You are on page 1of 6

2015 European Control Conference (ECC)

July 15-17, 2015. Linz, Austria

Automatic recovery from nonlinear oscillations in PID-controlled


anesthetic drug delivery
Zhanybai T. Zhusubaliyev1 , Margarida M. Silva,2 and Alexander Medvedev2

Abstract The problem of recovering from nonlinear oscil- question on whether typical nonlinear phenomena such as
lations in a PID-controlled drug delivery system for the neuro- sustained oscillations may appear. In fact, there is clinical
muscular blockade (NMB) in closed-loop anesthesia is consid- evidence of oscillatory behavior in the literature for cases
ered. The NMB system dynamics portrayed by a Wiener model
can exhibit sustained nonlinear oscillations for physiologically of closed-loop drug delivery in anesthesia, see e.g. [4]. The
feasible values of the model parameters and realistic PID gains. results of [5] provide mathematical support to the clinical
Such oscillations, also repeatedly observed in clinical trials, findings. The bifurcation analysis in [5] shows that, for
lead to under- and over-dosing of the administered drug and the case of Proportional-Integral-Derivative (PID)-controlled
undermine patient safety. This paper proposes an bifurcation NMB, complex nonlinear behaviors such as sustained os-
analysis-based tuning policy for the proportional PID gain that
ensures automatic recovery from those oscillations. It is shown cillations may indeed appear under certain conditions, as
that with a time-varying proportional PID gain, the type of a consequence of Andronov-Hopf bifurcation phenomena.
equilibria of the closed-loop system remains the same as in the These oscillations result in under- and over-dosing of anes-
case of constant controller gains and no additional types of thetics, and should be avoided. An expression for distance to
complex nonlinear behaviors are introduced by these auxiliary bifurcation is also suggested in [5], giving rise to a way of
dynamics. The recovery time is also evaluated in simulation
over a database of individualized patient models. monitoring how far the closed-loop system is from sustained
nonlinear oscillations. Even though a PID controller design
I. INTRODUCTION based on an individualized Wiener model [5] can initially
ensure a safe distance to bifurcation, the model parameters
Even though several attempts have been made to com- sometimes drift close to a critical zone as the surgery
pletely automatize the intravenous drug delivery to patients proceeds. The latter has been demonstrated on clinical data
under general anesthesia, see e.g. [1, Ch.4], no technique has in [6] by estimating the distance to the Andronov-Hopf
yet become widely accepted in the daily clinical practice. bifurcation by means of a particle filter.
The mismatch between research and practice is mainly due The main idea of the present paper is to ensure that the
to legal regulations. In fact, the chances of approval for closed-loop operating point of the drug delivery system is
any closed-loop control strategy by the regulatory entities moved away from the region of oscillatory dynamics. This
are likely to increase with a demonstration of a clinically is guaranteed by systematically changing the controller gains.
significant safety benefit with respect to the manual mode The structure of the nonlinear plant is exploited to calculate
[2], such as a reduction of the incidence of underdose-related the point where oscillations arise. The bifurcation analysis of
awareness, or of overdose-related post-operative complica- the nonlinear closed-loop system is instrumental in ensuring
tions. a bumpless transfer from the oscillatory unstable behavior to
For the case of the neuromuscular blockade (NMB), stability.
whenever a patient is underdosed, the patients muscles might The first contribution of this paper is hence the design of
unexpectedly contract in response to the surgical stimulation. a tuning procedure for a PID controller that ensures that the
This, in turn, might constrain the surgeons actions or even closed-loop system recovers from sustained oscillations in an
cause severe bodily harm at the moment of the unexpected autonomous way. This retuning guarantees that the system
contraction. Overdoses may, on the other hand, bring poten- recovers its reference tracking capabilities. Inspired by the
tial delays in the postoperative recovery and a risk of residual case of linear systems, where it is usually enough to lower the
blockade in the post-anesthetic period [3], that is highly loop gain to recover asymptotic stability [7], the proportional
undesired both for physiological and economical reasons. gain in the PID controller is here chosen to be time-varying.
The effect of drugs in anesthesia is an intrinsically non- Considering that the nonlinear case is more complex than
linear dynamical process. The incorporation of a controller the linear one, the second contribution of this paper is the
in the loop adds further nonlinearities, e.g. due to actuator nonlinear analysis of the impact of the time-varying gain on
saturation and sensor characteristics, as well as rises the the type of equilibria of the system. This auxiliary dynamical
*This work was supported by the European Research Council, Advanced
variable increases the order of the closed-loop system by
Grant 247035 (SysTEAM). one during the recovery period. It is then expected that the
1 Department of Computer Science, South West State University, 50 Years
complexity of the system dynamics as well rises. The third
of October Str., 94, 305040, Kursk, Russia zhanybai@hotmail.com contribution is a study of the system dynamics in the time
2 Department of Information Technology, Up-
psala University, SE-751 05 Uppsala, Sweden domain during the transition from oscillation to stability over
{margarida.silva,alexander.medvedev}@it.uu.se a database of individualized patient models.

978-3-9524269-3-7 2015 EUCA 2725


Regarding the novelty of the proposed approach, one has with
to take into account the abundance of the literature on
adaptive (self-tuning, time-varying, etc.) PID control [8]. It
K(t) = (K(t) K ) . (4)
is really difficult to claim that something is new in this
area. The matter of bumpless PID control is also studied
extensively with wide industrial applications in e.g. process Here e(t) = yr y(t) is the control error and yr is
control. Yet, the authors failed to find a single paper covering the reference. The initial condition for the gain K(0) = 0
oscillation recovery in closed-loop drug delivery, which topic is assumed to correspond to weakly damped closed-loop
is in the focus of the present contribution. This application oscillations and the value Kref that is chosen from the interval
area is quite different from traditional process control. For of variation for K is calculated to yield a safe distance to
instance, oscillation detection methods [9] based on signal bifurcation, for instance by using the PID design procedure
analysis are not applicable here. Indeed, when oscillations in proposed in [5]. The constant adjusts the transition rate
closed-loop drug delivery are evident, under- and over-dosing and is chosen in the PID design procedure.
of the patient has already occurred. Therefore, continuous The automatic PID-tuning in the present paper is part of
assessment of the risk for oscillation by e.g. the calculation the following scenario in the surgery room:
of distance to bifurcation from online data, as in [6], appears
1) A Wiener individualized NMB model expressed by (1),
to be a more suitable option.
(2) is estimated from a bolus response to the drug in
II. CLOSED-LOOP MODEL the beginning of surgery. Based on the model, desired
controller convergence rate, distance to bifurcation,
The results in [5] show that the minimally param-
and the gains of the PID controller K0 , Ti , Td are
eterized parsimonious (MPP) model for the NMB is
calculated by e.g. the design method in [5].
able to capture the same intrinsic nonlinear dynamics
2) The surgery proceeds under automatic muscle relaxant
of the system, e.g. limit cycles and sustained oscilla-
delivery. The Wiener model parameters , and the
tions, as a standard physiologically-based pharmacokinetic-
distance to bifurcation are estimated on-line by means
pharmacodynamic (PK/PD) model. Also, it is suitable for
of a suitable identification algorithm, e.g. the particle
individualization under limited exogenous system excitation,
filter or the Extended Kalman Filter, see e.g. [6].
as it is usually present in a real-life clinical scenario. This
3) When the estimated distance to bifurcation goes under
supports the use of the MPP models in the design of
a given threshold, a re-design of the PID controller
controllers for drug delivery in anesthesia.
is undertaken for the current estimated values of the
A state-space description of the linear dynamics of the
Wiener model to yield the same distance to bifurcation
MPP Wiener model [10] is given by
that the original design but for the current values of the
x 1 = k3 (x1 u(t)), x 2 = k2 (x1 x2 ), model parameters. This results in the PID controller
x 3 = k1 (x2 x3 ), (1) settings Kref , Ti , Td .
4) The recovery controller expressed by (3), (4) is ac-
where x1 and x2 are nonnegative auxiliary state variables, tivated for transferring the closed-loop system to the
and x3 , also nonnegative, is the output of the linear dynam- foreseen oscillation-free mode.
ical block of the Wiener model.
The system input (atracurium administered rate) u(t) is The mathematical model of the closed-loop NMB in (1)-
nonnegative and bounded 0 6 u(t) 6 umax , > 0 is the (4), can now be recast as a set of autonomous differential
patient-dependent parameter, while ki , i = 1, 2, 3 are positive equations [5]
constants. In the following analysis, the values chosen in [10]
k1 = 1, k2 = 4 and k3 = 10 are assumed. dx
= f (x), (5)
In the same way as in the PK/PD model, the effect of dt
the drug, as quantified by the measured NMB y(t) [%], is x = (x1 , x2 , x3 , x4 , x5 )T ,
modeled by the Hill function f (x) = (f1 , f2 , f3 , f4 , f5 )T ,

100 C50
y(t) = . (2)
C50 + x3 (t)

with
Several sensors are available in the clinical practice to
quantify the NMB in patients. For a brief summary see e.g. f1 = k3 x1 2 k1 k3 Td x5 0 (x3 )(x2 x3 )
[1, Ch.2]. k3
+ k3 x5 (yr (x3 )) + x4 x5 ,
A. Time-varying PID controller Ti
f2 = k2 (x1 x2 ), f3 = k1 (x2 x3 ),
Following [5], the control signal in (1) is assumed to be
generated by a PID controller f4 = yr (x3 ), f5 = (x5 K ),

1
Z
de(t)
 x1
u(t) = K(t) e(t) + e(s) ds + Td , (3) 0 (x3 ) = 3 2
(x3 ).
Ti dt 100 C50

2726
B. Bifurcation analysis caught by the nonlinear constraints. This kind of oscillations
The system (5) has a single equilibrium state x = typically arises in continuous smooth systems with negative
[x1 , x2 , x3 , x4 , x5 ]T , where feedback regulation when the loop gain and/or the delay in
the feedback becomes sufficiently high [12].
Ti For the case of PID-controlled drug delivery, the normal
x1 = x2 = x3 , x4 = x ,
K 3 operational conditions are the regime of a stable equilibrium
 1
state. However, due to the subject uncertainty, the equilib-

100
x3 = C50 1 , x5 = K . rium point may lose stability. As a result, a stable limit cycle
yr
softly arises from the equilibrium state [5].
The local stability of the equilibrium x is determined by the The eigenvalues si , i = 1, 5 for the Jacobian matrix are
eigenvalues of the Jacobian matrix determined by
 
fi det (Df(x ) sI) =
Df(x ) = , (6)
xj 16i65; 16j65 = (s4 + b1 s3 + b2 s2 + b3 s + b4 )(s + ) = 0.
where Considering that one eigenvalue of (6) is equal to < 0, for
f1 the equilibrium state to be stable, the following inequalities
= k3 , have to be satisfied
x1
f1 bi > 0, i = 1, 4, b1 b2 b3 > 0, (7)
= 2 k1 k3 K Td 0 (x3 ),
x2 b1 b2 b3 b21 b4 b23 > 0. (8)
f1
= k3 K 0 (x3 ) + 2 k1 k3 K Td 0 (x3 ), Note that bi > 0, i = 1, 4 are positive. Numerical experi-
x3
f1 k 3 K f1 k3 x4 k3 x3 ments show that the Jacobian matrix has one or two pairs of
= , = = , complex conjugate eigenvalues for the considered range of
x4 Ti x5 Ti K
f2 f2 parameters. The factor (s+) in characteristic polynomial (6)
= k2 , = k2 , indicates a relative independence (in small) of the controller
x1 x2
gain rate from the plant parameters.
f2 f2 f2
= = = 0, An Andronov-Hopf bifurcation occurs when the second
x3 x4 x5 condition (8) is violated. The surface
f3 f3 f3 f3
= = = 0, = k1 , (Ti , Td , K , , ) = b23 b1 b2 b3 + b21 b4 = 0 (9)
x1 x4 x5 x2
f3 f4 defines the boundary of the stability region for the equilib-
= k1 , = 0, i = 1, 2, 4, 5,
x3 xi rium state, see [5].
f4 f5 Substituting b1 , b2 , b3 and b4 in into (9), an expression
= 0 (x3 ), = 0, i = 1, 4,
x3 xi for the bifurcation surface that depends explicitly on the
f5 parameters Ti , Td , K , , and is as follows
= ,
x5 (Ti , Td , K , , ) =
and k1 k2 k3 [Td (k1 + k2 + k3 ) 1] (0 (x3 ))2 K2
1 1
+ k1 k2 k3 (k1 + k2 + k3 )2 Td

yr2

100
0 (x3 ) = 1 .
100 C50 yr + (k12 k2 + k12 k3 + k1 k22 + k1 k2 k3 + k1 k32 + (10)
The eigenvalues of Df(x ) can either be real or complex + k22 k3
+ k2 k32 )
(k1 + k2 + k3 )2 /Ti 0 (x3 )K

and their total number is equal to the dimension of the
phase space. The real part of an eigenvalue determines the (k1 + k2 )(k1 + k3 )(k2 + k3 ) = 0.
rate of growth of the corresponding perturbation away from
the equilibrium point, and the imaginary part determines the III. NUMERICAL SIMULATIONS
angular frequency of an oscillatory component. In order to assess the performance of the proposed time-
An equilibrium point is stable if all eigenvalues have neg- varying PID controller, numerical simulations were designed
ative real part, corresponding to an exponentially decreasing as follows:
response to perturbations. However, if an eigenvalue or a pair Given nominal values of (0 , ), calculate the PID
of complex conjugated eigenvalues crosses the imaginary controller gains Ti , Td , K0 with e.g. the Ziegler-
axis into the positive complex half-plane, the equilibrium Nichols step-response tuning procedure [8];
point becomes unstable. The transition in which a pair of Simulate the closed-loop system for the time interval
complex conjugated eigenvalues simultaneously crosses the 0 t < t1 in order to assess the time-domain response
imaginary axis from the left to the right complex half-plane of the closed-loop system;
is known as an Andronov-Hopf bifurcation [11]. This is the Given (Ti , Td , K0 , ) calculate the bifurcation value
transition that leads to emergence of self-sustained oscilla- = bif as the real positive root of equation
tions provided that the expanding oscillatory dynamics are (Ti , Td , K0 , , ) = 0 (10) with respect to ;

2727
At t = t1 , choose = 1 as a 5% perturbation of bif , affected by the value of Kref = Kbif than by the value of
so that the operating point is in the region of instability; , as expected from (6).
Simulate the closed-loop system for the time interval To exemplify the arising behaviors, a patient model from
t1 t < t2 in order to assess the time-domain response the patients database in [10], [14] was randomly chosen.
of the clsoed-loop system to this change; The parameters of the patient model are 0 = 0.0364
Given nominal values of (Ti , Td , 1 , ) calculate Kbif and = 4.24358. The PID gains for the given model are
as a the real negative root of (Ti , Td , K, 1 , ) = 0 Ti = 11.72262, Td = 2.930656 and K0 = 1.30697.
(10) with respect to K; Fig. 1a presents a bifurcation surface
Choose (Ti , Td , K , , ) = 0, (9) in the PID-controller
Kref = Kbif , (11) parameter space (Ti , Td , K ) calculated for this patient
with R+ , so that the system is in the region of model. In the diagram, the domain of stability for the
stability: (Ti , Td , Kref , 1 , ) = 0. equilibrium point is denoted by A. The space on the left
At t = t2 , apply K(t) from (4) in order to automatically
of surface (9), denoted by B, corresponds to the domain
recover stability of the closed-loop system; of existence of the self-sustained oscillatory dynamics. As
Simulate the closed-loop system for the time interval
mentioned above, this oscillatory mode appears through an
t > t2 in order to assess the time-domain response of Andronov-Hopf bifurcation.
the system to this change. The point denoted as 1 corresponds to the parameter values
Note that, in a real scenario, the time instant t2 occurs as Ti = 11.72262, Td = 2.930656 and K0 = 1.30697.

soon as the calculated distance to the bifurcation phenomena K0 was chosen to be less than Kref = 0.07816 because,
is smaller than a user-defined threshold, as mentioned in for this combination of parameters, (Ti , Td , K, , ) <

Sec. II-A. 0, K < Kref . As shown in this figure, Point 1 belongs
The previously described numerical simulations were run to the region A of stability for the equilibrium point. The
for a database of patient models ((i) , (i) ), i = 1, 41, arrow represents the direction in the parameter space along
obtained from clinical data [10]. The analysis presented here which the bifurcation surface is moved as the parameter
is intended to demonstrate the feasibility of the proposed ap- decreases from 0 = 0.0364 to 1 = 0.021435. Fig. 1b
proach for the automatic recovery from nonlinear oscillations shows a bifurcation surface for = 0.021435 after this
in a real application. It is important to recall that oscillations transition. As one can see from this figure, Point 1 moves
have been observed in real cases, see e.g. [13], and that now to the region of instability B.
the current practice is to switch the system to manual mode Fig. 2a and Fig. 2c illustrate in the time domain the effect
and to titrate the anesthetic dose manually afterwards. In an that this perturbation in the parameter has on the system.
automatic safety framework, it is recommended that the risk The time interval t < t1 corresponds to the case when Point 1
for oscillation is assessed online by e.g. the calculation of is in the stability region A. For simulation purposes, the
distance to bifurcation as in [6], and the re-tuning approach variables x1 , x2 and x3 in (1) are initialized at zero, and
proposed in this paper is applied before actual oscillations yr is made equal to 10. For t < t1 , the PID controller
arise. stabilizes the system around the reference yr , as shown in
The results of the numerical simulations for the database the time profile of x3 and y, in the middle and upper plots
of 41 models and different values of in (11) and in (4) of Fig 2a, respectively. The parameter is decreased to 1
are shown in Table I. at t1 , causing the bifurcation surface in Fig. 1a to slide to
the right (see Fig. 1b). Fig. 2d presents a magnification of
TABLE I: Mean values of the recovery time tc > t2 : t > the bifurcation surface in Fig. 1b for 1 = 0.021435. As
tc , x3 (t) I, with I = [0.9x3 , 1.1x3 ] for all 41 cases (i = seen from this figure, Point 1 now is in the region B of the
1, 41) in the simulation database. The values in brackets are instability for the equilibrium state. As illustrated by the time
[min, max]. diagrams for x3 and y in Fig. 2a, this results in the oscillatory
dynamics and the trajectories of the dynamical system (5)
tc [min]
converge to a limit cycle in the time interval t1 t < t2 .
((i) , (i) ) 0.1 0.01 6.321 [4.399, 7.980] As the parameter K is increased from K0 = 1.30697
0.1 0.1 4.422 [3.185, 7.008] to Kref 0.07816 (as shown in the bottom plot of Fig. 2c),
0.5 0.1 9.353 [4.475, 14.437]
while maintaining other parameters constant, Point 1 again
((i) , ) 0.1 0.1 4.075 [2.460, 5.674] becomes in the region of stability A. The scanning direction
is indicated in Fig. 2d by the arrow. The observed sequence
(, (i) ) 0.1 0.1 4.104 [ 2.379, 4.874]
of bifurcation transitions is shown in Fig. 2b. The upper
plot of Fig. 2b depicts the bifurcation diagram illustrating
As shown in Table I, the average recovery time tc > the transition from a limit cycle to an equilibrium point as
t2 : t > tc , x3 (t) I, with I = [0.9x3 , 1.1x3 ] for all the gain K increases from 1.30697 (Point 1 in Fig. 2d)
41 cases (i = 1, 41) in the simulation database is best for to 0.07816 (point 2 in Fig. 2d). The bottom plot shows
the combination x = 0.1 and = 0.1. Also, results indicate the variation of the real part of the complex-conjugated
that the behavior of the output during the recovery is more eigenvalues s1,2 = i.

2728
(a) (b)
Fig. 1: (a) Andronov-Hopf bifurcation boundary in the parameter space (Ti , Td , K ) for = 0.0364 and = 4.24358: A
is the domain of stability for an equilibrium state and B is the region where an equilibrium point is unstable. The point
1 corresponds to the parameter values: Ti = 11.72262, Td = 2.930656 and K0 = 1.30697 and it belongs to the region
of stability denoted as A. The arrow represents the direction in the parameter space along which the bifurcation surface is
moved as parameter decreases from 0.0364 to 0.021435. (b) Andronov-Hopf bifurcation boundary in the parameter space
(Ti , Td , K ) for = 0.021435 and = 4.24358. Now the point 1 becomes in the region of instability B.

For 1.30697 < K < Kbif , the system displays an to recover from oscillations that were previously observed in
unstable equilibrium state surrounded by a stable oscillatory clinical trials and that lead to under and overdosing of the
state (limit cycle). For this parameter interval, the Jacobian administered drug. By introducing a time-varying propor-
of system (5) has a pair of complex-conjugated eigenvalues tional PID gain, as shown via a bifurcation analysis of the
s1,2 = i with positive real parts > 0 (see Fig. 2b), and closed-loop, the type of equilibria remains the same as in the
three real and negative eigenvalues s3 < 0, s4 < 0, s5 < 0. case of constant controller gains. This is desirable since no
At K = Kbif , the equilibrium state undergoes a reverse additional complex nonlinear behavior types are introduced
Andronov-Hopf bifurcation. When the parameter K passes by this tuning. The recovery time after applying the proposed
through the point Kbif , a pair of complex-conjugated eigen- time-varying PID gain is evaluated in simulation to exemplify
values s1,2 = i crosses the imaginary axis into the the performance of the proposed strategy. For a realistic
negative real half-plane, and the equilibrium state becomes database of cases, the average recovery time is 4.22 min,
stable. which shows the feasibility of the PID re-tuning procedure.
This increase in the proportional gain K(t) of the PID
controller is simulated to occur at time t2 in Fig. 2c and R EFERENCES
Fig. 2a. The bottom plot of Fig. 2c shows the time-profile
[1] M. M. Silva, Nonlinear modeling and feedback control of drug deliv-
of the gain K(t), from K0 to Kref , following (4). The time- ery in anesthesia, Ph.D. dissertation, Uppsala dissertations from the
behavior of the state variable x3 and the output y in Fig. Faculty of Science, Acta Universitatis Upsaliensis, Uppsala University,
2c show that, for t t2 , the oscillations disappear and the 2014.
[2] P. Manberg, C. Vozella, and S. Kelley, Regulatory challenges facing
stable equilibrium behavior is recovered. The bottom plot of closed-loop anesthetic drug infusion devices, Clinical Pharmacology
Fig.2a is a zoom into the middle plot in the same figure for & Therapeutics, vol. 84, no. 1, pp. 166169, 2008.
t t2 . The amplitude of the oscillations decay, yielding a [3] T. Gilhuly, B. MacLeod, G. Dumont, A. Bouzane, and S. Schwarz,
Improved neuromuscular blockade using a novel neuromuscular
recovery time of 4.37 min. This value is close to the mean blockade advisory system: A randomized, controlled, clinical trial,
in the database for the same values of and , as shown in Anesthesia & Analgesia, vol. 107, no. 5, pp. 16091617, Nov. 2008.
Table I. [4] W. Haddad, K. Volyanskyy, J. Bailey, and J. Im, Neuroadaptive
output feedback control for automated anesthesia with noisy EEG
IV. CONCLUSIONS measurements, IEEE Transactions on Control Systems Technology,
vol. 19, no. 2, pp. 311326, Mar. 2011.
This paper presented a PID-tuning policy to automatically [5] Z. T. Zhusubaliyev, A. Medvedev, and M. M. Silva, Bifurcation
recover from undesired nonlinear oscillations in a drug deliv- analysis of PID-controlled neuromuscular blockade in closed-loop
anesthesia, Journal of Process Control, vol. 25, pp. 152 163, 2015.
ery system for the neuromuscular blockade (NMB) in anes- [6] O. Rosen, M. M. Silva, and A. Medvedev, Nonlinear estimation of a
thesia. The proposed tuning allows the closed-loop system parsimonious Wiener model for the neuromuscular blockade in closed-

2729
100
y 50
yr
0
0 t1 400 t2 800 1000
20
x3

10
0 t1 t2
0 400 800 1000
15
x3

10
5
590 t2 610 620
Time [min]
(a) Time-domain behavior of x3 and output y. The red dashed
line indicates 1.2 x3 , with x3 as the steady state value of the state (b) Bifurcation diagram and variation of the real part
variable x3 for t > t2 . of the complex-conjugated eigenvalues s1,2 = i
for 1.30697 K 0.07816 and = 0.021435.

0
0.03

1
0 t1 400 t2 800 1000
0
Kref
K

0.8
K0
0 t1 400 t2 800 1000
0
Kref
K

0.8
K0
590 t2 610 620
Time [min]
(c) Time-domain changes in and K, for = 0.1. (d) Zoom of the bifurcation boundary in Fig.2b for
The red square depicts the value of Kbif . 0 = 0.021435 and = 4.24358.

Fig. 2: Simulation results for K0 = 1.30697, Kref = 0.9Kbif = 0.07816, 0 = 0.0364, and 1 = 0.021435.

loop anesthesia, in Proc. 19th IFAC World Congress, Cape Town, [12] J. Laugesen and E. Mosekilde, Emergence of oscillatory dynamics,
South Africa, Aug. 2014, pp. 92589264. in Biosimulation in Biomedical Research, Health Care and Drug De-
[7] om and R. M. Murray, Feedback Systems: An Introduction for
K. J. Astr velopment, E. Mosekilde, O. Sosnovtseva, and A. Rostami-Hodjegan,
Scientists and Engineers. Princeton, NJ, USA: Princeton University Eds. Springer Vienna, 2012, pp. 6995.
Press, 2008. [13] J. A. Mendez, S. Torres, J. A. Reboso, and H. Reboso, Adaptive
[8] om and T. Hagglund, PID Controllers: Theory, Design, and
K. J. Astr computer control of anesthesia in humans, Computer Methods in
Tuning, 2nd ed. Research Triangle Park, NC: Instrument Society of Biomechanics and Biomedical Engineering, vol. 12, no. 6, pp. 727
America, 1995. 734, 2009.
[9] T. Hagglund, A control-loop performance monitor, Control Engi- [14] C. Rocha, T. Mendonca, and M. Eduarda Silva, Modelling neuro-
neering Practice, vol. 3, no. 11, pp. 1543 1551, 1995. muscular blockade: a stochastic approach based on clinical data,
[10] M. M. Silva, T. Wigren, and T. Mendonca, Nonlinear identification of Mathematical and Computer Modelling of Dynamical Systems, vol. 19,
a minimal neuromuscular blockade model in anesthesia, IEEE Trans. no. 6, pp. 540556, 2013.
Contr. Syst. Techn., vol. 20, no. 1, pp. 181188, 2012.
[11] J. Guckenheimer and P. Holmes, Nonlinear Oscillations, Dynamical
Systems, and Bifurcations of Vector Fields (Applied Mathematical
Sciences Vol. 42). New York: Springer, 1997.

2730

You might also like