIH) (Grant No. EB018818; Funder ID: 10.13039/100000002).

References

1.
Benjamin
,
E. J.
,
Virani
,
S. S.
,
Callaway
,
C. W.
,
Chamberlain
,
A. M.
,
Chang
,
A. R.
,
Cheng
,
S.
,
Chiuve
,
S. E.
,
Cushman
,
M.
,
Delling
,
F. N.
,
Deo
,
R.
,
De Ferranti
,
S. D.
,
Ferguson
,
J. F.
,
Fornage
,
M.
,
Gillespie
,
C.
,
Isasi
,
C. R.
,
Jiménez
,
M. C.
,
Jordan
,
L. C.
,
Judd
,
S. E.
,
Lackland
,
D.
,
Lichtman
,
J. H.
,
Lisabeth
,
L.
,
Liu
,
S.
,
Longenecker
,
C. T.
,
Lutsey
,
P. L.
,
MacKey
,
J. S.
,
Matchar
,
D. B.
,
Matsushita
,
K.
,
Mussolino
,
M. E.
,
Nasir
,
K.
,
O'Flaherty
,
M.
,
Palaniappan
,
L. P.
,
Pandey
,
A.
,
Pandey
,
D. K.
,
Reeves
,
M. J.
,
Ritchey
,
M. D.
,
Rodriguez
,
C. J.
,
Roth
,
G. A.
,
Rosamond
,
W. D.
,
Sampson
,
U. K. A.
,
Satou
,
G. M.
,
Shah
,
S. H.
,
Spartano
,
N. L.
,
Tirschwell
,
D. L.
,
Tsao
,
C. W.
,
Voeks
,
J. H.
,
Willey
,
J. Z.
,
Wilkins
,
J. T.
,
Wu
,
J. H. Y.
,
Alger
,
H. M.
,
Wong
,
S. S.
, and
Muntner
,
P.
,
2018
, “
Heart Disease and Stroke Statistics—2018 Update: A Report From the American Heart Association
,”
Circulation
, 137(12), pp.
e67
e492
.10.1161/CIR.0000000000000558
2.
Mceniery
,
C. M.
,
Cockcroft
,
J. R.
,
Roman
,
M. J.
,
Franklin
,
S. S.
, and
Wilkinson
,
I. B.
,
2014
, “
Central Blood Pressure: Current Evidence and Clinical Importance
,”
Eur. Heart J.
,
35
(
26
), pp.
1719
1725
.10.1093/eurheartj/eht565
3.
Ferguson
,
J. M.
,
Minas
,
J.
,
Siapantas
,
S.
,
Komesaroff
,
P. A.
, and
Sudhir
,
K.
,
2008
, “
Effects of a Fixed-Dose ACE Inhibitor-Diuretic Combination on Ambulatory Blood Pressure and Arterial Properties in Isolated Systolic Hypertension
,”
J. Cardiovasc. Pharmacol.
,
51
(
6
), pp.
590
595
.10.1097/FJC.0b013e31817a8316
4.
Roman
,
M. J.
,
Okin
,
P. M.
,
Kizer
,
J. R.
,
Lee
,
E. T.
,
Howard
,
B. V.
, and
Devereux
,
R. B.
,
2010
, “
Relations of Central and Brachial Blood Pressure to Left Ventricular Hypertrophy and Geometry: The Strong Heart Study
,”
J. Hypertens.
,
28
(
2
), pp.
384
388
.10.1097/HJH.0b013e328333d228
5.
Roman
,
M. J.
,
Devereux
,
R. B.
,
Kizer
,
J. R.
,
Lee
,
E. T.
,
Galloway
,
J. M.
,
Ali
,
T.
,
Umans
,
J. G.
, and
Howard
,
B. V.
,
2007
, “
Central Pressure More Strongly Relates to Vascular Disease and Outcome Than Does Brachial Pressure: The Strong Heart Study
,”
Hypertension
,
50
(
1
), pp.
197
203
.10.1161/HYPERTENSIONAHA.107.089078
6.
Sharman
,
J. E.
,
Lim
,
R.
,
Qasem
,
A. M.
,
Coombes
,
J. S.
,
Burgess
,
M. I.
,
Franco
,
J.
,
Garrahy
,
P.
,
Wilkinson
,
I. B.
, and
Marwick
,
T. H.
,
2006
, “
Validation of a Generalized Transfer Function to Noninvasively Derive Central Blood Pressure During Exercise
,”
Hypertension
,
47
(
6
), pp.
1203
1208
.10.1161/01.HYP.0000223013.60612.72
7.
Swamy, G., Xu, D., Olivier, N. B., and Mukkamala, R.
,
2009
, “
An Adaptive Transfer Function for Deriving the Aortic Pressure Waveform from a Peripheral Artery Pressure Waveform
,”
Am. J. Physiol. Heart Circ. Physiol.
,
297
(
5
), pp.
H1956
–H
1963
.10.1152/ajpheart.00155.2009
8.
Ding
,
F.-H.
,
Fan
,
W.-X.
,
Zhang
,
R.-Y.
,
Zhang
,
Q.
,
Li
,
Y.
, and
Wang
,
J.-G.
,
2011
, “
Validation of the Noninvasive Assessment of Central Blood Pressure by the SphygmoCor and Omron Devices Against the Invasive Catheter Measurement
,”
Am. J. Hypertens.
,
24
(
12
), pp.
1306
1311
.10.1038/ajh.2011.145
9.
Gallagher
,
D.
,
Adji
,
A.
, and
Orourke
,
M.
,
2004
, “
Validation of the Transfer Function Technique for Generating Central From Peripheral Upper Limb Pressure Waveform
,”
Am. J. Hypertens.
,
17
(
11
), pp.
1059
1067
.10.1016/j.amjhyper.2004.05.027
10.
Chen
,
C. H.
,
Ting
,
C. T.
,
Nussbacher
,
A.
,
Nevo
,
E.
,
Kass
,
D. A.
,
Pak
,
P.
,
Wang
,
S. P.
,
Chang
,
M. S.
, and
Yin
,
F. C.
,
1996
, “
Validation of Carotid Artery Tonometry as a Means of Estimating Augmentation Index of Ascending Aortic Pressure
,”
Hypertension
,
27
(
2
), pp.
168
175
.10.1161/01.HYP.27.2.168
11.
Chen
,
C. H.
,
Nevo
,
E.
,
Fetics
,
B.
,
Pak
,
P. H.
,
Yin
,
F. C.
,
Maughan
,
W. L.
, and
Kass
,
D. A.
,
1997
, “
Estimation of Central Aortic Pressure Waveform by Mathematical Transformation of Radial Tonometry Pressure
,”
Circulation
,
95
(
7
), pp.
1827
1836
.10.1161/01.CIR.95.7.1827
12.
Shih
,
Y.-T.
,
Cheng
,
H.-M.
,
Sung
,
S.-H.
,
Hu
,
W.-C.
, and
Chen
,
C.-H.
,
2013
, “
Comparison of Two Generalized Transfer Functions for Measuring Central Systolic Blood Pressure by an Oscillometric Blood Pressure Monitor
,”
J. Hum. Hypertens.
,
27
(
3
), pp.
204
210
.10.1038/jhh.2012.17
13.
Cheng
,
H.
,
Sung
,
S.
,
Shih
,
Y.
,
Chuang
,
S.
,
Yu
,
W.
, and
Chen
,
C.
,
2012
, “
Measurement of Central Aortic Pulse Pressure: Noninvasive Brachial Cuff-Based Estimation by a Transfer Function Vs. a Novel Pulse Wave Analysis Method
,”
Am. J. Hypertens.
,
25
(
11
), pp.
1162
1169
.10.1038/ajh.2012.116
14.
Hope
,
S. A.
,
Meredith
,
I. T.
,
Tay
,
D.
, and
Cameron
,
J. D.
,
2007
, “
Generalizability' of a Radial-Aortic Transfer Function for the Derivation of Central Aortic Waveform Parameters
,”
J. Hypertens.
,
25
(
9
), pp.
1812
1820
.10.1097/HJH.0b013e328277595d
15.
Cameron
,
J. D.
,
McGrath
,
B. P.
, and
Dart
,
A. M.
,
1998
, “
Use of Radial Artery Applanation Tonometry and a Generalized Transfer Function to Determine Aortic Pressure Augmentation in Subjects With Treated Hypertension
,”
J. Am. Coll. Cardiol.
,
32
(
5
), pp.
1214
1220
.10.1016/S0735-1097(98)00411-2
16.
Fetics, B., Nevo, E., Chen, C.-H. , and
Kass
,
D. A.
,
1999
, “
Parametric Model Derivation of Transfer Function for Noninvasive Estimation of Aortic Pressure by Radial Tonometry
,”
IEEE Trans. Biomed. Eng.
,
46
(
6
), pp.
698
706
.10.1109/10.764946
17.
Rashedi
,
M.
,
Fazeli
,
N.
,
Chappell
,
A.
,
Wang
,
S.
,
MacArthur
,
R.
,
McMurtry
,
M. S.
,
Finegan
,
B. A.
, and
Hahn
,
J. O.
,
2013
, “
Comparative Study on Tube-Load Modeling of Arterial Hemodynamics in Humans
,”
ASME J. Biomech. Eng.
,
135
, p.
31005
.10.1115/1.4023373
18.
Hope
,
S. A.
,
Tay
,
D. B.
,
Meredith
,
I. T.
, and
Cameron
,
J. D.
,
2003
, “
Use of Arterial Transfer Functions for the Derivation of Aortic Waveform Characteristics
,”
J. Hypertens.
,
21
(
7
), pp.
1299
1305
.10.1097/00004872-200307000-00017
19.
Abed-Meraim
,
K.
,
Qiu
,
W.
, and
Hua
,
Y.
,
1997
, “
Blind System Identification
,”
Proc. IEEE
,
85
(
8
), pp.
1310
1322
.10.1109/5.622507
20.
Ghasemi
,
Z.
,
Kim
,
C.
,
Ginsberg
,
E.
,
Gupta
,
A.
, and
Hahn
,
J.-O.
,
2017
, “
Model-Based Blind System Identification Approach to Estimation of Central Aortic Blood Pressure Waveform From Noninvasive Diametric Circulatory Signals
,”
ASME J. Dyn. Syst. Meas. Control
,
139
, p.
061003
.10.1115/1.4035451
21.
Swamy
,
G.
,
Ling
,
Q.
,
Li
,
T.
, and
Mukkamala
,
R.
,
2007
, “
Blind Identification of the Aortic Pressure Waveform From Multiple Peripheral Artery Pressure Waveforms
,”
Am. J. Physiol. Heart Circ. Physiol.
,
292
(
5
), pp.
H2257
H2264
.10.1152/ajpheart.01159.2006
22.
Abutaleb
,
A. S.
,
Waheed
,
M. E.-S.
, and
Elhamy
,
N. M.
,
2010
, “
Multichannel Blind Deconvolution Using the Stochastic Calculus for the Estimation of the Central Arterial Pressure
,”
Math. Probl. Eng.
,
2010
, pp.
1
21
.10.1155/2010/602373
23.
Hahn
,
J.-O.
,
Reisner
,
A. T.
, and
Asada
,
H. H.
,
2009
, “
Blind Identification of Two-Channel IIR Systems With Application to Central Cardiovascular Monitoring
,”
ASME J. Dyn. Syst. Meas. Control
,
131
, p.
051009
.10.1115/1.3155011
24.
Hahn
,
J. O.
,
McCombie
,
D. B.
,
Reisner
,
A. T.
,
Hojman
,
H. M.
, and
Harry
,
A.
,
2010
, “
Identification of Multichannel Cardiovascular Dynamics Using Dual Laguerre Basis Functions for Noninvasive Cardiovascular Monitoring
,”
IEEE Trans. Control Syst. Technol.
,
18
(
1
), pp.
170
176
.10.1109/TCST.2008.2009996
25.
Zhang
,
G.
,
Hahn
,
J. O.
, and
Mukkamala
,
R.
,
2011
, “
Tube-Load Model Parameter Estimation for Monitoring Arterial Hemodynamics
,”
Front. Physiol.
,
2
, pp.
1
18
.
26.
Hua
,
Y.
,
2002
, “
Blind Methods of System Identification
,”
Circuits, Syst. Signal Process.
,
21
(
1
), pp.
91
108
.10.1007/BF01211654
27.
Alberge
,
F.
,
Duhamel
,
P.
, and
Nikolova
,
M.
,
2002
, “
Adaptive Solution for Blind Identification/Equalization Using Deterministic Maximum Likelihood
,”
IEEE Trans. Signal Process.
,
50
(
4
), pp.
923
936
.10.1109/78.992140
28.
Gürelli
,
M. L.
, and
Nikias
,
C. L.
,
1995
, “
EVAM: An Eigenvector-Based Deconvolution of Input Colored Signals
,”
IEEE Trans. Signal Process.
,
43
(
1
), pp.
134
149
.10.1109/78.365293
29.
Chen
,
C. T.
,
1998
,
Linear System Theory and Design
,
Oxford University Press
,
New York
.
30.
Buntin
,
C. M.
, and
Silver
,
F. H.
,
1990
, “
Noninvasive Assessment of Mechanical Properties of Peripheral Arteries
,”
Ann. Biomed. Eng.
,
18
(
5
), pp.
549
566
.10.1007/BF02364617
31.
Ghasemi
,
Z.
,
Lee
,
J. C.
,
Kim
,
C. S.
,
Cheng
,
H. M.
,
Sung
,
S. H.
,
Chen
,
C. H.
,
Mukkamala
,
R.
, and
Hahn
,
J. O.
,
2018
, “
Estimation of Cardiovascular Risk Predictors From Non-Invasively Measured Diametric Pulse Volume Waveforms Via Multiple Measurement Information Fusion
,”
Sci. Rep.
,
8
(
1
), p.
10433
.10.1038/s41598-018-28604-6
32.
Fazeli
,
N.
,
Kim
,
C. S.
,
Rashedi
,
M.
,
Chappell
,
A.
,
Wang
,
S.
,
MacArthur
,
R.
,
McMurtry
,
M. S.
,
Finegan
,
B.
, and
Hahn
,
J. O.
,
2014
, “
Subject-Specific Estimation of Central Aortic Blood Pressure Via System Identification: Preliminary In-Human Experimental Study
,”
Med. Biol. Eng. Comput.
,
52
(
10
), pp.
895
904
.10.1007/s11517-014-1185-3
33.
Abdollahzade
,
M.
,
Kim
,
C.-S.
,
Fazeli
,
N.
,
Finegan
,
B. A.
,
Sean McMurtry
,
M.
, and
Hahn
,
J.-O.
,
2014
, “
Data-Driven Lossy Tube-Load Modeling of Arterial Tree: In-Human Study
,”
ASME J. Biomech. Eng.
,
136
(
10
), p.
101011
.10.1115/1.4028089
34.
Franklin
,
G. F.
,
Powell
,
J. D.
, and
Emami-Naeini
,
A.
,
1994
,
Feedback Control of Dynamic Systems
,
Pearson Higher Education
,
Upper Saddle River, NJ
.
35.
Chilali
,
M.
, and
Gahinet
,
P.
,
1996
, “
H∞ Design With Pole Placement Constraints: An LMI Approach
,”
IEEE Trans. Automat. Control
,
41
(
3
), pp.
358
367
.10.1109/9.486637