Computer Methods and Programs in Biomedicine 145 (2017) 35–43

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine

journal homepage: www.elsevier.com/locate/cmpb

Application of genetic algorithm for hemodialysis schedule

optimization
Jin Woo Choi a, Hajeong Lee b,c,d, Jung Chan Lee e,f,g,∗, Saram Lee h, Yon Su Kim b,c,d,
Hyung-Jin Yoon e,f,g, Hee Chan Kim e,f,g
a
Interdisciplinary Program in Bioengineering, Seoul National University Graduate School, Seoul, Korea
b
Department of Internal Medicine, Seoul National University Hospital, Seoul, Korea
c
Department of Internal Medicine, Seoul National University College of Medicine, Seoul, Korea
d
Kidney Research Institute, Medical Research Center, Seoul National University, Seoul, Korea
e
Department of Biomedical Engineering, Seoul National University Hospital, Seoul, Korea
f
Department of Biomedical Engineering, Seoul National University College of Medicine, Seoul, Korea
g
Institute of Medical and Biological Engineering, Medical Research Center, Seoul National University, Seoul, Korea
h
Biomedical Research Institute, Seoul National University Hospital, Seoul, Korea

a r t i c l e i n f o a b s t r a c t

Article history: Background: The conventional hemodialysis (HD) schedule has been used for decades, even though new
Received 22 September 2016 modalities have been introduced. Many reasons limit practices of frequent dialysis, such as patients’ en-
Revised 3 April 2017
vironments and unknown optimal schedules for each patient. This research provides a theoretical recom-
Accepted 7 April 2017
mendation of HD schedule through genetic algorithm (GA).
Methods: An end-stage renal disease (ESRD) with various dialysis conditions was modeled through a
Keywords: classic variable-volume two-compartment kinetic model to simulate an anuric patient, and GA was im-
Hemodialysis plemented to search for an optimal HD schedule for each individual considering and ignoring burden
Hemodialysis schedule consumption of each dialysis session. The adequacy of the optimized HD schedules through GA was as-
End-stage renal disease
sessed with time average concentration (TAC) and time average deviation (TAD).
Genetic algorithm
Results: While ignoring the burden of dialysis sessions, GA returned schedules with slightly improved val-
ues of adequacy criteria (EKRc and std Kt/V), compared to the conventional regular uniform HD schedules.
The optimized HD schedules also showed decreased TAC and TAD values compared to the conventional
regular uniform HD schedules. It showed that frequent dialysis resulted in more effective treatment and
higher fitness values. However, when burden was considered, less frequent dialysis schedules showed
better fitness value.
Conclusions: Through this research, GA confirmed that at least 12 h of dialysis should be conducted for a
week. The optimized schedules from GA indicated that evenly distributing the intervals amongst sessions
is efficient, and that scheduling a session at the start and end of a week is optimal to overcome a long
weekend interval. The theoretical optimal schedule of HD may help distribution of frequent dialysis and
provide more schedule options to patients.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction
For over half a century, the conventional hemodialysis (HD)
schedule of three times a week has been used as the stan-
dard modality to treat end-stage renal disease (ESRD) despite the
high mortality rate and the poor quality of life. To overcome
these drawbacks, an innovative technological changeover that re-
∗
Corresponding author at: Department of Biomedical Engineering, Seoul National
University College of Medicine 101 Daehakro, Jongro-gu, Seoul 03080, Republic of
Korea.
E-mail address: ljch@snu.ac.kr (J.C. Lee).
duces the time and resource consumption, and that provides dial-
ysis treatment at the optimal dose for each individual patient is
strongly desired. Therefore, various modalities, such as daily HD,
nocturnal HD, and continuous HD treatment, have been developed
for ESRD patients and new technologies have been applied in HD.
More frequent HD was assumed to be more physiological and to
better mimic the biological function of the kidneys, yet only re-
cently it was confirmed that more frequent HD improves dialy-
sis efficacy compared with conventional HD [1]. Reduced dialysis-
induced myocardial stunning with frequent HD compared to con-
ventional HD was reported [2] and improvements in hyperphos-
phatemia and hypertension control were also reported with fre-

http://dx.doi.org/10.1016/j.cmpb.2017.04.003
0169-2607/© 2017 Elsevier B.V. All rights reserved.
36 J.W. Choi et al. / Computer Methods and Programs in Biomedicine 145 (2017) 35–43
quent HD [1,3]. Several computer simulation studies predicted that
increased frequency or increased dialysis duration could increase
molecular clearance [4–7]. Moreover, previous works have shown
that frequent or longer dialysis not only decreases time average
concentration (TAC) in body, but also decreases time average devi-
ation (TAD), which denotes that those modalities are more physio-
logical than the conventional HD schedule [8,9]. Therefore, a short
daily HD seemed to be a good clinical alternative to conventional
HD [10]. However, there was some skepticism regarding the ef-
fectiveness of frequent HD in every case [11]. Therefore, the opti-
mal hemodialysis schedule that suits each patient must be identi-
fied and verified. However, to be treated with a more frequent HD
schedule, patients must be assigned to more days of clinical visits
and more medical resources must be utilized, which increases both
the burden of time and the cost for patients [12–14]. Also, each pa-
tient has their own lifestyle and circumstances related to dialysis
treatment, and the optimal treatment schedule may vary among
patients. There are innumerable possible combinations, since vari-
ous parameters such as duration, interval, and dosage must be con-
sidered for frequent HD scheduling. However, it is impossible to
perform a clinical trial to identify which frequent HD schedule is
suitable for a specific individual. Therefore, new HD modalities are
not widely accepted in clinical practices, even though theoretical
and clinical results have shown that these modalities have definite
advantages. Recently, the computational modeling has been in the
spotlight for the proof of theories. The Cobelli research group de-
veloped a simulator of type 1 diabetes mellitus (T1DM), which was
accepted as a substitute for preclinical trials by the Food and Drug
Administration (FDA) [15,16].
In this study, we propose a method to search for optimal in-
termittent and frequent HD schedules for individuals using mathe-
matical modeling. For this purpose, we applied genetic algorithm
(GA) to extract optimal scheduling solution from innumerable
combinations. GA is a global optimization algorithm that employs
the evolutionary processes found in nature, such as inheritance,
selection, crossover, and mutation [17,18]. Compared to other op-
timization algorithms, GA shows strength in large multi-parameter
optimization problems with non-linear, objective stochastic func-
tions. Moreover, GA is operated with a set of parameters in a par-
allel environment, so it is suitable for problems with multiple lo-
cal optima. Therefore, GA is utilized in various applications such as
flow shop scheduling, image optimization, robotics, optimal power
flow searching, and bankruptcy prediction modeling [18–22].
We implemented GA for optimal HD scheduling by consider-
ing the genes as the dialysis sessions and the chromosomes as the
weekly dialysis schedule. As a result, a search space of around 10
million combinations was created, so GA was an adequate candi-
date for solving this problem. The corrected equivalent renal clear-
ance (EKRc) and standard Kt/V (std Kt/V), which are dose mea-
suring parameters widely used to compare the efficacy of dialysis
modalities, were used to calculate a fitness function for GA. The
searched optimal schedule was evaluated using TAC and TAD to
assess its adequacy [8,9]. Lastly, we introduced the idea of burden,
and examined its effect on searching for the optimal HD schedule.
2. Methods
2.1. Mathematical model of anuric patient
Anuric patients were modeled using a classic variable-volume
two-compartment kinetic model [4], as described in our previous
work [6,23]. Several assumptions were made for the model. The
dry weight and total body water volume of the patients were set
to 60 kg, 70 kg, and 80 kg, and 30 l, 35 l, and 40 l, respectively. The
water volume distribution ratio of urea in the intracellular fluid, as
the non-perfused compartment, and the extracellular fluid, as the
perfused compartment, was assumed to be 2:1. The ultrafiltration
rate (Quf ) was set in order to maintain dry body weight after each
dialysis session. The removal (Jv ) and intake (Iw ) of water, and the
generation (G) and removal (Js ) of urea only occurred in the per-
fused compartment. The daily water intake was set to 1.5 l, and the
rate of urea nitrogen generation was assumed to be 6.25 mg/min
[4,24].
The dialysis clearance (Kd ) was adjusted to 178 ml/min, barely
enough to achieve the HEMO standard dose equivalent of
EKRc ≥ 13.8 ml/min and std Kt/V ≥ 2.29 with a conventional HD
schedule consisting of 4-h, three times a week [25]. The solute
transfer rates for urea between the compartments were expressed
in the form of inter-compartment clearance (Kic ) and the values
were assumed to be 600 ml/min [4,24,26]. The initial concentra-
tions of blood urea nitrogen in the compartments were assumed
to be 100 mg/dl.
2.2. Dialysis schedule variables
Several options were considered in scheduling HD. Three avail-
able sessions were assigned for each day, separated into a morn-
ing session (start at 9 a.m.), an early afternoon session (start at
1 p.m.), and a late afternoon session (start at 5 p.m.). As a result,
21 sessions were available over the course of a week. The op-
tions for dialysis duration of each session were 4-h, 3-h, and 2-
h. Also, two scheduling preferences were considered: the number
of dialysis sessions performed per week (Nd/wk ) with possible val-
ues between three and seven, and the number of sequential days
not assigned to dialysis (Nd/ns ) which represented whether dial-
ysis was performed on any day of the week (Nd/ns = 0), only on
Monday through Saturday (Nd/ns = 1), or only on Monday through
Friday (Nd/ns = 2). Also, a cost factor, α, was assigned as either 0
or 1 for not considering or considering the burden of hemodialy-
sis, respectively. It is an arbitrary number that represents the bur-
den of time, money, quality of life, and other difficulties that pa-
tients may experience through hemodialysis treatment. Over 200
million schedule combinations were created with these variables.
The search space of GA was reduced to around 10 million combina-
tions, which is still a large search space, by restricting the schedule
to only assign one HD sessions per day.
2.3. Genetic algorithm
Genetic algorithm is an adaptive metaheuristic search algorithm
that mimics Charles Darwin’s “survival of the fittest” concept in
nature. GA is described in the two following sections: GA popula-
tion and GA operation.
2.3.1. GA population
The GA searches through population that evolve over genera-
tions for an optimal solution. The population of GA contains a set
of chromosomes which consist of genes. For the HD schedule op-
timization problem, the genes represented HD sessions while the
chromosomes represented weekly dialysis schedules. Since 21 ses-
sions were available in a week, a chromosome was composed of
21-digit genotype from Sunday to Saturday, i.e. xxx-xxx-xxx-xxx-
xxx-xxx-xxx. Every gene in a 21-digit chromosome had four op-
tions of dialysis duration in each session: 4-h, 3-h, 2-h, and 0-
h. A short 1-h dialysis duration was not considered, because it
is not physiologically and clinically significant. Each dialysis dura-
tion was coded as 4, 3, 2, and 0, respectively, and 0 indicated that
a dialysis was not assigned for a given session. For example, the
conventional HD chromosome structure, consisting of 4-h HD on
Monday, Wednesday, and Friday, with only morning sessions, was
represented as 0 0 0-40 0-0 0 0-40 0-0 0 0-40 0-0 0 0, as shown in Fig. 1.
 J.W. Choi et al. / Computer Methods and Programs in Biomedicine 145 (2017) 35–43 37

Fig. 1. A sample of genetic algorithm chromosome structure (Conventional hemodialysis schedule).

Fig. 2. Genetic algorithm (a) Flow chart of genetic algorithm. (b) Crossover operator example. (c) Mutation operator example.

The first generation of a population with 20 chromosomes was ini- fused compartments with respect to time [7].
tialized by assigning random weekly dialysis schedules. With the  t
1
given variables Nd/wk , Nd/ns and α , Nd/wk sessions were randomly T AC = C p (t )dt (1)
t 0
selected among 21 genes, and dialysis durations were randomly se-
lected among 4-h, 3-h, and 2-h. where Cp (t) is the urea concentration profiles in the perfused com-
partment.
2.3.2. GA operation EKR was calculated as follows,
The GA searched for the fittest schedule through the follow- EKR = G/T AC (2)
ing operators: fitness evaluation and selection, crossover, and mu-
tation. These operators were repeated until the target generation EKRc, corrected value of EKR for the normalized water volume,
(Gen = 10 0 0) was reached, as shown in the Fig. 2(a)., and details of 40 l, was calculated as follows,
the operators are described. E K Rc = E KR/V × 40 (3)
Fitness evaluation and Selection: Each chromosome was evalu-
where V is the total volume of the compartments.
ated using an objective function to calculate a fitness value. This
Std Kt/V is defined as the inverse of the concentration main-
function was used to evaluate each chromosome’s fitness within
tained by the waste generation per unit volume of body water and
the population. First, the anuric patient model was treated with 4
was also used to assess the control of the peak solute concentra-
weeks of each chromosome, which represents a dialysis schedule.
tions [27,28]. Generally, std Kt/V can be obtained as follows,
After the 4 weeks of HD, the fitness value was calculated using
1−e−eqKt/V
EKRc, which assesses the control of the averaged solute concentra- 10080 × t
tion, and std Kt/V [19]. These parameters are widely used to evalu- st d Kt /V = 1−e−eqKt/V 10080
(4)
spKt/V
+ Nd/wk ×ts
− 1
ate the adequacy of different renal treatment modalities. The time
averaged concentration (TAC) of the last week was calculated by where eqKt/V and spKt/V are respectively the equilibrated Kt/V and
integrating the concentration profile of the molecule in the per- the single-pool Kt/V, and ts (min) is the time length of the dialysis
38 J.W. Choi et al. / Computer Methods and Programs in Biomedicine 145 (2017) 35–43
session. However, this formula is applicable only to the conven-
tional regular uniform dialysis schedule.
Therefore, the std Kt/V for an irregular non-uniform combined
dialysis schedule was calculated as follows,
st d Kt /V = G · t/MPC/V
where t denotes the total weekly time, 10,080 min, and the mean
pre-treatment concentration (MPC) is acquired from averaging the
pre-treatment concentrations in the perfused compartment during
the week.
Casino and Lopez et al. suggested a requirement of
EKRc ≥ 11.0 ml/min for renal treatments with an anuric pa-
tient, and EKRc ≥ 9.0 for patients with a residual renal function
[29]. In this study, the adequacy is defined as EKRc ≥ 13.8 and std
Kt/V ≥ 2.29, which is equivalent to the HEMO Study standard dose
[25]. The fitness function was defined as the ratio of adequacy
output (relative achievement of both in the EKRc and std Kt/V)
and cost input (burden consumption) as follows,
EKRc
+ st 2dKt
F itness = δ (EKRc−13.8 ) × δ (st dKt /V −2.29 ) × 13.8
Td/wk + α × Nd/wk
 culate EKRc, std Kt/V, fitness, TAC, and TAD for each conventional
0, n < 0
δ (n ) =
1, n ≥ 0
where Td/wk is the total dialysis time per week and α is a cost fac-
tor that takes the additional time consumed by commuting, wait-
ing and preparation for each dialysis session into consideration as
mentioned in Section 2.2. When the burden of hemodialysis was
considered, α = 1, the number of dialysis sessions per week was
reflected into the fitness function, which penalized frequent dial-
ysis schedules. A fitness of 0 indicated inadequacy either in EKRc
or stdKt/V parameters. When two different schedules achieved the
same adequacy output, the schedule that consumed less time was
considered to be the better solution.
After the fitness value was calculated for the population, the
selection was operated by selecting the top 70% of the fittest chro-
mosomes from the population. These chromosomes were passed
down to the next generation. The remaining 30% of the population
in the next generation was generated through crossover and muta-
tion operators.
Crossover: This process mimics the reproduction as seen in the
nature. Given two parent chromosomes, selected from the “fittest”
group, two child chromosomes were generated. The first child con-
tained the first 9-digits chromosomes of parent 1 and the last 12-
digits chromosomes of parent 2. The second child contained the
first 12-digit chromosomes of parent 2 and the last 9-digits chro-
mosomes of parent 1, as shown in Fig. 2(b). The algorithm ensured
that the correct number of Nd/wk was assigned to the week. This
operation was conducted three times for each generation to gen-
erate 6 chromosomes, which is 30% of the population for the next
generation. The crossover probability was set to 50%.
Mutation: In this operator, the children chromosomes from the
crossover were mutated with a probability of 10%. The dialysis du-
ration of each child chromosome was randomly reassigned among
4, 3, and 2 during the mutation operation, as shown in Fig. 2(c).
2.4. GA assessment
After 10 0 0 generations, the adequacy of the optimized dialysis
schedules were assessed by analyzing the time average concentra-
tion (TAC) and time average deviation (TAD). The TAC value was
calculated as previously described in Eq. (1) for the last week. The
TAD was calculated with following equation for the last week [9],
 t
1
T AD = |Cp (t ) − T AC|dt
t 0
where Cp is the concentration profiles in the perfused compart-
ment.
2.5. Numerical analysis
(5)
The fourth-fifth order Runge–Kutta method with variable time-
steps was used to solve the governing mass balance equations,
and was implemented using the ODE45 function, which uses the
Runge–Kutta method to solve the differential equation. The cal-
culation result provided an optimal scheduling solution with the
corresponding urea concentrations and water volume profiles in
the non-perfused and perfused compartments over a 4-week pe-
riod. The water volume profiles were used to verify the achieve-
ment of a suitable exchange volume at each treatment session. The
proposed work was implemented with MATLAB (The MathWorks,
Natick, MA, USA) on a Windows 7 OS. All computations were per-
formed on a computer with CPU i5-3570 (3.40 GHz) and RAM 8GB.
3. Results
/V
.29
,
A mathematical model of an anuric patient was used to cal-
(6) regular uniform dialysis schedule, and the results are presented in
a monotonous timetable (Table 1). Different combinations of the
number of sessions assigned to a week (Nd/wk ) and the duration
of each session were examined. All schedules, except for Nd/wk = 6
and Nd/wk = 7, did not contain weekend sessions and all the ses-
sions were arbitrarily assigned in the morning. As expected, the
conventional HD schedule, 3 × 4-h HD, achieved a standard dose
equivalent to EKRc ≥ 13.8 and std Kt/V ≥ 2.29. Other schedules also
satisfied the standard EKRc and std Kt/V, except for the schedules
3 × 3-h, 4 × 2-h, and 5 × 2-h.
To confirm that GA successfully searched towards the optimal
solution, the fitness values of each generation were analyzed. Since
GA always searches for higher fitness value, the fitness cannot de-
crease. As shown in Fig. 3, GA successfully searched towards higher
fitness and arrived at the optimal solution. For the most of cases,
GA converged after around 200 generations, but GA converged af-
ter 500 generations for the Nd/wk = 5 schedules. The computation
time ranged from 9 to 45 min to optimize each schedule condition.
GA searched for the optimal HD schedule using a mathematical
model of an anuric patient, without considering additional burden
consumption (α = 0). Different combinations of the number of ses-
sions assigned to a week (Nd/wk ), the number of sequential rest
days not assigned a dialysis session (Nd/ns ), and the duration of
each session were examined. The EKRc, std Kt/V, and fitness were
then calculated (Table 2). All optimal schedules satisfied the ade-
quacy criteria of EKRc ≥ 13.8, std Kt/V ≥ 2.29 and Td/wk of 12 h ex-
cept for the case of Nd/wk = 7. Also, to confirm that GA searched the
optimal HD schedule, neighboring schedules were compared with
those of the optimized HD schedule as shown in Table 3.
These optimized HD schedules were assessed with TAC and TAD
to evaluate the adequacy of the treatment. For each Nd/wk , TAC
and TAD are reported in Table 2. Additionally, the TAC and TAD
along with the fitness of the optimized HD schedules were plotted
with the conventional regular uniform HD schedules for compar-
ison in Fig. 4. In Fig. 4(b), TAC value of the conventional sched-
ule for Nd/wk = 5 showed significantly lower TAC, because the total
dialysis duration was longer compare to other schedules. The per-
formance of GA was also evaluated among patients with different
total body water: 30 l, 35 l, and 40 l, as shown in Fig. 5.
GA also searched for the optimal HD schedule while consider-
ing additional burden consumption (α = 1). Similar to the previous
results, EKRc, std Kt/V, fitness, TAC, and TAD are summarized in
Table 4. For Nd/wk = 5, 6, and 7, Td/wk was increased to 15 h, 19 h,
(7) and 24 h, respectively.
 J.W. Choi et al. / Computer Methods and Programs in Biomedicine 145 (2017) 35–43 39

Table 1
Adequacy of the conventional regular uniform dialysis schedule.

Nd/wk Weekly schedule EKRc (ml/min) Std Kt/V Td/wk (h) Fitness (α = 0) Fitness (α = 1) TAC (mg/dl) TAD (mg/dl)

Sun Mon Tue Wed Thu Fri Sat

3 – 400 – 400 – 400 – 13.81 2.36 12 0.1691 0.1354 51.68 13.68

– 300 – 300 – 300 – 10.84∗ 1.99∗ 9 0 0 65.84 13.62
4 – 300 300 – 300 300 – 13.99 2.64 12 0.1805 0.1354 51.04 12.87
– 200 200 – 200 200 – 9.84∗ 1.99∗ 8 0 0 72.54 12.50
5 – 300 300 300 300 300 – 16.66 3.30 15 0.1766 0.1324 42.86 13.60
– 200 200 200 200 200 – 11.95∗ 2.48 10 0 0 59.70 13.50
6 – 200 200 200 200 200 200 14.74 2.97 12 0.1970 0.1314 48.46 8.20
7 200 200 200 200 200 200 200 17.49 3.46 14 0.1983 0.1323 40.84 5.59

Nd/wk : number of dialysis sessions assigned to a week; the schedule code indicates the dialysis duration for each session, 0: no dialysis, 4: 4-h hemodialysis, 3: 3-h hemodial-
ysis, 2: 2-h hemodialysis; the digits in the schedule code indicate the dialysis session start time, the first digit: morning (start at 9 p.m.), the second digit: early afternoon
(start at 1 p.m.), the third digit: late afternoon (start at 5 p.m.). Asterisk indicates inadequacy with the criteria of EKRc ≥ 13.8 and std Kt/V ≥ 2.29, equivalent to the HEMO
Study standard dose.

Fig. 3. Progress of genetic algorithm over generations. Nd/wk : number of dialysis sessions assigned to a week. Nd/ns : number of sequential rest days not assigned to a dialysis
session.

Table 2.
Adequacy of the optimized schedules resulting from genetic algorithm when time factor α is 0, which means that only the dialysis time consumption with no
additional time consumption is considered to calculate the fitness function.

Nd/wk Nd/ns Weekly schedule EKRc (ml/min) Std Kt/V Td/wk (h) Fitness (α = 0) TAC (mg/dl) TAD (mg/dl)

Sun Mon Tue Wed Thu Fri Sat

3 0 040 000 004 000 000 400 000 13.95 2.35 12 0.1699 51.26 13.14
1 – 400 000 040 000 004 000 13.95 2.35 12 0.1699 51.32 13.16
2 – 400 000 040 000 004 – 13.95 2.35 12 0.1699 51.32 13.16
4 0 000 003 000 300 000 300 003 14.49 2.63 12 0.1833 49.36 10.16
1 – 003 000 300 000 300 003 14.49 2.63 12 0.1833 49.36 10.16
2 – 400 200 000 300 003 – 14.15 2.68 12 0.1829 50.62 12.57
5 0 002 000 030 000 300 020 002 14.69 2.85 12 0.1924 48.74 9.06
1 – 300 020 002 002 000 030 14.69 2.85 12 0.1924 48.71 9.06
2 – 300 200 200 200 003 – 14.42 2.86 12 0.1910 49.69 10.61
6 0 200 200 200 020 002 002 000 14.92 2.97 12 0.1980 47.98 7.26
1 – 200 200 200 020 002 002 14.91 2.97 12 0.1980 48.01 7.26
7 0 020 020 020 020 020 020 020 17.49 3.46 14 0.1983 40.84 5.59

Nd/wk : number of dialysis sessions assigned to a week; Nd/ns : number of sequential rest days not assigned to a dialysis session; the schedule code indicates the dialysis
duration for each session, 0: no dialysis, 4: 4-h hemodialysis, 3: 3-h hemodialysis, 2: 2-h hemodialysis; the digits in the schedule code indicate the dialysis session
start time, the first digit: morning (start at 9 p.m.), the second digit: early afternoon (start at 1 p.m.), the third digit: late afternoon (start at 5 p.m.). Adequacy criteria
are EKRc ≥ 13.8 and std Kt/V ≥ 2.29, equivalent to the HEMO Study standard dose.
40 J.W. Choi et al. / Computer Methods and Programs in Biomedicine 145 (2017) 35–43

Table 3.
Fitness comparison between the optimized schedule resulting from genetic algorithm and
its neighboring schedules.

Nd/wk Weekly schedule Fitness (α = 0)

Sun Mon Tue Wed Thu Fri Sat

3 GA – 400 000 040 000 004 – 0.1699

Nb1 – 400 000 400 000 004 – 0.1697
Nb2 – 400 000 004 000 004 – 0.1696

Nd/wk : number of dialysis sessions assigned to a week; GA: genetic algorithm; Nb1 and
Nb2: neighboring schedule 1 and 2 respectively.

Table 4.
Adequacy of the optimized schedules resulting from genetic algorithm when time factor α is 1, which means that dialysis time consumption and 1-h additional time
consumption per dialysis session are considered to calculate the fitness function.

Nd/wk Nd/ns Weekly schedule EKRc (ml/min) Std Kt/V Td/wk (h) Fitness (α = 1) TAC (mg/dl) TAD (mg/dl)

Sun Mon Tue Wed Thu Fri Sat

3 0 400 000 040 000 004 000 000 13.95 2.35 12 0.1359 51.29 13.14
1 – 040 000 004 000 000 400 13.95 2.35 12 0.1359 51.32 13.15
2 – 400 000 040 000 004 – 13.95 2.35 12 0.1359 51.32 13.16
4 0 000 300 003 000 003 000 300 14.49 2.63 12 0.1375 49.36 10.16
1 – 300 003 000 003 000 300 14.49 2.63 12 0.1375 49.36 10.16
2 – 400 000 300 200 003 – 14.21 2.67 12 0.1371 50.45 12.14
5 0 003 000 300 003 000 300 030 18.08 3.29 15 0.1373 35.60 8.35
1 – 300 030 003 000 300 003 18.08 3.29 15 0.1373 37.82 8.34
2 – 400 200 300 020 003 – 16.41 3.19 14 0.1360 44.15 11.46
6 0 003 000 400 300 030 030 003 22.67 4.11 19 0.1375 31.68 7.12
1 – 400 300 030 030 003 003 22.67 4.11 19 0.1375 31.57 7.12
7 0 030 300 040 300 040 030 040 28.53 5.05 24 0.1378 25.97 5.56

Nd/wk : number of dialysis sessions assigned to a week; Nd/ns : number of sequential rest days not assigned to a dialysis session; the schedule code indicates the dialysis
duration for each session, 0: no dialysis, 4: 4-h hemodialysis, 3: 3-h hemodialysis, 2: 2-h hemodialysis; the digits of the schedule code indicate the dialysis session
start time, the first digit: morning (start at 9 p.m.), the second digit: early afternoon (start at 1 p.m.), the third digit: late afternoon (start at 5 p.m.). Adequacy criteria
are EKRc ≥ 13.8 and std Kt/V ≥ 2.29, equivalent to the HEMO Study standard dose.

4. Discussion these schedules, and the optimized HD schedules showed better

In the analysis of the conventional regular uniform HD sched-
ule, the schedules that achieved both criteria for the EKRc and
std Kt/V involved at least 12 h of dialysis per week. However,
the schedules that did not satisfy the adequacy criteria involved
weekly dialysis time (Td/wk ) of 8, 9, and 10 h (Table 1). To con-
firm the minimum required weekly dialysis time, Td/wk = 11 h with
2 × 4-h and 1 × 3-h dialysis schedule was tested. This schedule
achieved EKRc = 12.77 ml/min and std Kt/V = 2.06, which did not
satisfy the minimum adequacy value. Therefore, the anuric patients
with 35l total body water required at least 12 h of dialysis through-
out the week.
GA successfully searched towards higher fitness and reached
the optimal solution as shown in Fig. 3, and the optimized sched-
ule was confirmed to have higher fitness value compared to neigh-
boring schedules (Table 3). Further analysis was conducted on the
optimized schedules of GA without considering the burden con-
sumption (α = 0). As shown in Table 1, the EKRc and fitness values
increased as the Nd/wk increased from 3 to 7, indicating that fre-
quent dialysis has a higher clearance rate. However, the std Kt/V
value did not change significantly. As previously reported, the EKRc
is known to be sensitive to interval among treatment sessions and
frequency of treatment, yet the std Kt/V is only sensitive to fre-
quency of treatment [25,30,31]. Moreover, all the schedules opti-
mized by GA, except for Nd/wk = 7, had a Td/wk of 12, which was
the minimum amount of time that anuric patients must be treated
to achieve an adequate dialysis dose both in the EKRc and std Kt/V.
This result may indicate that if the burden of time, price, or any
other form for each session is not considered, 12 h of dialysis treat-
ment per week is most suitable for 35 l patients.
These optimized schedules were compared with the conven-
tional regular uniform dialysis schedules (Tables 1 and 2). First, the
adequacy criteria of EKRc, std Kt/V, and fitness were compared for
overall performance than the conventional regular uniform dialy-
sis schedules. With three days of dialysis treatment similar to the
conventional dialysis schedule (Nd/wk = 3), GA returned the same
schedule for all Nd/ns = 1 and 2: 0 0 0-40 0-0 0 0-040-0 0 0-0 04-0 0 0.
The schedule was similar to the conventional HD schedule, yet the
interval among sessions was equally distributed. Therefore, EKRc,
std Kt/V, and fitness values were slightly higher than those of the
conventional HD. For the case of Nd/wk = 4, GA returned a schedule
with improved EKRc but slightly lower std Kt/V. However, the fit-
ness values of Nd/ns = 0, 1, and 2 were higher than those of the
conventional regular uniform dialysis schedule of Nd/wk = 4. For
Nd/wk = 5 and 6, EKRc and std Kt/V showed a similar trend, as GA
tried to distribute the interval equally among sessions. However,
GA returned exactly the same schedule as the conventional regular
uniform dialysis schedule for Nd/wk = 7.
EKRc and std Kt/V are capable of assessing a variety of HD
schedules, yet these values only evaluate the clearance of dialy-
sis. Therefore, TAC and TAD were also computed on the optimized
schedules of α = 0 to assess the concentration status and the con-
centration variations in the body. TACs decreased as the dialy-
sis frequency increased. This result was expected since the dial-
ysis clearance increases with increased frequency, and TAC is in-
versely correlated to EKRc. Therefore, less urea concentration re-
mained in the body. TACs of the optimized schedule were smaller
than those of the conventional regular uniform dialysis schedules,
which indicated that the optimized schedule more effectively re-
moved urea from a body. TAD, which expressed fluctuation of urea
concentration, values were also reduced for the optimized sched-
ule compared to those of the conventional regular uniform dialysis
schedule. This result is significant, because TAD reflects the dialysis
“unphysiology,” and a healthy person has low TAD value [8,9]. As
previously reported, the TAC and TAD values showed 15.85 mg/dl
and 1.54 mg/dl for a healthy person, and 40–70 mg/dl and 14–
 J.W. Choi et al. / Computer Methods and Programs in Biomedicine 145 (2017) 35–43
Fig. 4. Comparison of conventional regular uniform HD schedule (Conv) and opti-
mized HD schedule by genetic algorithm (GA). (a) Fitness. (b) Time average concen-
tration (TAC). (c) Time average deviation (TAD). Nd/wk : number of dialysis sessions
assigned to a week.
17 mg/dl when treated with the conventional HD schedule. The op-
timized schedule of GA achieved TAC and TAD values around 41–
51 mg/dl and 5.59–13.14 mg/dl, which corresponds to TAC/TAD val-
ues of short daytime hemodialysis, automated peritoneal dialysis,
and nocturnal peritoneal dialysis, according to the previous work
of Lopot et al. [8,9]. This improvement is significant, since the op-
timized schedule showed the TAC/TAD values that were closer to
those of a healthy person by reorganizing the schedule [8,9].
Another significance of the optimized schedules of α = 0 is that
all schedules for Nd/ns = 2 started and ended the week with dial-
ysis sessions. This extends from the fact that GA tried to evenly
distribute the time interval among the sessions, and that GA tried
to minimize the long interval over weekends. Higher sudden car-
41
Fig. 5. Comparison of conventional regular uniform HD schedule (Conv) and opti-
mized HD schedules by genetic algorithm (GA) for patients with different total body
water volume (30 l, 35 l, and 40 l) (a) Fitness. (b) Time average concentration (TAC).
(c) Time average deviation (TAD). Nd/wk : number of dialysis sessions assigned to a
week.
diac arrest rates in dialysis patients on Mondays and Tuesdays have
been reported, and was thought to be due to the increased volume
and potassium accumulated during the weekend, as well as the de-
velopment of post dialysis hypotension from the rapid removal of
large amounts of fluid [17,18]. Although cardiovascular burden and
phosphate kinetics were not incorporated into the mathematical
model in this study, we found that minimizing the weekend inter-
val is effective in optimizing the dialysis schedule to increase dialy-
sis adequacies within the limited time and resource. In comparison
of patients with different total body water contents (Fig. 5), higher
fitness values were observed in patients with lower total body wa-
ter content. However, significant differences were not observed,
except for the fact that as total body water content increased,
a longer minimum total dialysis time was required: Td/wk = 10 h,
12 h, and 14 h for 30l, 35 l, and 40 l total body water patient, re-
spectively.
The optimal schedules identified by GA while considering ad-
ditional burden consumption (α = 1) also showed higher EKRc, std
Kt/V, and fitness values compared to the conventional regular uni-
42 J.W. Choi et al. / Computer Methods and Programs in Biomedicine 145 (2017) 35–43
form dialysis schedule, and showed the similar TAC/TAD trends as
the optimized schedule of α = 0 (Table 4). However, longer dialysis
time per week (Td/wk ) was assigned as Nd/wk increased, so the EKRc
and the std Kt/V for Nd/wk = 5, 6, 7 increased significantly. More-
over, the difference in the fitness among all the schedules were in-
significant. The fitness of the optimized schedule of α = 1 did not
increase for more frequent dialysis schedule in contrast to the re-
sult of GA with α = 0, since longer dialysis time per week (Td/wk )
penalized the fitness. The fitness values of the optimized schedule
with parameters α = 0 for Nd/wk = 5, 6, 7 (Table 2) were recalcu-
lated with α = 1 to compare with the same Td/wk . The recalculated
fitness values were 0.1358, 0.132, and 0.1322, respectively. Among
all the schedules, the fitness values of Nd/wk = 4 and Nd/ns = 0, 1
with a 3-h schedule were the highest. Therefore, it seems to be
the ideal schedule when the burden consumption of treatment is
considered.
This work has shown that a more flexible dialysis schedule can
be utilized for chronic renal disease patients who desire or re-
quire more frequent dialysis using the optimized schedule of GA.
Although GA successfully found the optimal HD schedule, several
limitations should be noted. The results have shown that the bur-
den consumption, α , changes the fitness greatly, and that more fre-
quent dialysis is not always better. However, the cost factor in the
proposed work is objective, thus the value needs to be standard-
ized through quantified conversion. Reduction in the quality of life
may differ among patients, and the burden may vary among geo-
graphic regions due to the differences in the availability of trans-
portation, number of dialysis centers, insurance and other factors.
These should be reflected in the development of a standardized
function. Another limitation of the present study is the mathe-
matical model that represents anuric patients. The current model
was based on the classic variable-volume two-compartment ki-
netic model coupled with theoretical assumptions, which was a
simplified representation of a real patient. Therefore, various pa-
rameters related to the biological conditions, pathological condi-
tions, and various molecules, such as phosphate and creatinine,
should be considered in the future. Lastly, the computation time of
the proposed optimization method should be reduced. The classic
variable-volume two-compartment kinetic model was solved with
the fourth-fifth order Runge–Kutta method to compute the fitness
function of GA. However, the two-compartment kinetic model of
patient can be analytically solved [32,33], and this may reduce the
computation time of schedule evaluation. The reduced optimiza-
tion time will be beneficial for modalities that require adaptive and
frequent adjustment of schedule, such as in the case a wearable ar-
tificial kidney.
5. Conclusion
In this article, the genetic algorithm (GA) was implemented
to optimize hemodialysis schedules for ESRD patients. The algo-
rithm was able to find the optimized schedule for patients among
over 10 million combinations of Nd/wk , Nd/ns , and duration for each
session. Notably, Tte optimized schedules showed that GA tried
to maintain even intervals among sessions. Moreover, GA recom-
mended to start and end the week with dialysis sessions to over-
come issues arising from the long weekend interval, which might
reduce the sudden cardiac arrest rates over weekends. GA also im-
proved the adequacy of HD schedule by showing reduced TAC/TAD
values. Lastly, this research has shown that frequent dialysis is
more physiological and effective, but considering the burden, it
may not be practical. The introduced burden of the dialysis ses-
sion is objective, but this study shows that the burden should not
be ignored in order to adopt the practice of frequent dialysis. This
method will provide new possibilities to simulate the effects and
safety of the HD schedule and can be used to determine patient-
specific schedules for patients in various conditions and life-styles.
Furthermore, this research can be applied to wearable artificial kid-
neys (WAKs) by providing an optimal schedule to set the treating
algorithm.
Conflict of interest
The authors have no financial relationships relevant to this ar-
ticle to disclose. The authors declare no competing interests.
Acknowledgments
This study was supported by The Korean Health Technology
R&D Project of the Ministry of Health and Welfare, Republic of Ko-
rea (HI14C0559).
References
[1] The FHN Trial Group, G.M. Chertow, N.W. Levin, et al., In-center hemodialysis
six times per week versus three times per week, N. Engl. J. Med. 363 (24)
(2010) 2287–2300.
[2] H.J. Jefferies, B. Virk, B. Schiller, et al., Frequent hemodialysis schedules are
associated with reduced levels of dialysis-induced cardiac injury (myocardial
stunning), Clin. J. Am. Soc. Nephrol. 6 (6) (2011) 1326–1332.
[3] M.V. Rocco, R.S. Lockridge Jr., G.J. Beck, et al., The effects of frequent noctur-
nal home hemodialysis: the Frequent Hemodialysis Network Nocturnal Trial,
Kidney Int. 80 (10) (2011) 1080–1091.
[4] W.R. Clark, J.K. Leypoldt, L.W. Henderson, et al., Quantifying the effect of
changes in the hemodialysis prescription on effective solute removal with a
mathematical model, J. Am. Soc. Nephrol. 10 (3) (1999) 601–609.
[5] A.S. Goldfarb-Rumyantzev, A.K. Cheung, J.K. Leypoldt, Computer simulation
of small-solute and middle-molecule removal during short daily and long
thrice-weekly hemodialysis, Am. J. Kidney Dis. 40 (6) (2002) 1211–1218.
[6] J.C. Lee, C.Y. Park, S.W. Choi, et al., Computational dose predictions for com-
bined treatment of hemofiltration with weekly hemodialysis, Comput. Meth-
ods Programs Biomed. 89 (3) (2008) 275–281.
[7] A.C. Yamashita, H. Kawanishi, Kinetics and dose of daily hemofiltration, Blood
Purif. 22 (2) (2004) 14–19.
[8] F. Lopot, A. Valek, Quantification of dialysis unphysiology, Nephrol. Dial Trans-
plant 13 (6) (1998) 74–78.
[9] F. Lopot, B. Nejedly, S. Sulkova, Physiology in daily hemodialysis in terms of
the time average concentration/time average deviation concept, Hemodial Int.
8 (1) (2004) 39–44.
[10] W. Beck, F. Techert, H. Lebsanft, et al., Treatment frequency and efficiency in
hemodiafiltration, Blood Purif. 35 (1-3) (2013) 224–229.
[11] J.S. Berns, L.M. Dember, Can frequent hemodialysis be too frequent? J. Am. Soc.
Nephrol. 24 (3) (2013) 334–336.
[12] S. Klarenbach, M. Tonelli, R. Pauly, et al., Economic evaluation of frequent
home nocturnal hemodialysis based on a randomized controlled trial, J. Am.
Soc. Nephrol. 25 (3) (2014) 587–594.
[13] C.P. Lee, S.A. Zenios, G.M. Chertow, Cost-effectiveness of frequent in-center
hemodialysis, J. Am. Soc. Nephrol. 19 (9) (2008) 1792–1797.
[14] R.S. Suri, B. Larive, Y. Hall, et al., Effects of frequent hemodialysis on perceived
caregiver burden in the Frequent Hemodialysis Network trials, Clin. J. Am. Soc.
Nephrol. 9 (5) (2014) 936–942.
[15] R. Visentin, C. Dalla Man, B. Kovatchev, et al., The university of Virginia/Padova
type 1 diabetes simulator matches the glucose traces of a clinical trial, Dia-
betes Technol. Ther. 16 (7) (2014) 428–434.
[16] C.D. Man, F. Micheletto, D. Lv, et al., The UVA/PADOVA type 1 diabetes simula-
tor: new features, J. Diabetes Sci. Technol. 8 (1) (2014) 26–34.
[17] M.L. Smith, J.A. Scales, T.L. Fischer, Global search and genetic algorithms, Lead-
ing Edge 11 (1) (1992) 22–26.
[18] M. Tabassum, A genetic algorithm analysis towards optimization solutions, Int.
J. Digital Inf. Wireless Commun. 4 (1) (2014) 124–142.
[19] A.G. Bakirtzis, P.N. Biskas, C.E. Zoumas, et al., Optimal power flow by enhanced
genetic algorithm, IEEE Trans. Power Syst. 17 (2) (2002) 229–236.
[20] U. Aickelin, K.A. Dowsland, An indirect genetic algorithm for a nurse-schedul-
ing problem, Comput. Oper. Res. 31 (5) (2004) 761–778.
[21] K.S. Shin, Y.J. Lee, A genetic algorithm application in bankruptcy prediction
modeling, Expert Syst. Appl. 23 (3) (2002) 321–328.
[22] C.R. Reeves, A genetic algorithm for flowshop sequencing, Comput. Oper. Res.
22 (1) (1995) 5–13.
[23] D.K. Kim, J.C. Lee, H. Lee, et al., Calculation of the clearance requirements for
the development of a hemodialysis-based wearable artificial kidney, Hemodial
Int. 20 (2) (2016) 226–234.
[24] J.K. Leypoldt, B.L. Jaber, M.J. Lysaght, et al., Kinetics and dosing predictions for
daily haemofiltration, Nephrol. Dial Transplant 18 (4) (2003) 769–776.
[25] A. Vartia, Equivalent continuous clearances EKR and stdK in incremental
haemodialysis, Nephrol. Dial Transplant 27 (2) (2012) 777–784.
 J.W. Choi et al. / Computer Methods and Programs in Biomedicine 145 (2017) 35–43
[26] R.A. Ward, T. Greene, B. Hartmann, et al., Resistance to intercompartmental
mass transfer limits beta2-microglobulin removal by post-dilution hemodiafil-
tration, Kidney Int. 69 (8) (2006) 1431–1437.
[27] F.A. Gotch, The current place of urea kinetic modelling with respect to differ-
ent dialysis modalities, Nephrol. Dial Transplant 13 (6) (1998) 10–14.
[28] J.K. Leypoldt, B.L. Jaber, D.L. Zimmerman, Predicting treatment dose for novel
therapies using urea standard Kt/V, Semin. Dial 17 (2) (2004) 142–145.
[29] F.G. Casino, T. Lopez, The equivalent renal urea clearance: a new parameter to
assess dialysis dose, Nephrol. Dial Transplant 11 (8) (1996) 1574–1581.
43
[30] J.T. Daugirdas, J. Tattersall, Effect of treatment spacing and frequency on three
measures of equivalent clearance, including standard Kt/V, Nephrol. Dial Trans-
plant 25 (2) (2010) 558–561.
[31] A. Vartia, Effect of treatment frequency on haemodialysis dose: comparison of
EKR and stdKt/V, Nephrol. Dial Transplant 24 (9) (2009) 2797–2803.
[32] P. Korohoda, D. Schneditz, Analytical solution of multicompartment solute ki-
netics for hemodialysis, Comput. Mathemat. Meth. Med. 2013 (2013) 11.
[33] F. Grandi, G. Avanzolini, A. Cappello, Analytic solution of the variable-volume
double-pool urea kientics model applied to parameter estimation in hemodial-
ysis, Comput. Biol. Med 25 (1995) 505–518.