IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 54, NO.

1, JANUARY 2024 597

Positive Impulsive Control of Tumor

Therapy—A Cyber-Medical Approach
Levente Kovács , Senior Member, IEEE, Tamás Ferenci , Member, IEEE, Balázs Gombos, András Füredi,
Imre Rudas , Life Fellow, IEEE, Gergely Szakács, and Dániel András Drexler , Member, IEEE

Abstract—Chemotherapy optimization based on mathematical algorithms are compared to a generic protocol based on over-
models is a promising direction of personalized medicine. all survival. The experimental results show that the introduced
Personalizing, thus optimizing treatments, may have multiple algorithms significantly increased the overall survival of the mice,
advantages, from fewer side effects to lower costs. However, per- demonstrating that by control engineering methods an efficient
sonalization is a complicated process in practice. We discuss a modality of cancer therapy may be possible.
mathematical model of tumor growth and therapy optimization
algorithms that can be used to personalize therapies. The therapy Index Terms—Impulsive system, min–max therapy, optimal
generation is based on the concept of keeping the drug level over a treatment, positive system, therapy generation, tumor model.
specified value. A mixed-effect model is used for parametric iden-
tification, and the doses are calculated using a two-compartment
model for drug pharmacokinetics, and a nonlinear pharmaco-
I. I NTRODUCTION
dynamics and tumor dynamics model. We propose personalized
YBER-MEDICAL systems play an important role in
therapy generation algorithms for having a maximal effect and
minimal effective doses. We handle inter- and intra-patient vari-
ability for the minimal effective dose therapy. Results from mouse
C modern medicine, and their importance is growing. The
application of STEM in medicine offers prosperous results
experiments for the personalized therapy are discussed and the
in medical practice. For example, engineering methods can
be used for brain fatigue detection [1], [2], [3], prediction
Manuscript received 10 May 2023; accepted 8 September 2023. Date of of in-hospital death of trauma [4], skeleton maturity assess-
publication 20 September 2023; date of current version 19 December 2023. ment [5], [6], or Parkinson’s disease diagnosis [7]. System-
This work was supported in part by the European Research Council (ERC) theoretic methods are used in several drug dosing problems,
through the European Union’s Horizon 2020 Research and Innovation
Programme under Grant 679681; in part by the National Research, like control of anesthesia [8], or control of blood glucose level
Development and Innovation Fund of Hungary, financed under the 2019- with artificial pancreas [9].
1.3.1- Funding Scheme under Project 2019-1.3.1-KK-2019-00007; in part System-theoretic methods can also be utilized to optimize
by the Hungarian National Research, Development and Innovation Fund of
Hungary, financed under the TKP2021-NKTA-36 Funding Scheme and under drug dosing in cancer therapies. The therapies used in con-
Grant Horizon2020-2017-RISE-777911; and in part by the Collaboration ventional chemotherapy usually have a large resting time,
Between the Research Centre for Natural Sciences of the Eötvös Lóránd i.e., a long time between the injections and large injected
Research Network and the Szentágothai Research Centre of the University
of Pécs on Internationally Recognized Medical Research Projects. The work doses [10]. They use the maximum tolerable dose (MTD) in
of Dániel András Drexler was supported by the Starting Excellence Researcher order to achieve a maximal effect without killing the patient.
Program of Óbuda University, Budapest, Hungary. The work of András Füredi Another approach is the low-dose metronomic (LDM) ther-
was supported by the HORIZON.1.2—Marie Skłodowska-Curie Actions
(MSCA) Postdoctoral Fellowship (POC-TDM) under Grant 101065044. This apy, which applies lower doses with larger frequency. In some
article was recommended by Associate Editor C. Platania. (Corresponding cases, this approach was proven more effective against cancer
author: Dániel András Drexler.) cells, which often become resistant to the treatment [11], [12].
Levente Kovács, Imre Rudas, and Dániel András Drexler are with the
Physiological Controls Research Center, University Research and Innovation LDM therapy can also be cheaper with fewer side effects. We
Center, Óbuda University, 1034 Budapest, Hungary (e-mail: drexler.daniel@ aim to provide algorithms for the mathematical model-based
uni-obuda.hu). generation of LDM therapies.
Tamás Ferenci is with the Physiological Controls Research Center,
University Research and Innovation Center, Óbuda University, 1034 Budapest, Scheduling LDM therapy and providing the required doses
Hungary, and also with the Department of Statistics, Corvinus University of is a challenging task. A promising engineering approach is
Budapest, 1093 Budapest, Hungary. to use a mathematical model describing the effect of the
Balázs Gombos is with the Drug Resistance Research Group, HUN-REN
Research Centre for Natural Sciences, 1117 Budapest, Hungary, and also with drug on tumor growth and use this model to generate an
the Molecular Medicine PhD School, Semmelweis University, 1085 Budapest, optimal therapy. There are numerous models in the literature
Hungary. (see [10], [13], [14], [15] and several therapy generation algo-
András Füredi is with the Drug Resistance Research Group, HUN-REN
Research Centre for Natural Sciences, 1117 Budapest, Hungary, and also rithms [10], [16], [17], [18]). A specific characteristic of this
with the Microsystems Laboratory, HUN-REN Centre for Energy Research, physiological problem is that the input is the injection, which
1121 Budapest, Hungary. is positive, and the system is impulsive. Such systems are
Gergely Szakács is with the Center for Cancer Research, Medical University
of Vienna, 1090 Vienna, Austria, and also with the Drug Resistance Research rare in engineering practice, and thus handling them requires
Group, HUN-REN Research Centre for Natural Sciences, 1117 Budapest, unconventional solutions [19], [20], [21]. Besides therapy gen-
Hungary. eration, the usage of nanorobots in cancer treatments is also
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TSMC.2023.3315637. exploited by Shi et al. [22], [23], [24], while robotic capsules
Digital Object Identifier 10.1109/TSMC.2023.3315637 are used for site-specific drug delivery in [25].

c 2023 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
For more information, see https://creativecommons.org/licenses/by/4.0/
598 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 54, NO. 1, JANUARY 2024

Pérez-García et al. [10] used their model and heuristic

scheduling based on in silico tests. Tse et al. considered
multidrug optimization problems in [26]. Cacace et al. [18]
solved the therapy generation as an optimization problem to
minimize a quadratic cost function combined with a state
observer. They generate antiangiogenic therapy based on the
model created by Hahnfeldt et al. [27]. The authors of this
manuscript created an optimization algorithm for the antian-
giogenic therapy based on the Hahnfeldt model in [16], where
the aim was to track a reference path that results from max-
imal dosing with a predefined maximal deviation from that
path. The authors used an approach based on pharmacokinetic Fig. 1. Procedure of the therapy optimization: the MIC is specified based
on a chosen strategy, and the minimal doses are calculated, which ensure that
(Section II-C) and pharmacodynamic (Section II-B) model to the drug level is over MIC.
optimize chemotherapy in [17] and [28], based on the results
of Kusuoka et al. [29]. Here, we develop the results further version of the second one, which uses the minimal and max-
and provide several model-based approaches for the therapy imal values of the parameters for a specific population and
generation problem. finds the minimal dosages to ensure that the tumor volume
The fundamental component of optimization is the under- does not increase for the worst-case parameter combination
lying mathematical model. Our algorithm depends on the (Section III-D). This strategy is based on interval arithmetics,
pharmacokinetics and pharmacodynamics of the drug; how- briefly summarized in Section II-F. One can switch between
ever, several strategies discussed here also depend on the tumor the strategies based on the tumor response automatically, or
dynamics. The literature on tumor models is rich; see the the switching can be done manually by an expert (Fig. 1).
works of [30], [31], and [32]; in our work, we use a tumor We carry out in vivo experiments with 29 mice with breast
model described by ordinary differential equations. cancer (Section IV) and show with a log-rank test, that our
We use a fourth-order model (Section II-A) created to algorithm significantly increases overall survival compared to
describe measurements from animal experiments first for a generic therapy.
antiangiogenic therapy using bevacizumab [33], [34], and
later for chemotherapy using pegylated liposomal doxorubicin II. M ETHODS
(PLD) in [35] and [36] based on the experiments from [37].
The latest model has four state variables, two state variables to The mathematical methods used for the therapy
describe the living and dead tumor volume dynamics, and two optimization are based on the tumor model discussed in
state variables to describe the pharmacokinetics of the drug Section II-A. The pharmacodynamics of the model is detailed
as a two-compartment model. We formulate the optimization in Section II-B, which is needed to calculate the drug level
problem as maintaining a predefined drug level during the ther- that has to be maintained during the therapy. The required
apy with the lowest amount of injections, i.e., we carry out dosages are calculated based on the pharmacokinetics of the
optimal impulsive control of the two-compartment pharma- model discussed in Section II-C. Based on the pharmacoki-
cokinetic model and propose three strategies to calculate this netic model, the minimal dosages required to keep the drug
predefined drug level. level over a specified limit are calculated as a solution to an
Impulsive control of compartment systems is not new in optimization problem discussed in Section II-D. In order to
the literature; e.g., the work of Pierce and Schumitzky from personalize the therapy, identification is carried out with a
1976 [38]. Kusuoka et al. [29] created an algorithm in 1981 for mixed effect model given in Section II-E. Interval arithmetics
finding minimal doses that keep the drug level over a specified is reviewed in Section II-F, which will be used to design
lower limit, discussed in Section II-D. Our work is based on robust therapy when only the model parameter ranges are
this algorithm and contributes by defining the lower limit for known.
the drug level, which we will call minimal inhibitory concen-
tration (MIC) after [39] (see Fig. 1). We test the algorithms A. Tumor Model
in vivo using mouse experiments discussed in Section IV-A. The therapy generation algorithm will be based on a fourth-
We propose three strategies to define the lower limit of the order model describing tumor dynamics, pharmacodynamics,
drug level: two personalized therapies and one robust ther- and pharmacokinetics [35], [36], defined by the equations
apy created for a specific population. The first strategy is a x1 x3
personalized therapy aiming to maximize the effect of the ẋ1 = (a − n)x1 − b (1)
ED50 + x3
drug using low dosages and only uses the pharmacokinetic and x1 x3
pharmacodynamic model (Section III-B). The second strategy ẋ2 = nx1 + b − wx2 (2)
ED50 + x3
is a personalized therapy, which finds the minimal effective ẋ3 = −(c + k1 )x3 + k2 x4 (3)
dosages that ensure that the tumor volume does not increase
ẋ4 = k1 x3 − k2 x4 (4)
during the therapy for constant and time-varying (but known)
parameters (Section III-C). This strategy requires knowledge where x1 , x2 , x3 , and x4 are the time functions of the living
of the tumor dynamics as well. The third strategy is the robust tumor volume, dead tumor volume, drug level in the central
KOVÁCS et al.: POSITIVE IMPULSIVE CONTROL OF TUMOR THERAPY—A CYBER-MEDICAL APPROACH 599

TABLE I
N OTATIONS , NAMES , AND D IMENSIONS OF THE PARAMETERS of the therapy optimization will be to keep the drug level in
OF THE T UMOR G ROWTH M ODEL the central compartment over the MIC.
One possible strategy, described in Section III-B, is to
define the limit as a constant multiple of the ED50 param-
eter, i.e., MIC = κED50 . If κ is sufficiently large, the value
of the Hill function is close to 1, i.e., close to the maximal
effect of the drug. For example, for the value of κ = 100, the
value of the Hill function is 0.99; thus, the drug has at least
99% effect during the therapy.
Another strategy is to calculate the MIC such that the drug
prevents the tumor from growing with the lowest dosages
during the therapy, which we discuss in Section III-C for per-
sonalized treatment, in the case of intrapatient variability for
known model parameters, and as a worst-case treatment for a
population to treat interpatient variability in Section III-D.

C. Pharmacokinetic Model
The pharmacokinetics of the model described by (3) and (4)
is a linear time-invariant system. The input of the system is
positive and impulsive and has an effect on x3 as described
by (5). Since the drug level in the central compartment (x3 )
has a direct effect on the tumor as described by the term on
the right-hand side of (1), the output yp of the pharmacoki-
netic model is x3 . Thus, the state-space representation of the
pharmacokinetic model is
      
ẋ3 −c − k1 k2 x3 1
compartment (the blood), and drug level in the peripheral com- = + u (7)
ẋ4 k1 −k2 x4 0
partment (the tissues), respectively. The unit of the volumes  
is mm3 , while the drug levels are in mg·kg−1 , compatible with A

the unit of the injected doses. The parameters of the model yp = x3 (8)
are positive and are listed in Table I. with u being the sum of impulsive inputs that is written using
The injections are considered as impulsive effects on the Dirac-delta distributions δ [20], i.e.,
central compartment x3 , i.e., let the K number of injections K−1
take place at time instants tk ≥ 0, k = 0, 1, 2, . . . , K − 1 with u(t) = uk δ(t − tk ) (9)
t0 < t1 < · · · < tK−1 and injected doses uk ≥ 0, then there is k=0
a discontinuity of the first kind in x3 at time tk , such that at t ≥ 0, where K is the total number of injections, tk ≥ 0,
    k = 0, 1, . . . , K −1 is the time of injections with doses uk ≥ 0,
x3 tk+ = x3 tk− + uk . (5)
k = 0, 1, . . . , K − 1, and δ is the Dirac delta distribution, i.e.,
The output of the system is the total tumor volume δ(t) = 0, if t = 0 (10)
∞
y = x1 + x2 (6)
δ(τ )dτ = 1. (11)
since x1 and x2 cannot be measured separately in the −∞
experiments, only the total tumor volume, which is their sum. The output of the pharmacokinetic subsystem (7) produced
from the impulsive inputs (9) at time t can be written as the
B. Pharmacodynamic Model sum of impulse responses w of the system as
The pharmacodynamics of the drug is defined by the Hill K−1

function x3 (ED50 + x3 )−1 in (1) and (2), which is a common yp (t) = w(t − tk )uk . (12)
function used to model the effect of the drug [8], [40]. This k=0
function expresses that the effect of the drug is saturated, and In order to avoid ambiguity, we note that from now on, we
after a given limit, increasing the drug level yields a very use the letter w to denote the impulse response of the pharma-
low increase in the drug effect. The median effective dose cokinetic subsystem (and not the washout rate from Table I,
parameter ED50 is the drug concentration where the effect is which we will not need in the remaining of this article). The
50%, i.e., the value of the Hill function is 0.5. impulse response of the pharmacokinetic subsystem at time
The desired value of the drug level in the central compart- t ≥ 0 is
ment that should be maintained by the therapy discussed in λ1 + k2 λ1 t λ2 + k2 λ2 t
w(t) = e + e (13)
Section II-D depends on the value of ED50 . An important goal λ1 − λ2 λ2 − λ1
600 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 54, NO. 1, JANUARY 2024

where λ1 and λ2 are the eigenvalues of the system matrix A parameters for a subject—mouse—are random variates from a
in (7) and can be expressed with the parameters as distribution characterized by their mean and variance (and typ-
ically assumed to be normal), and the focus of the estimation
−(c + k1 + k2 ) ± (c + k1 + k2 )2 − 4ck2 lies in these population parameters. These approaches cap-
λ1,2 = . (14)
2 ture the intraindividual correlations of repeated measured data
The pharmacokinetic subsystem is kinetic, thus it is posi- and can be considered as a middle ground between estimating
tive [41], [42], [43], [44], which implies that the impulse individual parameters and one global parameter set.
response of the system is also positive for all t ≥ 0. Individual variates are obtained similarly to residuals. In
Theorem 1: The pharmacokinetic subsystem (7) is asymp- fact, a mixed model can be written (with a single level of
totically stable and nonoscillatory [28]. grouping) as
yi = Xi β + Zi bi + ε i
D. Optimal Impulsive Therapy
One way to optimize impulsive therapy is to find the where yi is the response variable, Zi describes the grouping
minimal injection doses in order to maintain a predefined structure, bi represents the so-called random effects, while
drug level in the patient, i.e., look for the optimal injection β is the vector of fixed effects, with Xi being the usual—
doses uk given at time instants tk , k = 0, 1, . . . , K − 1 such fixed effects—design matrix, and εi being the usual error term.
that x3 (t) ≥ MIC for all t ≥ 0. This problem has been Typically, bi ∼ N (0, ) and ε i ∼ N (0, σ 2 I) is assumed [45].
addressed by Kusuoka et al. [29], who formulated and solved In the present case, yi means the measured tumor volumes, Xi
this optimization problem for compartmental systems with a and Zi contain the times of the measurements, while β col-
constraint that the input should also be positive. lects the parameters of the differential equations (with bi being
Let u = (u0 , u2 , . . . , uK−1 ) and 1 = (1, 1, . . . , 1) be a their associated random effects). The likelihood implied by the
column vector with elements of one and length K, and let  above model can be calculated, and maximum-likelihood esti-
be the matrix of impulse responses constructed as mation is readily possible (although often other variants are
  used).
 = w ti − tj−1 i,j (15) A further complication in such parameter identification
problems is that the model is specified through a system
where i, j = 1, 2, . . . , K, and tK is a time instant after the last
of differential equations from which no explicit formula can
injection, i.e., tK > tK−1 . The purpose of tK is to maintain the
be derived for the response variable. The stochastic approx-
predefined drug level for a certain time after the last injection.
imation expectation–maximization (SAEM) approach can be
The goal is to have yp (t) ≥ MIC for all t ≥ 0, where MIC is
used, which is an extension of the classic EM algorithm [46],
the desired lower limit for the drug level, with the constraint
widely used to carry out maximum-likelihood estimation
that the injections are positive, and the goal of minimizing
(locally) when the model depends on unobserved parameters,
the total amount of injections. Since the pharmacokinetic sub-
as described by Delyon et al. [47]. This is widely used to esti-
system is positive, asymptotically stable, and nonoscillatory,
mate models described by differential equations with mixed
yp (t) ≥ MIC for all t ≥ 0 is equivalent to yp (tk ) ≥ MIC for
effects [48].
all tk , k = 0, 1, . . . , K. Thus, the optimization problem can be
Calculations in the in vivo experiments given in Section IV
written as
of this study are carried out under R statistical environment
minimize 1 u version 4.1.0 [49] using the nlmixr package version 2.0.4 [50].
u
subject to u ≥ MIC · 1
F. Interval Arithmetics
u ≥ 0. (16)
Interval arithmetics is an efficient tool to analyze expres-
The solution to this optimization problem [29] is sions in which the variables are perturbed, but their lower and
upper limits are known. We will use interval arithmetics in
ũ = MIC · −1 1. (17) Section III-D to calculate the min–max therapy in the pres-
This constrained optimization problem minimizes the sum ence of parametric perturbations, i.e., calculate the minimal
of the drug dosages u (i.e., the cumulative dose) with the con- doses that ensure that the drug has inhibiting/killing effect all
straints such that the drug level in the central compartment is the time in the worst-case parameter combinations. We will
not less than the MIC at the time of the injections (before the use the notations and definitions from the work of Alefeld
drug is injected), and the drug dosages are non-negative. and Mayer [51].
Let a denote the lower limit of the variable a, and a denote
E. Mixed-Effect Model the upper limit of the variable a, and let [a] denote the closed
interval of the possible values of a, i.e.,
The application of the therapy generation algorithms  
requires the knowledge of the model parameters, i.e., prior [a] = a, a . (18)
parametric identification has to be carried out. If the mod-
Let ◦ be a binary operation on intervals, such that ◦ ∈
eled population has similar intrinsic and extrinsic parameters,
{+, −, ·, /} defined as the set
then the mixed effect model is an efficient tool for paramet-
ric identification [45]. In brief, such models assume that actual [a] ◦ [b] = {a ◦ b, a ∈ [a], b ∈ [b] } (19)
KOVÁCS et al.: POSITIVE IMPULSIVE CONTROL OF TUMOR THERAPY—A CYBER-MEDICAL APPROACH 601

where 0 ∈/ [b], if the operation is division. The results of the Steady-state 1 is the origin, which is ideally the goal of
operations are the intervals the therapy. Steady-state 2 is a nonzero equilibrium that is
  achieved if the drug effect and the tumor dynamics are bal-
[a] + [b] = a + b, a + b (20) anced. Note that in steady-state 2, the variables can be zero,
 
[a] − [b] = a − b, a − b (21) but in this case, we get steady-state 1.
 
[a] · [b] = min ab, ab, ab, ab . (22) The Jacobian of (30) and (31) is
 
a − n − E∗ 0
In the case if a > 0 and b > 0, we have that [a]·[b] = [ab, ab]. J= (34)
n + E∗ −w
The multiplicative inverse is defined as
   with the eigenvalues
1 1 
= b ∈ [b] , if 0 ∈
/ [b] (23) λ1 = a − n − E ∗
[b] b (35)
λ2 = −w. (36)
which can be calculated as
1  −1  Since the parameters are positive, we have that λ2 < 0. In
= b , b−1 . (24) steady-state 2, we have λ1 = 0, thus the steady-state is on the
[b]
line x2∗ = (a/w)x1∗ = 0, if initially x1 (0) > 0. Steady-state 1 is
The standard interval functions are ψ ∈ F = asymptotically stable if λ1 < 0, i.e., if E∗ > a − n. Note that
{sin, cos, tan, arctan, exp, ln, abs, sqrt}, which are defined by we are interested in the case when a − n > 0, i.e., the tumor
their range as grows without therapy.
ψ([x]) = {ψ(x), x ∈ [x]}. (25) Theorem 2: If E(t) > a − n for all t ≥ 0, then the origin of
the model (30), (31) is asymptotically stable.
Functions composed of standard interval functions and binary Proof: Let E1 and E2 be time functions that satisfy that
(1)
operations can be used in interval arithmetics based on the E2 > E1 > 0, i.e., E2 (t) > E1 (t) > 0 for all t ≥ 0, and let x1
definitions. (2)
be the solution to (30) with input E1 and x1 be the solution
If a function depends on an interval variable more than once, with input E2 to (30) with initial condition x1 (0) > 0. Due to
then we should use interval variables defined as the convex the positivity of x1 and E, we have that ∀t ≥ 0
combination of their lower and upper limits [52], i.e.,
(a − n)x1 (t) − x1 (t)E2 (t) < (a − n)x1 (t) − x1 (t)E1 (t) (37)
ã = αa + (1 − α)a, α ∈ [0, 1]. (26)
thus x1(1) (t) > x1(2) (t) whenever t > 0. Thus, the tumor dynam-
Let ψ be a function of a, then ics model is monotonous in E. Let E1 > a−n with E1 ≡ const,
  and E2 be an arbitrary function with E2 (t) > E1 for all t ≥ 0.
[ψ] = ψ, ψ (27) Since E1 > a − n, we have that a − n − E1 < 0, thus the
solution to (30) with initial condition x1 (0)
with the lower limit defined as the solution of
(1)
x1 (t) = e(a−n−E1 )t x1 (0), t≥0 (38)
ψ = min ψ(ã) (28)
α tends to zero; thus, the origin is asymptotically stable. Due
and the upper limit as the solution of to the monotonicity, x1(2) (t) < x1(1) (t) for all t > 0, thus the
solution with E2 also converges to zero, since if E2 > E1 , then
ψ = max ψ(ã) (29)
(2)
α
x1 (t) < e(a−n−E1 )t x1 (0), t≥0 (39)
with α constrained to the interval [0, 1].
where the right-hand side tends to zero whenever
E1 > a − n.
III. M AIN R ESULTS Thus, if the net effect E is kept over the limit a−n during the
A. Tumor Dynamics whole therapy, then the total tumor volume tends to zero. For
Consider the first two equations (1) and (2) of the tumor impulsive systems with linear pharmacokinetics and nonlinear
model and denote the drug effect by E = bx3 (ED50 + x3 )−1 . pharmacodynamics, the effect of the drug can be kept over a
These equations describe the tumor dynamics specified limit with the method discussed in Section II-D. The
following section will provide strategies to define this specific
ẋ1 = (a − n)x1 − Ex1 (30) limit. In the original model E = bx3 (ED50 + x3 )−1 , thus we
ẋ2 = nx1 + Ex1 − wx2 . (31) will provide limit for x3 denoted as MIC.
Since the maximum of x3 (ED50 + x3 )−1 is 1, the direct
The equilibrium points of this model are the solutions to consequence of Theorem 2 follows.
  Theorem 3: Suppose that a − n > 0. There exists x3 such
0 = a − n − E∗ x1∗ (32)
  that the origin of the model (30), (31) is asymptotically stable
0 = n + E∗ x1∗ − wx2∗ . (33) if and only if b > a − n.
The solutions are as follows. Proof: Suppose that b > a−n. In the original model, we use
1) x1∗ = 0 and x2∗ = 0. x3 (t)
2) E∗ = a − n, and x2∗ = (a/w)x1∗ . E(t) = b . (40)
ED50 + x3 (t)
602 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 54, NO. 1, JANUARY 2024

Thus, in the applied tumor model, the effect E is in the interval method is the high level of toxicity that may result from the
[0, b), provided that x3 is always positive. Since large doses. In order to avoid toxicity, the application of lower
x3 doses that still have inhibiting/killing effect may be desirable.
lim b =b (41)
x3 →∞ ED50 + x3
C. Personalized Therapy With Minimal Drug Level
we can always find a function x3 which satisfies Theorem 2.
Now, suppose that the origin of the model is asymptotically One can also calculate MIC as the minimal drug level that
stable. We show indirectly that it implies b > a − n. Suppose ensures that the tumor volume does not increase between injec-
that b < a − n, and the origin is asymptotically stable. Thus, tions. This can be calculated in a personalized fashion if the
we have the limits model parameters of the patient are known. After the MIC
  is acquired, the algorithm defined in Section III-C is used to
x3
lim ẋ1 = lim a − n − b x1 generate the required doses.
x3 →∞ x3 →∞ ED50 + x3 Theorem 4: The therapy ensures that the tumor does not
= (a − n − b) x1 (42)
  grow using minimal doses if the drug level is kept over
>0
  MIC = κED50 , where
x3 a−n
lim ẋ1 = lim a − n − b x1 κ≥ . (46)
x3 →0 x3 →0 ED50 + x3 b+n−a
= (a − n)x1 . (43)
Proof: The drug level ensuring that the tumor does not grow
Since the function x3 (ED50 + x3 )−1 is strictly monotonously can be calculated based on (1), by considering that with the
increasing and a − n − b < a − n due to the positivity of the optimized therapy x3 (t) ≥ MIC for all t ≥ 0, thus the net
parameters, the x1 solution to (1) with initial condition x1 (0) growth rate is
is bounded as x3 (t) MIC
a−n−b ≤a−n−b (47)
e(a−n−b)t x1 (0) ≤ x1 (t) ≤ e(a−n)t x1 (0) (44) ED50 + x3 (t) ED50 + MIC
which implies that if we ensure that the net growth rate is
where e(a−n−b)t tends to infinity due to (42); thus, the origin smaller or equal to zero at the time of the next injection (prior
is unstable, which leads to a contradiction. to the injection), then the net growth rate is smaller than zero
Due to the nonlinearity of E, the fourth-order model between the injections since the PK model is asymptotically
described by (1)–(4) can have rich dynamics; it was shown stable without oscillatory behavior by Theorem 1 [29]. Thus,
in [53], that if the therapy is considered to be continuous, the tumor volume does not increase if we have
i.e., we add u to the state x3 , and the applied control method
MIC
is P-type control based on the total tumor volume x1 + x2 , a−n−b ≤ 0. (48)
which is ED50 + MIC
From this equation, we can express
u = −k(x1 + x2 ) (45)
ED50
MIC ≥ (49)
then the closed-loop system has a nontrivial equilibrium, and b
the system can have bifurcations with realistic parameter val- −1
a−n
ues. Moreover, the qualitative property of the equilibrium is
and calculate the minimum value of κ by dividing the right-
independent of k.
hand side with ED50 .
This result can be used to generate the minimal dose therapy
B. Personalized Therapy With Maximal Effect that ensures that the tumor does not grow during the therapy.
The critical parameter of the therapy design is the mini- However, this approach only works if the parameters do not
mal drug level (denoted by MIC) that has to be maintained. change during the therapy.
If the MIC is known, the algorithm discussed in Section II-D Now consider the case of intrapatient variability of the
only requires the knowledge of the pharmacokinetic parame- parameters, i.e., when the tumor model parameters change
ters, and can be used to calculate the minimal doses. The value during the therapy. Suppose that the parameters are con-
of MIC can be acquired intuitively or based on the pharma- stant between injections and can only change at the time of
codynamic parameters. For example, choosing MIC = κED50 injections. Let ai , ni , bi , and EDi50 denote the value of the
with κ sufficiently large ensures high efficiency for the ther- parameters a, n, b, and ED50 prior to the ith injection with
apy. The choice of κ = 100 results in 99% efficiency based i = 0, 1, . . . , K − 1, respectively. Let wi denote the impulse
on the pharmacodynamic model. response function of the PK model with the PK model param-
The advantage of the maximal effect is that it is robust eter values prior to the ith injection, and x3i (0− ) be the drug
against parametric perturbations due to oversized doses. The level at the time of the ith injection, without the injected dose.
other advantage is that we only need to know the pharmacoki- Note that we will use t = 0 as the time of the ith injection.
netic (PK) and pharmacodynamic (PD) parameters and do not Let di denote the dose of the ith injection and Ti denote the
require knowledge about the other tumor model parameters. time between the ith injection and (i+1)th injection. The drug
Thus, the method can also be applied to other tumor mod- level in the central compartment at time Ti is composed of the
els with similar PK and PD models. The disadvantage of the drug level due to the ith injection with a value of di wi (Ti )
KOVÁCS et al.: POSITIVE IMPULSIVE CONTROL OF TUMOR THERAPY—A CYBER-MEDICAL APPROACH 603

and the drug level due to the previous injections, denoted by while the second eigenvalue is
x̃3 (Ti ), which can be calculated as
⎛ ⎞    2
i−1 i − c̃ + k̃1 + k̃2 − c̃ + k̃1 + k̃2 − 4c̃k˜2
x̃3 (Ti ) = dj wj ⎝ Tk ⎠ (50) λ̃2 = (59)
2
j=0 k=j
and the impulse response in the new variables is
if i > 0, and x̃3 (T0 ) = 0. Thus, the drug level in the central
compartment at time Ti is λ̃1 − k̃2 λ̃2 − k̃2
w̃(Ti ) = eλ̃1 Ti + eλ̃2 Ti . (60)
x3 (Ti ) = di w (Ti ) + x̃3 (Ti ).
i
(51) λ̃1 − λ̃2 λ̃2 − λ̃1
In the worst-case PK parameter combination, the drug depletes
Theorem 5: The therapy ensures that the tumor does not
fast thus we need the w(Ti ) lower limit for the impulse
grow with minimal doses in the case of changing parameters
response at time Ti after the ith injection, which can be
if the drug level is calculated as
 i  calculated by solving the minimization problem
a − ni EDi50 − x̃3 (Ti )
di ≥   . (52) minimize w̃(Ti )
wi (Ti ) bi + ni − ai αc ,αk1 ,αk2

Proof: Between the ith and (i + 1)th injections, we have to subject to αc , αk1 , αk2 ∈ [0, 1]. (61)
ensure that the net growth rate is not positive, which can be This problem can be solved using numerical optimization
ensured if we have for systems with nonlinear PK equations as well by replac-
x3 (Ti )
ai − ni − bi ≤0 (53) ing the impulse response function with the solution of the
ED50 + x3 (Ti )
i
nonlinear differential equations depending on the uncertain
for all i = 0, 1, . . . , K − 1. Substituting (51) yields parameters.
The worst-case value of the drug level in the central
di wi (Ti ) + x̃3 (Ti )
ai − ni − bi ≤0 (54) compartment Ti time after the ith injection is
EDi50 + di wi (Ti ) + x̃3 (Ti )
x3 (Ti ) = di wi (Ti ) + x̃3 (Ti ) (62)
from which we can express di to get (52).
with
D. Min–Max Therapy for Uncertain Parameters ⎛ ⎞
i−1 i
In the case of unknown parametric perturbations, we pro- x̃3 (Ti ) = dj wj ⎝ Tk ⎠. (63)
vide the minimal drug level dosing described in Section III-C, j=0 k=j
considering the worst-case scenario of parameter changes. This
way, the therapy can handle interpatient variabilities in a robust Theorem 6: The tumor volume does not grow during the
manner. Let a, b, n, and ED50 be the lower limits, while a, b, n, therapy for the worst-case parameter combination if the
and ED50 be the upper limits of the parameters a, b, n, and applied dose is
 
ED50 , respectively. Let c, k1 , and k2 be the lower limits, while a − n ED50 − x̃3 (Ti )
c, k1 , and k2 be the upper limits of the parameters c, k1 , and di ≥   . (64)
w(Ti ) b + n − a
k2 , respectively. Let di denote the drug dose of the ith injection
and Ti denote the time between the ith injection and (i + 1)th Proof: In order to guarantee that the tumor does not grow
injection. between two injections, we must have
The time evolution of the drug level is characterized by x3 (Ti )
the impulse response of the PK system, which has to be ana- a−n−b ≤0 (65)
ED50 + x3 (Ti )
lyzed with interval analysis to calculate the worst-case value
of the impulse response at time Ti . Let this worst-case value be where we supposed that the current injection takes place at
denoted by w(Ti ), which is the lowest value considering all the t = 0, and the next injection occurs at t = Ti . Then, the value
parameter combinations. Since the variables appear more than of the left-hand side of (65) is in the interval
one time in the expression (13), we must define the interval [1, 1]
variables [a] − [n] − [b] (66)
[ED50]
[1, 1] +
c̃ := cαc + c(1 − αc ) (55) [x3 (Ti )]
  [1, 1]
k̃1 := k1 αk1 + k1 1 − αk1 (56) = [a, a] − [n, n] − [b, b]  
  ED50
ED50
k̃2 := k2 αk2 + k2 1 − αk2 (57) [1, 1] + ,
x3 (Ti ) 3 (Ti )
x
as the convex combinations of the lower and upper limits of  
the parameters, i.e., we have αc , αk1 , αk2 ∈ [0, 1]. Then, the := ϕ, ϕ . (67)
first eigenvalue of the system matrix is
 
The therapy ensures that the tumor does not grow in the worst
 2
− c̃ + k̃1 + k̃2 + c̃ + k̃1 + k̃2 − 4c̃k˜2 case if
λ̃1 = (58) ϕ ≤ 0. (68)
2
604 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 54, NO. 1, JANUARY 2024

Application of interval arithmetics discussed in Section II-F

yields that
x3 (Ti )
ϕ =a−n−b . (69)
x3 (Ti ) + ED50

Substituting (62) into (69) yields

di wi (Ti ) + x̃3 (Ti )
ϕ =a−n−b . (70)
di wi (Ti ) + x̃3 (Ti ) + ED50
Rearranging this equation to di and combining it with the
inequality (68) yields the result.
Fig. 2. Overall survival of the mice (in days). Time is calculated from the
second injection, and the survival probability is calculated as the ratio of the
IV. I N V IVO T EST OF THE R ESULTS living mice. The p value is the significance resulting from the log-rank test
of the control group and the groups S1 and S2.
A. Experimental Setup
The results were tested on a clinically relevant, geneti- tumor volume was still over 200 mm3 . Otherwise, the treat-
cally engineered mouse model of breast cancer. In this model, ment stopped until the tumor trigger, i.e., when the volume
Brca1, a DNA repair gene, and p53, a regulator of cell cycle reached 200 mm3 again. This protocol uses large doses (the
and genome stability, were knocked out in breast epithe- maximal tolerable dose of PLD is 8 mg·kg−1 ) with relatively
lial cells. The resulting mammary tumors highly resemble large resting times (i.e., not less than ten days).
the Brca1-linked, triple-negative, hereditary breast cancer in The mice in the experiments are genetically identical; thus,
humans: the molecular, immunohistochemical, morphological, similar model parameters are expected. They were implanted
and genetic characteristics are almost indistinguishable from tumor pieces derived from the same tumor, and their treatment
their human counterpart [54]. started when their tumor first reached 200 mm3 . The tumor
Moreover, these tumors respond to chemotherapy very width and length were measured using calipers three times
similarly; initial treatment with doxorubicin, docetaxel, or cis- a week (on Monday, Wednesday, and Friday), and the tumor
platin significantly reduces tumor size and induces remission. volume was approximated using the formula [33]
However, long-term therapy often fails due to the emergence π
of drug resistance [55]. Despite we showed that PLD increases V = (width · length)(2/3) . (71)
3
relapse-free and overall survival by 6- and 3-fold, respectively,
these tumors cannot be cured using conventional chemother-
apy regimens [37]. Findings obtained by using this model B. Results of the Experiment
are frequently translated to human cancer clinics due to their The tumors reacted well to the first dose (4 mg·kg−1 in the
similarity to human breast cancer. case of group S1, and 6 mg·kg−1 in the case of group S2),
The ideal therapy is personalized therapy, which is tailored and they shrank to a small volume (tumor remission). When
to the specific patient. This requires knowledge of the patient- the tumors started to grow back (tumor relapse) and reached
specific model parameters. However, the identification of the 200 mm3 again, we initiated a protocol based on our results
parameters requires several measurements, which is a large presented here in groups S1 and S2. The mice received two
drawback of the method. The measurements can be used to injections per week, one on Tuesday and one on Friday. We
carry out the parametric identification using a mixed-effect set 6 mg·kg−1 as the maximal dose that can be given in one
model described in Section II-E. The acquired parameters single injection.
can be used to design optimal therapy using the algorithm At the time of the first relapse, we carried out para-
described in Section II-D. metric identification for the mice in groups S1 and S2
There were 29 mice divided into three groups. based on the previous measurements using the algorithm in
1) Control group (C): Eight mice were treated with a Section II-E. We generated maximal effect therapies as given
standard protocol. in Section III-B using κ = 100, i.e., the minimal drug level
2) Group S1: Eleven mice, received 4 mg·kg−1 drug when was 100 times the median effective dose, resulting in 99%
the tumor first reached 200 mm3 , then the therapy was effect of the drug.
generated using the methods in Section III after tumor We repeated the parametric identification on day 82 of
relapse. the experiment. Based on expert knowledge, we changed the
3) Group S2: Ten mice, received 6 mg·kg−1 drug when treatment strategy based on the following heuristics.
the tumor first reached 200 mm3 , then the therapy was 1) If the tumor volume was still over 200 mm3 , we contin-
generated using the methods in Section III after tumor ued applying the maximal effect therapy. If the resulting
relapse. doses were larger than 6 mg·kg−1 with the new param-
The control group (C) treated with the standard protocol eters, then we decreased the parameter κ to 10, i.e., the
received 6 mg·kg−1 PLD, when the tumor reached 200 mm3 new minimal drug level was ten times the median
volume, and the injections were repeated every 10 days if the effective dose, resulting in 90% efficiency.
KOVÁCS et al.: POSITIVE IMPULSIVE CONTROL OF TUMOR THERAPY—A CYBER-MEDICAL APPROACH 605

Fig. 3. Tumor volume measurements, weight measurements, and the injected Fig. 5. Tumor volume measurements, weight measurements, and the injected
doses for mouse 6 from group S1. The measurements are denoted with x-es, doses for mouse 1 from group C. The measurements are denoted with x-es,
which are linearly interpolated for visualization. which are linearly interpolated for visualization.

Fig. 4. Tumor volume measurements, weight measurements, and the injected Fig. 6. Tumor volume measurements, weight measurements, and the injected
doses for mouse 1 from group S2. The measurements are denoted with x-es, doses for the mouse 5 from group C. The measurements are denoted with x-es,
which are linearly interpolated for visualization. which are linearly interpolated for visualization.

2) If the tumor volume was under 200 mm3 , we switched selected from groups S1 and S2, and two from the con-
to minimal drug level therapy described in Section III-C. trol group C. Compared to the control group, the injections
For comparison, we calculated the survival time of each are more frequent in groups S1 and S2 and have smaller
specimen starting from the second injection since this is the doses.
time when the different treatment protocols started. The overall In these examples, the maximal tumor volumes were larger
survival is shown in Fig. 2 for the three groups. The sur- after relapse for the mice from groups S1 and S2, possibly due
vival curves were estimated with the nonparametric method of to the lower dosages of the drug. However, after the tumor
Kaplan and Meier [56]. The equality of survival curves among remission, the tumor volume was kept at a low value with
different groups was tested with the nonparametric log-rank small dosages, while in the case of the control group, there
test [57], which resulted in a p-value of 0.031. This analysis were no injections after remission; thus, the tumor grew back
showed that the strategies described in this article significantly after a short time. The results illustrate that the tumor can be
increased the overall survival of the mice. kept at a low size with more frequent but smaller doses.
We show measurements for four selected mice in Figs. 3–6. Based on these results, a suggestion for the application of
The figures show the approximated tumor volumes, mea- the results is as follows.
sured mouse weights, and injected doses. Two mice are 1) Collect measurements.
606 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 54, NO. 1, JANUARY 2024

2) Carry out parametric identification using the algorithm [3] E. Q. Wu, C.-T. Lin, L.-M. Zhu, Z. R. Tang, Y.-W. Jie, and G.-R. Zhou,
presented in Section II-E. “Fatigue detection of pilots’ brain through brains cognitive map and
multilayer latent incremental learning model,” IEEE Trans. Cybern.,
3) If the parameters of a specific patient are known, vol. 52, no. 11, pp. 12302–12314, Nov. 2022.
calculate the maximal effect therapy discussed in [4] G. Kong, D.-L. Xu, J.-B. Yang, T. Wang, and B. Jiang, “Evidential rea-
Section III-B. If toxicity is an issue, either decrease the soning rule-based decision support system for predicting ICU admission
and in-hospital death of trauma,” IEEE Trans. Syst., Man, Cybern., Syst.,
effect by decreasing κ, or use the minimal dose therapy vol. 51, no. 11, pp. 7131–7142, Nov. 2021.
discussed in Section III-C. [5] D. B. Larson, M. C. Chen, M. P. Lungren, S. S. Halabi, N. V. Stence, and
4) If the patient-specific parameters are not known, but C. P. Langlotz, “Performance of a deep-learning neural network model
in assessing skeletal maturity on pediatric hand radiographs,” Radiology,
we have a priori measurements from a population with vol. 287, no. 1, pp. 313–322, 2018.
similar characteristics, we can use the ranges of the [6] S. Wang et al., “An ensemble-based densely-connected deep learning
parameters to calculate the worst-case therapy discussed system for assessment of skeletal maturity,” IEEE Trans. Syst., Man,
Cybern., Syst., vol. 52, no. 1, pp. 426–437, Jan. 2022.
in Section III-D. [7] M. Gazda, M. Hireš, and P. Drotár, “Multiple-fine-tuned convolutional
neural networks for Parkinson’s disease diagnosis from offline handwrit-
ing,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 52, no. 1, pp. 78–89,
V. C ONCLUSION Jan. 2022.
[8] C.-M. Ionescu, “A computationally efficient hill curve adaptation strat-
Cyber-medical systems in therapy generation may revolu- egy during continuous monitoring of dose-effect relation in anaesthe-
tionize the treatment of numerous diseases, including cancer. sia,” Nonlinear Dyn., vol. 92, no. 3, pp. 843–852, 2018.
[9] P. H. Colmegna, R. S. Sánchez-Peña, R. Gondhalekar, E. Dassau, and
We proposed therapy generation algorithms based on a tumor F. J. Doyle, “Switched LPV glucose control in type 1 diabetes,” IEEE
model, supposing that the parameters of the model are known Trans. Biomed. Eng., vol. 63, no. 6, pp. 1192–1200, Jun. 2016.
from an identification procedure. [10] V. M. Pérez-García et al., “Computational design of improved standard-
ized chemotherapy protocols for grade II oligodendrogliomas,” PLoS
The algorithms are based on an optimization that calculates Comput. Biol., vol. 15, no. 7, 2019, Art. no. e1006778.
the minimal injection doses, ensuring the drug level is kept [11] T. Browder et al., “Antiangiogenic scheduling of chemotherapy improves
over a specified limit based on personalized pharmacokinetic efficacy against experimental drug-resistant cancer,” Cancer Res.,
vol. 60, pp. 1878–1886, Apr. 2000.
model parameters. We proposed strategies to determine this [12] L. Kovács et al., “Experimental closed-loop control of breast cancer in
limit. mice,” Complexity, vol. 2022, pp. 1–10, May 2022. [Online]. Available:
The maximal effect therapy is used to keep the drug level https://www.hindawi.com/journals/complexity/2017/5985031/
[13] I. Smalley et al., “Leveraging transcriptional dynamics to improve BRAF
over a constant multiple of the median effective dose param- inhibitor responses in melanoma,” eBioMedicine, vol. 48, pp. 178–190,
eter to ensure that the drug has a predefined efficiency all the Oct. 2019.
time. The advantage of this method is that it only requires [14] A. R. Akhmetzhanov, J. W. Kim, R. Sullivan, R. A. Beckman, P. Tamayo,
extra knowledge of the pharmacodynamics, and by choosing and C.-H. Yeang, “Modelling bistable tumour population dynamics
to design effective treatment strategies,” J. Theor. Biol., vol. 474,
a large multiplier, the robustness can be increased at the cost pp. 88–102, Aug. 2019.
of toxicity. [15] J. M. Greene, J. L. Gevertz, and E. D. Sontag, “Mathematical approach
We provide a formula for the minimal drug level that to differentiate spontaneous and induced evolution to drug resistance
during cancer treatment,” JCO Clin. Cancer Inform., vol. 3, pp. 1–20,
ensures that the tumor volume does not increase during the Apr. 2019.
therapy. These results were generalized for the cases when [16] D. A. Drexler, J. Sápi, and L. Kovács, “Optimal discrete time control
intrapatient and interpatient variability of the parameters are of antiangiogenic tumor therapy,” IFAC-PapersOnLine, vol. 50, no. 1,
pp. 13504–13509, 2017.
present, providing a robust impulsive therapy solution. [17] D. A. Drexler and L. Kovács, “Optimization of impulsive discrete-time
The results were tested in in vivo experiments using tumor chemotherapy,” in Proc. IEEE 1st Int. Conf. Societal Autom.,
29 mice. The therapy generated with the proposed methods 2019, pp. 1–7.
[18] F. Cacace, V. Cusimano, and P. Palumbo, “Optimal impulsive control
was compared to a generic protocol used in clinical practice. with application to antiangiogenic tumor therapy,” IEEE Trans. Control
The superiority of our results was proved with a log-rank test, Syst. Technol., vol. 28, no. 1, pp. 106–117, Jan. 2020.
which showed a significant increase in overall survival. [19] H. Zhu, X. Li, and S. Song, “Input-to-state stability of nonlinear
impulsive systems subjects to actuator saturation and external distur-
The combination of the strategies given in Section III was bance,” IEEE Trans. Cybern., vol. 53, no. 1, pp. 173–183, Jan. 2023.
carried out using heuristics in the experiments. This can [20] X. Li, H. Zhu, and S. Song, “Input-to-state stability of nonlinear systems
be developed further by using feedback of the tumor vol- using observer-based event-triggered impulsive control,” IEEE Trans.
Syst., Man, Cybern., Syst., vol. 51, no. 11, pp. 6892–6900, Nov. 2021.
ume, which can change the MIC after every sampling, or
[21] E. Ahmadi, J. Zarei, and R. Razavi-Far, “Robust l1 -controller design
one can track the parameter changes and use the strategy in for discrete-time positive T–S fuzzy systems using dual approach,” IEEE
Section III-C to personalize the therapy and track intrapatient Trans. Syst., Man, Cybern., Syst., vol. 52, no. 2, pp. 706–715, Feb. 2022.
variability. [22] S. Shi, Y. Chen, and X. Yao, “In vivo computing strategies for tumor
Sensitization and targeting,” IEEE Trans. Cybern., vol. 52, no. 6,
pp. 4970–4980, Jun. 2022.
[23] S. Shi, N. Sharifi, Y. Chen, and X. Yao, “Tension-relaxation in vivo
R EFERENCES computing principle for tumor sensitization and targeting,” IEEE Trans.
Cybern., vol. 52, no. 9, pp. 9145–9156, Sep. 2022.
[1] E. Q. Wu, P. Xiong, Z.-R. Tang, G.-J. Li, A. Song, and L.-M. Zhu, [24] S. Shi, Y. Chen, and X. Yao, “NGA-inspired nanorobots-assisted detec-
“Detecting dynamic behavior of brain fatigue through 3-D-CNN- tion of multifocal cancer,” IEEE Trans. Cybern., vol. 52, no. 5,
LSTM,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 52, no. 1, pp. 2787–2797, May 2022.
pp. 90–100, Jan. 2022. [25] H. Zhou and G. Alici, “A magnetically actuated novel robotic capsule
[2] W. Dang, Z. Gao, D. Lv, X. Sun, and C. Cheng, “Rhythm-dependent for site-specific drug delivery inside the gastrointestinal tract,” IEEE
multilayer brain network for the detection of driving fatigue,” IEEE J. Trans. Syst., Man, Cybern., Syst., vol. 52, no. 6, pp. 4010–4020,
Biomed. Health Inform., vol. 25, no. 3, pp. 693–700, Mar. 2021. Jun. 2022.
KOVÁCS et al.: POSITIVE IMPULSIVE CONTROL OF TUMOR THERAPY—A CYBER-MEDICAL APPROACH 607

[26] S.-M. Tse, Y. Liang, K.-S. Leung, K.-H. Lee, and T. S.-K. Mok, “A [50] M. Fidler et al., “Nonlinear mixed-effects model development and
memetic algorithm for multiple-drug cancer chemotherapy schedule simulation using nlmixr and related R open-source packages,” CPT
optimization,” IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 37, Pharmacomet. Syst. Pharmacol., vol. 8, no. 9, pp. 621–633, Sep. 2019.
no. 1, pp. 84–91, Feb. 2007. [51] G. Alefeld and G. Mayer, “Interval analysis: Theory and applications,” J.
[27] P. Hahnfeldt, D. Panigrahy, J. Folkman, and L. Hlatky, “Tumor develop- Comput. Appl. Math., vol. 121, no. 1, pp. 421–464, 2000.
ment under angiogenic signaling: A dynamical theory of Tumor growth, [52] W. Lodwick (Univ. Colorado Denver, Denver, CO, USA). Constrained
treatment response, and postvascular dormancy,” Cancer Res., vol. 59, Interval Arithmetic. (1999). [Online]. Available: http://www-math.
no. 19, pp. 4770–4775, 1999. ucdenver.edu/ccm/reports/index.shtml
[28] D. A. Drexler and L. Kovács, “Optimization of low dose metronomic [53] D. A. Drexler, I. Nagy, and V. Romanovski, “Bifurcations in a closed-
therapy based on pharmacological parameters,” IFAC-PapersOnLine, loop model of tumor growth control,” in Proc. 21th IEEE Int. Symp.
vol. 54, no. 15, pp. 221–226, 2021. Comput. Intell. Inform., 2021, pp. 329–334.
[29] H. Kusuoka et al., “Optimal control in compartmental systems and its [54] X. Liu et al., “Somatic loss of BRCA1 and P53 in mice induces mam-
application to drug administration,” Math. Biosci., vol. 53, nos. 1–2, mary tumors with features of human BRCA1-mutated basal-like breast
pp. 59–77, 1981. cancer,” Proc. Nat. Acad. Sci., vol. 104, no. 29, pp. 12111–12116, 2007.
[30] J. S. Lowengrub et al., “Nonlinear modelling of cancer: Bridging the [55] S. Rottenberg et al., “Selective induction of chemotherapy resistance
gap between cells and tumours,” Nonlinearity, vol. 23, no. 1, pp. R1–R9, of mammary tumors in a conditional mouse model for hereditary breast
2010. cancer,” Proc. Nat. Acad. Sci., vol. 104, no. 29, pp. 12117–12122, 2007.
[31] P. M. Altrock, L. L. Liu, and F. Michor, “The mathematics of can- [56] E. L. Kaplan and P. Meier, “Nonparametric estimation from incomplete
cer: Integrating quantitative models,” Nat. Rev. Cancer, vol. 15, no. 12, observations,” J. Amer. Stat. Assoc., vol. 53, no. 282, pp. 457–481, 1958.
pp. 730–745, 2015. [57] R. Peto and J. Peto, “Asymptotically efficient rank invariant test pro-
[32] A. M. Jarrett et al., “Mathematical models of tumor cell proliferation: A cedures,” J. Royal Stat. Soc. A, Gen., vol. 135, no. 2, pp. 185–207,
review of the literature,” Expert Rev. Anticancer Ther., vol. 18, no. 12, 1972.
pp. 1271–1286, 2018.
[33] J. Sápi, L. Kovács, D. A. Drexler, P. Kocsis, D. Gajári, and Z. Sápi,
“Tumor volume estimation and quasi-continuous administration for most
Levente Kovács (Senior Member, IEEE) received
effective bevacizumab therapy,” PLoS One, vol. 10, no. 11, pp. 1–20,
the M.Sc. degree in electrical engineering from the
2015.
“Politehnica” University of Timis, oara, Timis, oara,
[34] D. A. Drexler, J. Sápi, and L. Kovács, “Modeling of tumor growth
Romania, in 2000, and the M.Sc. degree in biomed-
incorporating the effects of necrosis and the effect of beva-
ical engineering and the Ph.D. degree in electri-
cizumab,” Complexity, vol. 2017, Nov. 2017, Art. no. 5985031.
cal engineering from the Budapest University of
[Online]. Available: https://www.hindawi.com/journals/complexity/
Technology and Economics, Budapest, Hungary, in
2022/9348166/
2011 and 2008, respectively.
[35] D. A. Drexler, T. Ferenci, A. Füredi, G. Szakács, and L. Kovács, He is the Rector/President of Óbuda University,
“Experimental data-driven tumor modeling for chemotherapy,” IFAC- Budapest. He is a Full Professor with the John von
PapersOnLine, vol. 53, no. 2, pp. 16245–16250, 2020. Neumann Faculty of Informatics, Óbuda University,
[36] D. A. Drexler, T. Ferenci, A. Lovrics, and L. Kovács, “Tumor dynam- where he defended his habilitation with merit in 2013. He founded the
ics modeling based on formal reaction kinetics,” Acta Polytechnica Physiological Controls Research Center, Óbuda University in 2013, where
Hungarica, vol. 16, no. 10, pp. 31–44, 2019. he is the Deputy Head. He has published more than 500 articles in interna-
[37] A. Füredi et al., “Pegylated liposomal formulation of doxorubicin over- tional journals and refereed international conference papers, accumulating an
comes drug resistance in a genetically engineered mouse model of breast impact factor of over 120 and H-index of 30. His fields of interest are modern
cancer,” J. Controlled Release, vol. 261, pp. 287–296, 2017. control theory and physiological controls.
[38] J. G. Pierce and A. Schumitzky, “Optimal impulsive control of compart- Prof. Kovács was the President of IEEE Hungary Section from 2017 to 2020
ment models, I: Qualitative aspects,” J. Optim. Theory Appl., vol. 18, and has been the President since 2022. In 2016, he founded the Cybernetics
no. 4, pp. 537–554, Apr. 1976. Technical Committee of the IEEE SMC Society, the Cyber-Medical Technical
[39] M. Jacobs, “Optimisation of antimicrobial therapy using pharmacoki- Committee, promoting the theory and practice of personalized healthcare.
netic and pharmacodynamic parameters,” Clin. Microbiol. Infect., vol. 7,
no. 11, pp. 589–596, 2001.
[40] B. Meibohm and H. Derendorf, “Basic concepts of pharmacoki-
netic/pharmacodynamic (PK/PD) modelling,” Int. J. Clin. Pharmacol. Tamás Ferenci (Member, IEEE) received the
Therapeut., vol. 35, no. 10, pp. 401–413, 1997. Ph.D. degree in biostatistics from Óbuda University,
[41] A. I. Vol’Pert, “Differential equations on graphs (in Russian),” Math. Budapest, Hungary, in 2013.
USSR-Sbornik, vol. 17, no. 4, p. 571, 1972. He is with the Physiological Controls Research
[42] P. Érdi and J. Tóth, Mathematical Models of Chemical Reactions. Theory Center, Óbuda University. He is also with the
and Applications of Deterministic and Stochastic Models. Princeton, NJ, Department of Statistics, Corvinus University of
USA: Princeton Univ. Press, 1989. Budapest, Budapest, and the Department of Public
[43] J. Tóth, A. L. Nagy, and D. Papp, Reaction Kinetics: Exercises, Health, Semmelweis University, Budapest. He has
Programs and Theorems. New York, NY, USA: Springer, 2018. published more than 90 journal papers. His current
[44] D. A. Drexler, E. Virágh, and J. Tóth, “Controllability and reachabil- research interests include clinical biostatistics, epi-
ity of reactions with temperature and inflow control,” Fuel, vol. 211, demiology, and disease modeling.
pp. 906–911, Jan. 2018. Dr. Ferenci is an Editor of the European Journal of Epidemiology.
[45] J. C. Pinheiro and D. M. Bates, Mixed-Effects Models in S and S-PLUS.
New York, NY, USA: Springer, 2000.
[46] A. P. Dempster, N. M. Laird, and D. B. Rubin, “Maximum likelihood Balázs Gombos received the B.S. degree in biol-
from incomplete data via the EM algorithm,” J. Royal Stat. Soc. B, ogy from the University of Debrecen Clinical Center,
Methodol., vol. 39, no. 1, pp. 1–22, 1977. Medical Microbiology, University of Debrecen,
[47] B. Delyon, M. Lavielle, and E. Moulines, “Convergence of a stochastic Debrecen, Hungary, in 2017, and the M.S. degree
approximation version of the EM algorithm,” Ann. Stat., vol. 27, no. 1, in molecular biology from the Faculty of Medicine,
pp. 94–128, 1999. Department of Biochemistry and Molecular Biology,
[48] S.-M. Chow, Z. Lu, A. Sherwood, and H. Zhu, “Fitting nonlinear ordi- University of Debrecen in 2020. He is currently
nary differential equation models with random effects and unknown pursuing the Ph.D. degree in biology with the
initial conditions using the stochastic approximation expectation– Drug Resistance Research Group, Research Centre
maximization (SAEM) algorithm,” Psychometrika, vol. 81, no. 1, for Natural Sciences, Budapest, Hungary, and with
pp. 102–134, 2016. the Molecular Medicine Ph.D. School, Semmelweis
[49] (R Foundation for Statistical Computing, Vienna, Austria). R: A University, Budapest, Hungary.
Language and Environment for Statistical Computing. (2021). [Online]. His current research interests include personalized chemotherapy and
Available: https://www.R-project.org/ algorithm-assisted therapy design.
608 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 54, NO. 1, JANUARY 2024

András Füredi received the M.Sc. degree in bio- Gergely Szakács received the medical degree
physics from the University of Szeged, Szeged, and Ph.D. degree in pathobiochemistry from
Hungary, in 2010, and the Ph.D. degree in molecular Semmelweis University, Budapest, Hungary, in 1996
medicine from Semmelweis University, Budapest, and 2001, respectively.
Hungary, in 2018. During his postdoc with National Cancer Institute
He is a Marie Curie Fellow with the Microsystems (NIH), Bethesda, MD, USA, he discovered com-
Laboratory, HUN-REN Centre for Energy Research, pounds targeting multidrug resistant (MDR) cancer,
Budapest, while also a Project Leader with the Drug receiving two Federal Technology Transfer Awards.
Resistance Research Group, HUN-REN Research In 2005, he returned to Hungary to establish his lab
Centre for Natural Sciences, and the Cybermedical with the support of the EMBO Young Investigator
Competence Center. He is also a Scientific Advisor Program, a Marie Curie Fellowship, and grants from
for several Hungarian and international research and development companies. the Leukemia & Lymphoma Society and NIH. In 2012, he obtained an ERC
He is an expert in modeling human cancer in mice and developing novel drugs Starting grant to support his research on innovative approaches to target MDR
and treatment strategies to combat malignancies. cancer. He has published several high-profile original papers and reviews
on the “Achilles’ heel” of multidrug resistant cancer. His research revolves
around membrane transporters and drug resistant cancer. Based on his find-
ings demonstrating the paradoxical sensitivity of cancer cells overexpressing
the MDR exporter P-gp, he has aimed to exploit the collateral sensitivity of
MDR cells.
Dr. Szakács received the Momentum Award established by the Hungarian
Imre Rudas (Life Fellow, IEEE) received the bach- Academy of Sciences with the aim to attract outstanding young scientists to
elor’s degree in mechanical engineering from Don’t Hungary in 2010. He is the Chair of the Excellence-Based Graduate Training
Bànki Technical College, Budapest, Hungary, in Program (International Ph.D. Program for Translational Oncology) at the
1971. He received the master’s degree in mathe- Medical University of Vienna.
matics from Eötvös Loránd University, Budapest,
Hungary, in 1977, the Ph.D. degree in robotics and
the Doctor of Science degree in computer science
from the Hungarian Academy of Sciences, Budapest,
in 1987 and 2004, respectively, and the Doctor Dániel András Drexler (Member, IEEE) received
Honoris Causa degrees in computer science from the the Ph.D. degree in electrical engineering from the
Technical University of Košice, Košice, Slovakia, in Budapest University of Technology and Economics,
2001; the “Politehnica” University of Timis, oara, Timis, oara, Romania, in 2005; Budapest, Hungary, in 2015.
Óbuda University, Budapest, Hungary, in 2014; and the Slovak University of He is the Vice Deputy Head of the Physiological
Technology in Bratislava, Bratislava, Slovakia, in 2016. Controls Research Center, Óbuda University,
He is a Rudolf Kalman Distinguished Professor, a Rector Emeritus, and a Budapest, where he is also an Associate Professor
Professor Emeritus with Óbuda University, Budapest. He has edited and/or with the John von Neumann Faculty of Informatics.
published 22 books, published more than 890 papers in international scien- He has published 119 papers in international
tific journals, conference proceedings, and book chapters, and received more scientific journals, conference proceedings, and
than 8000 citations. His present areas of research activities are Computational book chapters and received over 1000 citations. His
Cybernetics, Cyber Physical Control, Robotics, and Systems of Systems. research interests include modeling and control of physiological, positive,
Dr. Rudas was awarded by the Honorary Professor Title in 2013 and and impulsive systems.
Ambassador Title by the Wrocław University of Science and Technology. Dr. Drexler has been the Membership Development Officer of the IEEE
He is the Junior Past President of the IEEE Systems, Man, and Cybernetics Hungary Section since 2017 and the Vice Chair of the IEEE Hungary Section
Society. He is a Fellow of the International Fuzzy Systems Association. since 2021.