Particle filterbased robust adaptive control for closedloop administration of sodium nitroprusside
 Nicolò Malagutti^{1}Email author
DOI: 10.1186/2194399018
© Malagutti; licensee Springer. 2014
Received: 1 August 2013
Accepted: 3 February 2014
Published: 25 March 2014
Abstract
Abstract
Automatic closedloop administration of medicinal drugs has been the subject of intense research for decades due to its undisputed potential benefits in terms of cost savings and improved patient outcomes. However, concerns still exist about the ultimate safety of engineered feedback controllers. Manual methods remain dominant in clinical practice. In this context, we present a novel feedback control architecture, which combines multiple robust controllers with a particle filterbased method for realtime tracking of a patient’s doseresponse characteristic. The proposed method is applied to the case of the drug sodium nitroprusside, a vasodepressor used in the treatment of acute hypertension in intensive care and surgery, which is modelled as having a lineartimevarying doseresponse characteristic. Our design takes into account the uncertainty in the patient response parameters, as well as potential nonzeromean disturbances in the baseline arterial pressure and several possible time trends in the variation of the doseresponse model. The performance and safety of the new approach are evaluated through an extensive computational simulation campaign. The results show that the proposed system can achieve adequate and safe feedback control of mean arterial pressure, thus validating our analysis and design. Our findings also highlight the fundamental  and possibly clinically overlooked  role of system excitation in ensuring that successful simultaneous identification and control of timevarying drug administration systems can be achieved.
Keywords
Biomedical engineering Drug delivery Feedback systems Adaptive control Robust control Closedloop Hypertension Sodium nitroprusside Particle filteringBackground
In many clinical settings, the administration of suitable medicinal drugs is required to maintain important biological signals within an acceptable range. Unfortunately, significant variability exists in the way different patients respond to the same drug dose (interpatient variability), and variations can even occur in the response of one individual over time (intrapatient variability). Appropriate dosing of drugs characterised by a narrow range of therapeutic concentrations and high variability in the response is therefore a challenging task. In these cases, dose titration protocols are followed, by which clinical operators administer an initial drug amount, generally according to population statistics (e.g. age, gender, height, weight,...), followed by close monitoring and periodic manual adjustment of the dose on the basis of the clinical observations. This manual form of closedloop regulation is prone to human error [1] and can be very timeconsuming in clinical environments where timely intervention and/or staff levels can be an issue. In this context, the successful development of safe and effective automatic control methods could bring the significant benefits of improved patient outcomes and better allocation of clinical resources [2].
Over the last 30 years, much research has been devoted to the development of strategies for closedloop feedback control of drug administration for a variety of applications, including control of haemodynamics, insulin for diabetes management, chemotherapy, anaesthesia and neuromuscular blockade [3–5]. Remarkably, despite promising results in clinical research publications, closedloop methods for automatic drug administration have not yet been embraced by the medical community, and manual control remains, to date, the standard of care in the clinical setting. Indeed, leading scholars in the field have recently acknowledged that “Success in the development of [...] closedloop biomedical devices, will be contingent on the development of robust, verifiable advanced control algorithms”[6]. This suggests that insufficient guarantees of system robustness may be a key reason  in terms of risk assessment considerations  underpinning clinicians’ continued preference for manual methods.
This paper considers the specific problem of intravenous infusion of the drug sodium nitroprusside (SNP). The SNP case study typifies research into closedloop drug administration. SNP is a fastacting vasodepressor used in the treatment of acute hypertension in intensive care and surgery patients. The drug is highly effective in reducing mean arterial pressure (MAP); however, incorrect dosing may lead to undesired hypotensive peaks or metabolic toxicity [7], hence the interest in reliable automatic drug dosing. Following the clinical validation of a doseresponse model by Slate [8], administration of SNP was the subject of intense research during the 1980s and 1990s, and a wide range of control engineering approaches were proposed, including selftuning regulators [9], model reference adaptive control [10], multiple model adaptive control [11, 12], fuzzy control [13] and model predictive control [14] (see [15] for a review). A commercial device (IVAC titrator; IVAC Medical Systems, Inc., San Diego, CA, USA) developed for SNP titration in intensive care applications reported success in clinical trials [16] but was never successfully marketed (as discussed in [17]). To the best of our knowledge, the automatic management of MAP through SNP and other drugs remains to this day an experimental pursuit.
The work herein presents a novel approach to the control of automatic infusion of SNP with a strong focus on robustness. The proposed methodology is an extension of an earlier approach which featured multiple robust feedback controllers designed with μ synthesis [18] and used Kalmanfilterbased estimators [19] to determine controller selection in the context of a robust multiple model adaptive control architecture [20]. While retaining the same robust controllers as the previous work, the new approach uses particle filtering to generate an estimate of the doseresponse characteristic in real time and exploits the estimation result to inform appropriate feedback control. The new method has been named Robust Adaptive Control with Particle Filtering (RACPF) and we have reported on a nonclinical case study featuring this architecture in [21], where we positively compared the performance of RACPF with that of a Kalmanfilterbased approach. We deem particle filters to be better suited to the estimation of systems characterised by timevarying parameters, nonGaussian disturbances and even nonlinearities, such as pharmacokineticpharmacodynamic systems [22]. As a more general estimation tool, particle filtering allows us to remove undesirable ad hoc filter adaptations present in our earlier work. The resulting architecture can be readily transposed to other applications as required. This article provides an overview of the main characteristics of RACPF and reports on an extensive computational simulation campaign designed to assess the viability of the new methodology and its effectiveness in delivering safe automatic feedback control of SNP infusion for a broad range of response characteristics and disturbances.
The manuscript is structured as follows: the Methods Section comprises of three subsections detailing the SNP doseresponse model, the proposed RACPF control architecture, and the characteristics of the computational simulations, respectively; the Results Section reports on the outcomes of the simulation campaign; finally, the Discussion and Conclusion Sections comment on the results and the potential of the proposed approach, and outline future research directions.
Methods
Model and problem description
Doseresponse model
where v (t) ∼ N (0,1), w (t) ∼ N (0,2) are normally distributed random noise signals at the input (actuation noise) and the output (measurement noise), respectively. The signal y_{meas} (t) represents the measurable MAP value.
Control performance requirements
The following performance requirements apply [23]:

A settling time of preferably 10 min or less, but no more than 15 min

MAP should be contained within ±10 mmHg of the desired setpoint value at most times

Under no circumstances should the system display resonant (persistent oscillatory) or unstable behaviour or cause MAP to drop below a predetermined dangerous threshold (set at 60 mmHg for the purpose of this work)

To ensure that SNP toxicity is prevented, the infusion rate should not exceed a predetermined value (set here at 200 ml/h)

Highfrequency dynamics in the control signal should be limited since drug delivery is generally provided through mechanically actuated infusion pumps
The control approach
Controller design
The controllers are designed using μ synthesis. This is an advanced control engineering method used in the design of linear feedback controllers for applications where stability and performance must be guaranteed in the face of uncertainty in the model of the system to be controlled (robustness). In μ synthesis, a dynamical system is considered along with a model or bounding function of any uncertainty it features, whether structured (e.g. parametric) or unstructured (e.g. arising from the presence of delays and/or nonlinearities). Any performance specifications are also expressed in terms of frequency domain bounds. The technique involves a numerical search over the space of stabilising controllers and enables the designer to compute a feedback controller capable of delivering robust performance in the closedloop configuration, if one such controller exists [24].
Results of controller design
Controller  To suit K  To suit K  To suit  To suit 

number i  (best performance)  (acceptable performance)  α  delay 
(mmHg/(ml/h))  (mmHg/(ml/h))  T (s)  
1  0.250.57  0.250.78  0.250.75  050 
2  0.571.25  0.371.77  0.250.75  050 
3  1.252.30  0.753.45  0.250.75  050 
4  2.304.32  1.336.50  0.250.75  050 
5  4.329.50  2.459.50  0.250.75  050 
Particle filtering
Particle filtering is a sequential Monte Carlo method which iteratively computes the conditional probability distribution of the state of a partially observed dynamical system described by a stochastic statespace model [25]. In this application, the main focus is not to estimate the state but rather to identify the parameters (particularly patient sensitivity K) so that the correct controller can be used in the feedback loop. To this end, the estimation problem is recast as a nonlinear tracking problem to include the uncertain parameters in the state, as shown in (7) below. We refer to the original system states as ‘linear’ states x^{ l }, as opposed to the ‘nonlinear’ states x^{ n }representing the parameters. Such a mixed linear/nonlinear formulation is well suited to the application of a specialised form of particle filtering called marginalised particle filtering[26].
where $k=\lfloor \frac{t}{{T}_{s}}\rfloor $ (⌊·⌋ is the floor operator). The subscript d indicates zeroorder hold discretisation of continuoustime model (5).
The aim of filtering is to obtain the posterior probability density of the state conditioned on the observations up until that time point, i.e. p(x(k)Z(0 : k)), where p indicates probability and $Z(0:k)\equiv {\left\{\left[u\right(i\left)\phantom{\rule{0.3em}{0ex}}\phantom{\rule{0.3em}{0ex}}\phantom{\rule{0.3em}{0ex}}\phantom{\rule{0.3em}{0ex}}{y}_{\text{meas}}\right(i\left)\right]\right\}}_{i=0}^{k}$. represents the observations. The posterior probability can be calculated analytically in the case of linearGaussian systems using Kalman filters, while, for more general cases, numerical approximation methods must be used [26]. Particle filters are one such method, which approximates the posterior probability with a finite number of samples (particles).
Marginalised particle filters exploit the fact that a subset of the states can be treated as conditionally linear (x^{ l }here). This methodology involves an estimation of these states using the optimal Kalman filter result (marginalisation), while the other states (x^{ n }) are estimated by the particle filter. As the dimensionality of the numerical problem is thus reduced, marginalised particle filtering is associated with a comparatively lower computational burden. Here follows a description of our algorithm, which combines a bootstrap particle filter implementation [27] with marginalisation.
 (a)Initialisation. State x _{ i }(0) for particle i = 1,…,N is set as$\begin{array}{ll}{x}^{l}\left(0\right)\sim (0,P(0\left)\right)& {x}^{n}\left(0\right)\sim \left[\begin{array}{l}U(0.25,9.5)\\ U(10,50)\\ U(0.25,0.75)\end{array}\right]\end{array}$
 (b)Weighting. For i = 1,…,N compute the estimated output from each particle as ${\u0177}_{i}\left(k\right)={p}_{0}{C}_{d}\left({x}_{i}^{n}\right(k\left)\right){x}_{i}^{l}\left(k\right)$. Then, evaluate the particles’ normalised importance weights $\stackrel{~}{q}\left(k\right)$$\begin{array}{ll}{q}_{i}=p\left(y\right(k\left)\right{\u0177}_{i}\left(k\right))& {\stackrel{~}{q}}_{i}\left(k\right)=\frac{{q}_{i}\left(k\right)}{\sum _{j=1}^{N}{q}_{j}\left(k\right)}.\end{array}$
 (c)
Resampling. Resample N particles on the basis of the weights obtained in step (b) using a residual resampling algorithm [28].
 (d)
Time update. For each particle i=1,…,N
 (i)Kalman filter correction of the linear state estimate using the available observation Z(k)${x}_{i}^{l}\left(k\rightk)={x}_{i}^{l}(kk1)+{H}_{i}\left(k\right)\left(\phantom{\rule{0.8pt}{0ex}}y\right(k){p}_{0}{C}_{d,i}(k\left){x}_{i}^{l}\right(kk1)),$
 (i)
Sample χ via (9) and update the nonlinear states via (8), then calculate ${A}_{d,i}\left({x}_{i}^{n}\right(k\left)\right)\left(k\right)$ and ${B}_{d,i}\left({x}_{i}^{n}\right(k\left)\right)\left(k\right)$ using (7).
 (i)
Time update of the marginalised states (${x}_{i}^{l}(k+1k)$) via (6) and the state estimate covariance matrix using ${P}_{i}(k+1k)={A}_{d,i}(k\left){P}_{i}\right(k\leftk\right){A}_{d,i}^{T}\left(k\right)+Q,$ where Q is the covariance matrix of the state/input noise v.
 (e)
Iteration. Increase k → k+1 and repeat over from step (b).
Controller selection
where π_{ j }is the probability of the true parameters belonging to subset j and N is the total number of particles.
Theoretical proofs of robustness
The use of the μ synthesis controller design techniques gives a mathematical guarantee of robust stability and performance as long as the true patient sensitivity value is matched by the correct feedback controller (Table 1). Proper operation of the closedloop system hinges, therefore, on achieving adequately accurate identification of the patient’s individual response characteristic.
The inferred parameter probability distribution generated by the particle filter approaches the true p(x(k)Z(0 : k)) for N → ∞[29] and can be deemed a suitable approximation if a sufficiently large amount of particles is used. Therefore, it is fair to expect for the estimate to converge and, thus, for the system to deliver the required performance, asymptotically in time. The presence of a numerical tool in the architecture, however, makes it difficult  if at all possible  for a mathematical proof to be developed. To the best of the author’s knowledge, no theoretical proofs of robust performance exist for the several ‘robust adaptive control’ approaches described in the control literature to date [30], and proofs of asymptotic stability have been developed in restricted cases only (e.g. for lineartimeinvariant Gaussian systems in [31]).
In a general formulation such as we consider here, not only would available proofs not apply, but also asymptotic results would be clinically inadequate as the patient must be successfully identified and controlled over a bounded time horizon. In light of the above, while RACPF is equipped with conservatively designed controllers and very general estimation tools, its ability to maintain stability and deliver the required level of performance can, at this time, only be evaluated heuristically. We do so, in this work, using computer simulations.
Numerical simulations
A broad computational simulation campaign was conducted in order to evaluate the ability of the proposed adaptive control approach to control MAP through SNP administration in a wide variety of conditions. To reflect the unpredictable nature of blood pressure disturbances and patient parameter variations, a large number of cases were randomly generated and simulated. All simulations involved a control horizon of 10,000 s (approximately 2 h and 45 min), with a target MAP of 100 mmHg for the time period of 0 to 4,000 s and 80 mmHg for 4,000 to 10,000 s.
 (a)
Relatively ‘settled’ patients, i.e. displaying elevated MAP (p _{0} = 120 mmHg at t = 0) with p _{dist}(t) modelled as a random, zeromean, lowintensity additive disturbance (in the range ±6 mmHg)
 (b)
More ‘unsettled’ patients, i.e. displaying the same initial MAP as (a) but greater intensity of random fluctuations (in the range ±15 mmHg), as well as two step increases in p _{dist} (t) (+20 mmHg) at 2,000 and 5,500 s, modelling a worsening hypertensive condition
Scenarios of controlled MAP reduction of 20 to 80 mmHg were deemed quite general while clinically plausible.
The p_{dist} (t) signal used in the simulations was generated by filtering the superposition of a Gaussian white noise process and step signals (the latter only for stream b) with a suitable lowpass filter to meet the frequency domain bound assumption of (4). The seeds for the generation of all random processes were also randomised and changed for every simulation: no two simulations in the campaign, therefore, would exhibit the same p_{0}(t), w(t) and v(t) traces.
 (i)
Forty simulations in which midrange parameter values were chosen, i.e. K = 4 mmHg/(ml/h), α = 0.5, and T = 30 s and held fixed throughout the control horizon
 (ii)
Eighty simulations in which the three model parameters K, α, and T were randomly selected from the allowed ranges as per (3) and held constant throughout the control horizon
 (iii)
Eighty simulations in which the initial values of the parameters were randomly selected as in (ii) with piecewise linear variations throughout the control horizon. A random number of slope changes between 2 and 5 was selected for each run. The slope of the ramp was also selected at random each time ensuring it would not exceed the constraints of (3)
 (iv)
Eighty simulations in which the initial values of the parameters were randomly selected as in (ii) with piecewise exponential variations throughout the control horizon. Between 2 and 5 changes were again randomly selected for each run. The exponent was also randomly selected in such a way that the slope of the resulting curve would not exceed the constraints of (3)
 (v)
Eighty simulations in which the initial values of the parameters were randomly selected as in (ii) with piecewise quartersinusoidal variations throughout the control horizon. Between 2 and 5 changes were again randomly selected for each run. The target value for each change was randomly selected in such a way that the slope of the resulting curve would not exceed the constraints of (3)
 1.
Measures concerning control performance

${t}_{{c}_{1}},{t}_{{c}_{2}}$  the convergence time (10% to 90% of transition) as the two MAP setpoint changes are imposed

t_{±5},t_{±10},t_{±15}  the time (out of 10,000 s) y_{meas} remained within ±5, ±10 and ±15 mmHg of the setpoint r(t), respectively

ε_{max},ε_{min},ε_{avg}  the positive and negative peaks and average setpoint tracking error recorded over the control horizon

t_{pp}, t_{npp}  the time the RACPF controller and the true patient system form a provably performant (pp) or nonprovably performant (npp) pair as evaluated through μ analysis [24], i.e. a pp closedloop pair would achieve a performance index A≥1 in Figure 4 (note: due to the inherent conservativeness of μ analysis, a npp pair does not necessarily indicate an unstable or underperforming closedloop condition, but only that there is no mathematical proof of robust performance. At npp times, loop performance should be evaluated from other indices)

t_{sat}  the time the infusion rate signal remained at the maximum allowed value (controller saturation)
 2.
Measures concerning estimation of the response characteristic

t_{0}  time (out of 10,000 s) for which the K subinterval deemed the most likely by the particle filter, i.e. with the greatest π_{ i } (intervals as per best performance in Table 1) contains the true simulated value of K

t_{1}  time for which the K subinterval deemed the most likely by the particle filter is a neighbouring interval to that containing the true value of K

t_{2}  time for which the K subinterval deemed the most likely by the particle filter is 2 intervals away from that containing the true value of K

K_{rel},T_{rel},α_{rel}  the mean relative estimation error for parameters K, T and α, respectively (the mean of the particles is used to determine the estimate)
The simulation environment was programmed using Matlab and Simulink (MathWorks, Natick, MA, USA) and run on a standard desktop computer (Intel Core 2 Quad CPU, 3.0 GHz).
Results
Aggregate results for simulation vb54
Parameter  Result 

t_{0} (s)  5,924 
t_{1} (s)  3,442 
t_{2} (s)  634 
t_{pp} (s)  9,716 
t_{npp} (s)  284 
K_{rel} (%)  81 
α_{rel} (%)  25 
T_{rel} (%)  22 
${t}_{{c}_{1}}$ (s)  186 
${t}_{{c}_{2}}$ (s)  332 
t_{sat} (s)  0 
t_{±5} (s)  6,558 
t_{±10} (s)  9,368 
t_{±15} (s)  9,980 
ε_{max} (mmHg)  16.7 
ε_{min} (mmHg)  14.07 
ε_{avg} (mmHg)  1.08 
Summary of results of the simulation campaign
Batchstream  t _{0}  t _{1}  t _{2}  t _{pp}  t _{npp}  K _{rel}  α _{rel}  T _{rel}  ${t}_{{c}_{1}}$  ${t}_{{c}_{2}}$  t _{sat}  t _{±5}  t _{±10}  t _{±15}  ε _{max}  ε _{min}  ε _{avg} 

[s]  [s]  [s]  [s]  [s]  (%)  (%)  (%)  [s]  [s]  [s]  [s]  [s]  [s]  [mmHg]  [mmHg]  [mmHg]  
ia  604  9,396  0  10,000  0  69  11  24  396  360  0  9,442  9,995  10,000  8.43  8.33  0.16 
(128)  (130)  (0)  (0)  (0)  (5.0)  (3.0)  (5.0)  (46)  (59)  (0)  (104)  (14)  (0)  (0.72)  (0.95)  0.05  
ib  5,986  4,014  0  9,640  360  40  4.0  20  397  332  0  8,752  9,968  9,999  11.00  10.28  0.43 
(788)  (788)  (0)  (512)  (512)  (5.0)  (3.0)  (3.0)  (74)  (72)  (0)  (212)  (147)  (0.6)  (1.46)  (1.18)  (0.06)  
iia  4,926  3,790  1,280  9,826  167  63  3  25  390  375  0  9,356  9,944  9,994  8.56  8.43  0.21 
(4,338)  (3,821)  (2,414)  (252)  (252)  (79)  (24)  (54)  (126)  (101)  (0)  (612)  (146)  (34)  (1.98)  (1.98)  (0.22)  
iib  4,804  4,443  716  9,478  522  42  2  27  392  359  0  8,856  9,907  9,990  11.06  9.85  0.43 
(3,038)  (2,682)  (1,324)  (576)  (576)  (54)  (22)  (57)  (101)  (75)  (0)  (868)  (87)  (40)  (3.24)  (2.98)  (0.25)  
iiia  2,971  5,358  1,617  9,898  102  85  14  10  421  385  0  9,289  9,941  9,992  9.32  8.97  0.24 
(2,297)  (2,166)  (1,911)  (204)  (204)  (73)  (19)  (26)  (120)  (110)  (0)  (435)  (114)  (42)  (2.10)  (2.72)  (0.23)  
iiib  4,710  4,460  731  9,396  604  49  2.0  10  428  380  0  8,710  9,898  9,991  12.03  10.94  0.60 
(1,938)  (1,772)  (882)  (694)  (694)  (39)  (15)  (24)  (124)  (104)  (0)  (612)  (151)  (41)  (2.44)  (2.90)  (0.35)  
iva  2,152  5,802  1,938  9,860  140  77  12  9  408  384  0  9,405  9,974  10,000  8.78  8.51  0.21 
(2,096)  (1,945)  (2,135)  (207)  (207)  (59)  (21)  (25)  (89)  (96)  (0)  (398)  (88)  (15)  (1.86)  (1.72)  (0.15)  
ivb  4,701  4,474  729  9,349  651  45  3.0  11  408  366  0  8,860  9,936  9,995  11.44  10.15  0.50 
(1,730)  (1,744)  (970)  (796)  (796)  (38)  (17)  (25)  (94)  (101)  (0)  (471)  (90)  (31)  (2.15)  (1.95)  (0.21)  
va  2,834  5,350  1,689  9,898  102  94  13  12  419  404  0  9,164  9,930  9,994  9.95  8.95  0.37 
(2,159)  (2,030)  (1,344)  (205)  (205)  (62)  (18)  (25)  (110)  (97)  (0)  (117)  (116)  (26)  (2.30)  (1.71)  (0.36)  
vb  4,529  4,568  848  9,298  702  48  0.0  14  390  405  0  8,450  9,846  9,976  13.15  11.76  0.68 
(1,663)  (1,534)  (882)  (871)  (871)  (31)  (15)  (26)  (107)  (124)  (0)  (674)  (220)  (80)  (3.64)  (3.84)  (0.47) 
The results for batches ia and ib confirm the repeatability of results using the proposed method. The standard deviation values are much lower than for other batches, indicating that the system was able to deliver consistent outcomes, as required, when applied multiple times to the same case. The fact that the standard deviation values are not 0 can be explained in terms of the stochastic features present in both the patient model and the particle filter. As no two simulations are identical, some variability in the results exists even with parameters K, α and T being exactly the same.
When considering the other batches (i to v a/b), the results show that control of MAP was successfully achieved in all cases, with no instability or dangerous pressure drops observed in any of the computed simulations. The controlled MAP trace was maintained within ±10 mmHg of the desired setpoint for over 94% of the time in any given batch of simulations (considering ${M}_{{t}_{\pm 10}}2\sigma $, i.e. 9,404.74 s for the worstperforming batch vb). Similarly, the occurrence of temporary deviations > ±15 mmHg from the setpoint was prevented for over 98% of the time and peak tracking errors (ε_{max},ε_{min}) had a magnitude of less than 20 mmHg at all times. The tracking of the setpoint was achieved with negligible bias as shown by the very low value of ε_{avg} for all batches. It should be clarified that the above statistics refer to the whole simulation (including the initial condition and the two setpoint changes), thus suggesting that control performance remained entirely adequate throughout the simulation campaign. Furthermore, performance was consistent across simulation batches, indicating that the system can deal with timevarying parameters regardless of the shape of the variations, as long as the assumptions of (3) are met. Transition times ${t}_{{c}_{1,2}}$ also met the specifications and were contained below 10 min in almost all cases, with the longest observed transition taking approximately 13 min to settle. As could be expected, due to the simpler nature of the control task (lower disturbance), the performance results for stream a were better than those for stream b, although only marginally.
The results associated with the identification of patient parameters by the particle filter show that the system is tracked as it evolves through time (with t_{0} and t_{1} combined representing about 90% of the total simulation time on average). However, the system slightly but consistently overestimates K and delivers estimation errors K_{rel} of about 50% to 70% on average. The large standard deviation values for α_{rel} and T_{rel}, on the other hand, suggest that the latter two parameters are not dependably identified (there is poor convergence of the particle scatter with respect to α and T). Although the estimation errors may seem significant, the precision in identification is adequate in the context of the required performance, since each controller can cater for variations of two to fourfold in the value of K, and for all possible values of α and T (Table 1). The t_{pp} result confirms this by showing that the system pairs the patient with a suitable controller at most times, with closer inspection of simulation data showing that t_{npp} is accrued mostly in the initial stages when the particle filter has not yet converged. Notably  and perhaps counterintuitively at a first glance  identification of the patient response is more accurate for the more challenging cases of stream b than for stream a as indicated by higher values of t_{0}.
Each simulation took approximately 8 min to execute (corresponding to a simulation timetorealtime ratio of 1:20) using 1,000 particles. This number of particles was found to deliver a reasonable compromise between accuracy in the results and computation time. A number of test simulations conducted ahead of the campaign did not show noticeable improvements in the estimation results when run using 5,000 particles.
Discussion
We have presented RACPF, a novel approach for the control of uncertain, timevarying systems, and tailored it to the case study of automatic closedloop SNP administration for the management of acute hypertension. Having adopted an underlying response model which is, to our knowledge, the most general ever adopted in the literature (with regard to the ranges of variability of parameters, the allowed rates of change and shapes of variations, and the presence of both random and nonzeromean changes in p_{0}(t)), the system delivered the required performance with no evidence of unsafe behaviour throughout the extensive simulation campaign presented herein. We therefore propose that RACPF may represent a viable solution to assist with automated control of hypertension not only in the traditionally considered postoperative setting, but also in the more challenging intraoperative context, where rapid changes in a patient’s response characteristic may occur [32].
Only a limited number of authors have considered automatic control of MAP during surgery (although not specifically with SNP). The challenge of online adaptation to variations in the patient response has generally been addressed through ad hoc rules [33] which mimic physicians’ clinical decision process [34]. The RACPF method is fundamentally different. By adopting a doseresponse model which incorporates uncertainty, μ synthesis allows the designer to conclusively determine a priori the minimum number of controllers required to achieve the desired level of robust performance, while particle filtering tracks the patient’s response as it evolves through time to inform correct controller selection. In addressing the argument that methods for automatic drug administration may have been dismissed on grounds of perceived safety in the past, we deem the transparent nature of RACPF, where an explicit relationship is retained between the estimated posterior probability distribution of the patient response and the chosen control action, to be well suited to clinical applications. A further advantage of this work is that the framework is not specific to SNP and the same methodology could be used to design controllers and estimators, given a reference doseresponse model, for a different drug delivery problem.
The results also support the viability of particle filtering in a realtime pharmacological application. Particle filters are a very general estimation tool and can be applied to linear as well as nonlinear problems, but are known to be computationally onerous. The fact that simulations ran much faster than real time without requiring particular hardware or software optimisation was therefore a notable positive finding. Particle filter methods have been proposed to assist with the estimation of physiological systems and draw inference from patient data in order to support clinical decisionmaking processes (see, e.g. [35]). To our knowledge, however, such methods have been generally reserved for offline computations. Our work highlights the potential of particle filter as a viable tool for realtime closedloop system identification. As pharmacokinetic/pharmacodynamic models commonly combine linear and nonlinear dynamics [22], we anticipate that marginalisation would be applicable to a broad class of problems in this area.
On the point of identifiability of the system, a number of relevant comments can be made. Perhaps counterintuitively, the results show good control performance, but the identification statistics point to somewhat imprecise parameter estimates. As well as being a result of the underdetermined nature of the estimation problem (multiple uncertain parameters and a single output), this outcome stems from a wellknown tradeoff between control and identification in feedback systems [36]. The very purpose of a feedback controller is to suppress some of the dynamics from the controlled system, thus the stronger the control action (performance), the lesser inputoutput information is available to estimate the response characteristic. An extreme situation would be that of a controller capable of forcing a perfect flatline MAP output at all times: in such a situation no knowledge of the patient characteristic could be gathered (arguably, adaptive control would not be required, either). The necessary imposition of performance constraints in our design, therefore, affects the precision of system identification. The fact that the controllers were designed to cater for all values of α and T and, as a result, the identification of these parameters was inaccurate, is a clear example of this tradeoff. Furthermore, accuracy in identification depends on the level of excitation of the system, i.e. the amount of energy delivered by exogenous inputs (in this case, r(t) and p_{0}(t)), which elicits observable dynamics at the output. Since r(t) is mostly a constant signal, excitation is provided mainly by the changes in p_{0}(t). This explains why identification is more accurate (greater t_{0}) for the more disturbed simulations of stream b. It should be remarked that the tradeoff between identifiability and excitation is an inescapable one and affects both traditional manual methods and computerbased solutions equally. A notable question arising in the light of these considerations is whether the commonly stated clinical goal of maintaining a patient in a settled state is a desirable one or whether this is actually counterproductive when changes in a patient’s response must be timely identified and acted upon. Our analysis suggests that a clinically acceptable, timevarying MAP target would deliver greater system excitation and thus improve ongoing system identification.
Conclusion
Through the extensive simulation campaign presented in this manuscript, we have validated, in silico, the methodology underpinning the novel RACPF architecture and demonstrated its applicability to the problem of automatic closedloop administration of SNP for the control of MAP in a model of acute hypertension. These results provide a strong motivation for the approach to be tested further in in vivo models (e.g. animal testing).
We have also proven the viability of realtime computation of particle filters on consumer hardware for the relatively simple SNP doseresponse system. The analysis of other case studies, including highly nonlinear and multipleinputmultipleoutput systems, will be essential to identify any limitations to the practical use of particle filtering in estimating more complex pharmacological response models.
Finally, our analysis has highlighted the known and inescapable tradeoff between identification and control in feedback systems and the role of system excitation in adaptive control. While identification proved adequate for control in the SNP case study, this could only be verified heuristically through simulations. In future work we plan to research mathematical methods to quantify the tradeoff and design clinically acceptable inputs to guarantee dependable system identification, and thus greater theoretical guarantees of robustness, for this and other automatic drug delivery applications.
Authors’ information
NM received the B.E. degree in Biomedical Engineering in 2005 and the M.E. degree in Industrial Engineering (Biomedical) in 2007 from Politecnico di Milano, Milan, Italy. He is currently a lecturer in Digital Systems at The Australian National University, Canberra, Australia, where he has recently completed a Ph.D. degree with a focus on drug delivery, robust adaptive control and cardiovascular modelling.
Declarations
Acknowledgements
The Author would like to thank Dr. Jochen Trumpf for his precious advice in structuring this paper.
Authors’ Affiliations
References
 Cosgrove DM, Petre JH, Waller JL, Roth JV, Shepherd C, Cohn LH: Automated control of postoperative hypertension: a prospective, randomized multicenter study. Ann Thorac Surg 1989, 47(5):678–683. 10.1016/00034975(89)901173View ArticleGoogle Scholar
 Bailey JM, Haddad WM: Drug dosing control in clinical pharmacology. IEEE Control Syst Mag 2005, 25(2):35–51.View ArticleGoogle Scholar
 Parker RS, Doyle FJ: Controlrelevant modeling in drug delivery. Adv Drug Deliv Rev 2001, 48: 211–228. 10.1016/S0169409X(01)001144View ArticleGoogle Scholar
 Haddad WM, Bailey JM: Closedloop control for intensive care unit sedation. Best Pract Res Clin Anaesthesiol 2009, 23: 95–114. 10.1016/j.bpa.2008.07.007MathSciNetView ArticleGoogle Scholar
 Mendonca T, Lemos J, Magalhaes H, Rocha P, Esteves S: Drug delivery for neuromuscular blockade with supervised multimodel adaptive control. IEEE Trans Control Syst Technol 2009, 17(6):1237–1244.View ArticleGoogle Scholar
 Doyle FJ, Bequette BW, Middleton R, Ogunnaike B, Paden B, Parker RS, Vidyasagar M: Control in biological systems. In The Impact of Control Technology. Edited by: Samad T, Annaswamy AM. Piscataway: IEEE Control Systems Society; 2011.Google Scholar
 Varon J, Marik PE: Perioperative hypertension management. J Vasc Health Risk Manage 2008, 4(3):615–627.Google Scholar
 Slate JB, Sheppard LC, Rideout VC, Blackstone EH: A model for design of a blood pressure controller for hypertensive patients. In Proceedings of the 1st Annual Conference, IEEE Engineering in Medicine and Biology Society, Denver, Colorado. Piscataway: IEEE; 1979.Google Scholar
 Mansour NE, Linkens DA: Poleassignment selftuning control of blood pressure in postoperative patients: a simulation study. IEEE Proc Control Theory Appl 1989, 136: 1–11. 10.1049/ipd.1989.0001MATHMathSciNetView ArticleGoogle Scholar
 Pajunen GA, Steinmetz M, Shankar R: Model reference adaptive control with constraints for postoperative blood pressure management. IEEE Trans Biomed Eng 1990, 37(7):679–687. 10.1109/10.55674View ArticleGoogle Scholar
 He WG, Kaufman H, Roy R: Multiple model adaptive control procedure for blood pressure control. IEEE Trans Biomed Eng 1986, 33: 10–19.View ArticleGoogle Scholar
 Martin JF, Schneider AM, Smith NT: Multiplemodel adaptive control of blood pressure using sodium nitroprusside. IEEE Trans Biomed Eng 1987, 34(8):603–611.View ArticleGoogle Scholar
 Ying H, McEachern M, Eddleman DW, Sheppard LC: Fuzzy control of mean arterial pressure in postsurgical patients with sodium nitroprusside infusion. IEEE Trans Biomed Engs 1992, 39(10):1060–1070. 10.1109/10.161338View ArticleGoogle Scholar
 Yu C, Roy RJ, Kaufman H, Bequette BW: Multiplemodel adaptive predictive control of mean arterial pressure and cardiac output. IEEE Trans Biomed Eng 1992, 39(8):765–778. 10.1109/10.148385View ArticleGoogle Scholar
 Yu YC: Blood pressure, automatic control of. In Encyclopedia of Medical Devices and Instrumentation. Edited by: Webster JG. New York: Wiley; 2006:490–500.Google Scholar
 Chitwood WR, Cosgrove DM, Lust RM: Multicenter trial of automated nitroprusside infusion for postoperative hypertension. Ann Thorac Surg 1992, 54(3):517–522. 10.1016/00034975(92)90446BView ArticleGoogle Scholar
 Bequette BW: Modeling and control of drug infusion in critical care. J Process Control 2007, 17(7):582–586. 10.1016/j.jprocont.2007.01.015View ArticleGoogle Scholar
 Malagutti N, Dehghani A, Kennedy RA: Improved robust performance in a system for automatic administration of a vasoactive drug. In Proceedings of BIOSTEC Biosignals Conference, Vilamoura, Portugal. Montreal: SciTePress; 2012:282–290.Google Scholar
 Malagutti N, Dehghani A, Kennedy RA: Robust control design for automatic regulation of blood pressure. IET Control Theory Appl 2013, 7(9):387–396.MathSciNetView ArticleGoogle Scholar
 Fekri S, Athans M, Pascoal A: Issues, progress and new results in robust adaptive control. Int J Adaptive Control Signal Process 2006, 20(10):519–579. 10.1002/acs.912MATHMathSciNetView ArticleGoogle Scholar
 Malagutti N, Hassani V, Dehghani A: An adaptive multicontroller architecture using particle filtering. In Proceedings of the 2012 Australian Control Conference, Sydney, Australia. Barton: Engineers Australia; 2012:392–398.Google Scholar
 Bonate P: PharmacokineticPharmacodynamic Modeling and Simulation. Berlin: Springer; 2011.View ArticleGoogle Scholar
 Malagutti N, Dehghani A, Kennedy RA: A robust multiplemodel adaptive control system for automatic delivery of a vasoactive drug. Proceedings of 9th IASTED Conference on Biomedical Engineering 2012, 74–79.Google Scholar
 Skogestad S, Postlethwaite I: Robust stability and performance analysis for MIMO systems. In Multivariable Feedback Control, Analysis and Design. New York: Wiley; 2006:309–368.Google Scholar
 Cappe O, Godsill SJ, Moulines E: An overview of existing methods and recent advances in sequential Monte Carlo. Proc IEEE 2007, 95(5):899–924.View ArticleGoogle Scholar
 Schon T, Gustafsson F, Nordlund PJ: Marginalized particle filters for mixed linear/nonlinear statespace models. IEEE Trans Signal Process 2005, 53(7):2279–2289.MathSciNetView ArticleGoogle Scholar
 Doucet A, de Freitas N, Gordon N: An introduction to sequential Monte Carlo methods. In Sequential Monte Carlo Methods in Practice. New York: Springer; 2001:3–13.View ArticleGoogle Scholar
 Douc R, Cappe O: Comparison of resampling schemes for particle filtering. In Proceedings of the 4th International Symposium on Image and Signal Processing and Analysis. Faculty of Electrical Engineering and Computing, University of Zagreb, Croatia; 2005:64–69.Google Scholar
 Doucet A, Johansen AM: A Tutorial on Particle Filtering and Smoothing: Fifteen Years Later, Volume 12. Oxford: Oxford University Press; 2009.Google Scholar
 Ioannou PA: Robust adaptive control: she search for the Holy Grail. In Proceedings of the 47th IEEE Conference on Decision and Control. Piscataway: IEEE; 2008:12–13.Google Scholar
 Hassani V, Hespanha J, Athans M, Pascoal AM: Stability analysis of robust multiple model adaptive control. In Proceedings of The 18th IFAC World Congress. Milan: IFAC; 2011.Google Scholar
 Meijers RH, Schmartz D, Cantraine FR, Barvais L, dH́ollander AA, Blom JA: Clinical evaluation of an automatic blood pressure controller during cardiac surgery. J Clin Monit Comput 1997, 13(4):261–268. 10.1023/A:1007389631055View ArticleGoogle Scholar
 Furutani E, Araki M, Kan S, Aung T, Onodera H, Imamura M, Shirakami G, Maetani S: An automatic control system of the blood pressure of patients under surgical operation. Int J Control, Automation, Syst 2004, 2: 39–54.Google Scholar
 Hoeksel SA, Blom JA, Jansen JR, Maessen JG, Schreuder JJ: Automated infusion of vasoactive and inotropic drugs to control arterial and pulmonary pressures during cardiac surgery. Crit Care Med 1999, 27(12):2792–2798. 10.1097/0000324619991200000031View ArticleGoogle Scholar
 Zenker S, Rubin J, Clermont G: From inverse problems in mathematical physiology to quantitative differential diagnoses. PLoS Comput Biol 2007, 3(11):e204. 10.1371/journal.pcbi.0030204MathSciNetView ArticleGoogle Scholar
 Forssell U, Forssell U, Ljung L, Ljung L: Closedloop identification revisited. Automatica 1997, 35: 1215–1241.MathSciNetView ArticleGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License(http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.