This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3106856, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI

Adaptive Individualized Modeling from

Limited Clinical Data for Precise Anemia
Management
AFFAN AFFAN1 , (Student Member, IEEE), JACEK M. ZURADA2 ,(Life Fellow, IEEE), and
TAMER INANC.3 , (Senior Member, IEEE)
1
Electrical and Computer Engineering Department, University of Louisville,KY, USA (e-mail: affan.affan@louisville.edu)
2
Electrical and Computer Engineering Department, University of Louisville,KY, USA (e-mail: jacek.zurada@louisville.edu)
3
Electrical and Computer Engineering Department, University of Louisville,KY, USA (e-mail: t.inanc@louisville.edu)
Corresponding author: Tamer Inanc (e-mail: t.inanc@louisville.edu).
This work was supported by NSF under grant 1722825.

ABSTRACT It is challenging in practice to achieve a steady-state value for external human recombinant
erythropoietin (EPO) dosage to be administrated to maintain Hemoglobin (Hb) level within the desired range
of 11-12 g/dl based on population-based models for anemia management due to inter-and intra-variability of
the patients. On the other hand, Pharmacokinetic (PK) and Pharmacodynamic (PD) characteristics can vary
for the patients over the course of treatment due to aging and other life changes. To address the inter-and
intra-variability in anemia management, the semi-blind robust identification approach is proposed to obtain
individualized patient models using limited number of clinical patient data. Semi-blind robust identification
utilizes the effect of the initial condition during system identification to reduce the identification error. To
reflect the patient’s true dose-response relation as time passes and ensure the suitability of the individualized
model for the controller, the model (In)validation technique is discussed to provide appropriate mathematical
evidence about the suitability of the individualized model for dose prediction and controller design via
testing it on new clinical data of the particular patient. One-step-ahead prediction results are shown for
identified individualized patient models. The individualized patient models provide decision support to the
clinicians about EPO dosage to avoid undershoot or overshoot of Hb level. Minimum mean squared error
(MMSE) is calculated for the predicted values obtained by the models identified using semi-blind robust
identification with and without the model (In)validation against clinically acquired EPO-Hb data.

INDEX TERMS Anemia management, adaptive modeling, drug dosing, individualized patient modelling,
model (In)validation, robust system identification.

I. INTRODUCTION level at an appropriate range, the CKD patients are treated

HE chronic kidney disease (CKD) is a heterogeneous
T disorder affecting the structure and functionality of hu-
man kidneys [1]. The kidneys of patients with CKD lose the
ability to filter the blood gradually over time and reduce the
production of erythropoietin (EPO). The EPO is the glyco-
protein produced by kidneys to regulate the functionality of
red blood cells in bone marrow [2]. The patients of CKD
suffer from low hemoglobin (Hb) levels and deficiency of
oxygen-carrying red blood cells. Other chronic diseases that
require anemia management include cancer, diabetes, and
heart diseases. The patients suffering from these diseases
are required to be treated for a longer time with a stable
and robust drug management system. To maintain the Hb
with external human recombinant erythropoietin (EPO) [3].
According to the National Kidney Foundation’s Dialysis, the
Hb level should be in the range of 11 and 12 g/dL in the
response of EPO dosage [4].
Presently, the medical centers have developed their anemia
management protocols (EPO management systems) based
on the average (population-based) response to the medica-
tion. These traditional rule-based dose adjustment protocols
usually lead to inappropriate EPO doses and increased Hb
fluctuations. These phenomena have been linked to an in-
creased risk of serious cardiovascular problems and throm-
botic events [5], [6]. Feedback control techniques for EPO
dose individualization would therefore be of great impor-

VOLUME 4, 2016 1

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3106856, IEEE Access
A. Affan et al.: Adaptive Individualized Modeling from Limited Clinical Data for Precise Anemia Management
tance to make the anemia treatment more targeted and ef-
ficient [7]– [14]. Most of the existing techniques however
are population-based [7]– [11] and treat individual patients
with a one-size-fits-all model [15]. While population-based
models may be useful in the analysis of drug properties at
large, however, they are not well suited to guide the treatment
of individual patients. To address the impact of inter-and
intra-individual variability in dose-response characteristics,
individualized patient-specific models are needed instead of
traditional population-based models, and assurance about
model suitability, i.e., model (In)validation for controller
design is required as well.
Many researchers have proposed different techniques to
develop does-response models for number of different dis-
eases. Bayesian-based drug delivery using population patient
data is discussed in [16]. Bayesian framework for warfarin
dosing is presented in [17]. In [18], control oriented models
are developed for Type 1 Diabetes to capture the intra-patient
variability of insulin drug patients. Artificial Intelligence-
based neural network models are discussed in [19]– [21].
Some researchers attempted to identify individual patient
models in [22]. In [5] the focus was to predict the value
for EPO instead of Hb, which is not desired as predicting
Hb level gives values of EPO but not true for vice versa.
However, most of these models are based on predetermined
model structure and noise distribution, that are not suit-
able for anemia management as each patient poses different
characteristics. The models obtained by the classical system
identification techniques do not yield good results as they
assume that predefined model structure and model order are
close to the actual system (patient) and one mathematical
model works for all dynamics regardless of complexity and
disturbances in the system (patient). The modeling stage
should include the effect of all disturbances and uncertainties
which are being introduced during operations. In contrast
to classical identification techniques, robust system identi-
fication considers system (patient) uncertainties, unmodeled
dynamics, and model complexity, i.e., there is no assumption
on the model order, uncertainties, and noise affecting data.
The robust system identification procedure does not re-
quire large measurement data set or information of mea-
surement noise nor the information on the structure of the
system (patient) model to be identified is required. In this
technique, the information on the maximum gain of the
model, K, the stability margin of the model response r,
and a bound on the noise is required. Depending on the
nature of a posteriori information, the robust identification
techniques may lead to different identification methods. The
selection of a robust system identification method depends
on the data type as time-domain data and frequency domain
data pose different characteristics and different techniques
are applied to obtain a model for the data type [6], [9].
For frequency domain data, the H∞ -identification technique
is implemented, which calculates the uncertainty bounds
in terms of the H∞ norm. For time-domain data, the l1 -
identification technique is implemented, which calculates an
2 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
l1 -error bound. However, sometimes both time and frequency
domain data may be available for a single system. In this case,
the mixed H∞ /l1 -robust identification procedure is used [10].
The behavior of a system (patient) is highly affected by the
initial conditions at t=0. In anemia management, these could
be due to previous drugs used by the patient. The techniques
such as l1 and H∞ in robust identification assume zero initial
conditions for the system, which highly affects the result
of the identification process. To solve this issue, the semi-
blind identification technique is introduced in [11]– [13],
which accommodates the effect of the initial conditions of
the patient in identification process.
As the patients of CKD are treated with EPO for an
extended time. It is reported in the literature that the Pharma-
cokinetic (PK) and Pharmacodynamic (PD) characteristics
vary over the course of treatment due to aging, the addition of
new medication, new disease, and change in life habits [23]–
[25]. As the patient’s PK and PD changes, the individualized
patient model identified earlier may (In)valid to predict and
not suitable for controller design. Therefore, the individual-
ized patient models are required to be adaptive in order to
reflect the current patient’s dose-response status. The model
(In)validation technique tests the identified model on the
new clinically acquired data. The model is adapted if exiting
model (In)validates [11].
In this research work, the individual patient models for
CKD patients are obtained using a robust system identifica-
tion technique and patient-specific time-domain clinical data.
For model identification, the semi-blind system identification
technique is used [11]. The system model is a combination
of a nominal model and uncertainty bound [15]. The identi-
fication technique develops a nominal model and uncertainty
bound using patient-specific dose-response (EPO-Hb) data.
For the efficient controller, an appropriate mathematical ev-
idence is required about the model suitability for controller
design. To address this requirement, the identified models
are (In)validated using new patient-specific clinical data. We
summarize our contributions as follows:
• We present a semi-blind robust system identification
technique to identify individualized patient models for
CKD patients using limited number of time-domain
clinical data.
• We present the model (In)validation framework to show
the suitability of models for prediction and controller
design by testing the models on the new clinical dataset.
The remainder of this paper is organized as follows; First,
we discuss the semi-blind robust system identification in
section-II. Section III discusses the model (In)validation and
section-IV provides a complete algorithm of the semi-blind
robust identification with the model (In)validation. Section-V
presents the patient model results obtained by the one-step-
ahead prediction along with the error analysis followed by
the conclusion.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3106856, IEEE Access

A. Affan et al.: Adaptive Individualized Modeling from Limited Clinical Data for Precise Anemia Management

II. SYSTEM IDENTIFICATION

The system (patient) model is the set of mathematical equa-
tions that describes its behavior in response to the input.
The purpose of system identification is to develop a low
order model of the system (patient) by using finite, noisy
measurement of clinical dose-response data, which can be
controllable and observable by implementing a suitable con-
troller and observer, respectively. In this work, the system
(patient) model is the combination of a parametric and non-
parametric portion given as:
FIGURE 1. Framework for semi-blind identification
G = Gp + Gnp (1)
The non-parametric portion Gnp describes the internal be-
i−1
havior of the system. This portion requires less prior knowl- where g0 = DG ; gi = CG (AG ) BG . The solution to (4)
edge of the system such as a model for internal behavior involves solving a Bi-Affine matrix, which is a non-convex,
of physiological systems. On the other hand, the parametric NP-hard problem. The above problem can be converted to the
portion Gp describes the input to output relation of the convex problem as mentioned in [11], which preserves the
system. The performance of system identification is highly controllability and observability of the system. The convex
affected by the initial conditions due to parametric portion.In problem can be defined as follows:
most of application of system identification zero initial con- Problem 3: Determine G (z) = Gp (z) + Gnp (z), which is
ditions are considered. However, in anemia management, compatible with a priori and a posteriori information, such
it is important to take into account that patient may have that τ is a non-empty set:
medical history due to other diseases and medications. It n
−

o
τ (y) = G (z) ∈ S : yi − TgN u+ i + ΓN


is also important to mention that for this application, the gu i
(5)
patient data is not steady state-data, therefore, the effect of 

− 
where  ΓN gu  ≤ γ Ku ; i = 0, 1, · · · , N − 1 and TgN

past inputs is still present. Therefore, to reduce the error in i
identification, it is important to use the initial conditions in is the Toeplitz matrix and ΓN g is the Hankel matrix. The first
the model identification process for the application of anemia part of the τ set corresponds to the plant (patient) response
management. To incorporate the effect of initial conditions, for input u and the latter part provides information for system
the semi-blind identification technique is introduced. Next, response for past inputs u− . This problem can be solved by
we formally define the problem [6]. following LMIs as presented in [11]:
Problem 1: Considering (1), obtain the patient model G
KR−2 (TgN )T
 
which gives output yk as Hb level in response to input uk M (g) = ≥ 0,
(TgN ) KR2
of EPO dosage to patients of CKD.
y − (TuN pP + TuN g) − ΓN −
 
yi = Gui + ηi , (2) g u ∈ N,
−
|ηi | ≤ i , (3) −γKu 6 ΓN
g u 6 γKu ,

where γ, Ku , p, P represent system gain, bound on the norm

where ηi is the model noise and i is the maximum bound on
of sequence u− , affine parameters, and the parametric portion
noise.
of the model, respectively. For model identification, u = vi −
vi−1 , where v is the EPO input to the system. Fig. 1 shows the
A. SEMI-BLIND ROBUST IDENTIFICATION
framework for semi-blind identification. As the measurement
The response of dynamic systems is highly affected by its
of Hb is also affected by the past inputs u− , the parametric
state at t = 0. The semi-blind robust identification technique
part should be an integrator to accommodate all the effects of
incorporates the effect of initial conditions of the system
the inputs before t = 0, i.e initial conditions.
(patient) [11], [12]. The problem for semi-blind identification
Heretofore, we have discussed the semi-blind robust sys-
can be defined as follows:
tem identification to obtain individualized patient models
Problem 2: Given input sequence u, output sequence y, noise
from limited number of patient-specific clinical data, pos-
bound ∈ N , maximum stability gain and characteristics of
sibly corrupted by noise. However, as the the time passes
past input u− , determine G (z) = Gp (z) + Gnp (z), which
the patient’s characteristics may change due to aging, change
is compatible with priori and posteriori information, such that
in food habits, and new medication for a new disease. The
τ is a non-empty set. where τ is defined as:
change in patient’s status may reflect in the patient’s drug-
N dose response model identified at an early stage, and the
. X
−1 −
gi uN −i + Cg AN ΓN


τ (y) = yi = g g u i=0
(4) model may not match the patient’s current dose-response
i=0 characteristics. Therefore, this model may not be suitable
VOLUME 4, 2016 3

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3106856, IEEE Access
A. Affan et al.: Adaptive Individualized Modeling from Limited Clinical Data for Precise Anemia Management
for controller. To adapt the model to patient’s current status,
the next section discusses the model (In)validation technique
which provides mathematical evidence about the suitability
of model for controller design by testing the model on data
unseen by identification process.
III. MODEL (IN)VALIDATION
The actual plant consists of a nominal system model and
uncertainty. Once the individualized system (patient) model
has been identified by using a semi-blind robust identification
technique, then the model needs to be validated (or invali-
dated) by utilizing additional patient-specific data that is not
seen by the identification process before it is used in the con-
troller design. The theory offers the tools to design the model
which is robust and stable in theory but no information is
available regarding the stability and robustness of the model
in practice. This issue arises due to the system uncertainties,
unmodeled dynamics and model complexity in the system.
The model (In)validation techniques provide evidence about
the usability of the model under these uncertainties for
controller design to some extent by testing the model on a
new experimental data set. By assuming multiplicative and
additive noise, the problem can be stated as follows:
Problem 4: Given M experimental data points (yi , ui ), the
nominal model G (z) ∈ S, descriptions of admissible noise
N , uncertainty ∆ and initial conditions x0 , determine if there
exists at least one triple (η, ∆, x0 ) that can reproduce the
available experimental data by the following equation [11]:
y = (I + ∆) Tg u+ + Tgic x0 + η,


(6)
where u+ is the input after t = 0, Tg maps the input to the
output, whereas Tgic maps the initial conditions to the output.
The above problem has a term Tgic x0 , where x0 is the initial
condition and unknown. During the identification, this term
FIGURE 2. The (In)validation framework for semi-blind robust identification.
can be replaced with some term representing the effect of
these initial conditions such as u− ∈ U− , Therefore, (6) can
be modified as follows:
− (In)validation conditions are not satisfied, i.e., the model is
y = (I + ∆) Tg u+ + ΓM


g u + η, (7)
where U− and ΓM g denote the past admissible inputs and Han-
kel matrix, respectively. Figure 2 shows the framework for
the (In)validation for the models obtained by the semi-blind
robust identification technique. To avoid the issues related to
4 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
the solution of the non-convex problem, the following convex
relaxation is considered as discussed in [11]. Fig. 3 shows
the alternative setup for the model (In) validation, where
.
measurement noise is also affected by ∆, η = (1 + ∆) η̄.
Equation (7) can be modified as follows:
−
y = (I + ∆) Tg u + ΓM


g u + η̄ (8)
Equation (8) is satisfied if a triple (u− , η̄, ∆) exists and
k∆k∞ < 1. Further details of convex relaxation of the model
(in)validation process, please see [11].
FIGURE 3. The convex relaxed (In)validation framework for semi-blind robust
identification.
IV. ALGORITHM
Section II and III provide theoretical insight on semi-blind
robust identification, and model (In)validation. Algorithm-
1 shows the semi-blind robust system identification frame-
work for individualized predictive dose-response modeling.
The algorithm starts with the patient ID/tag, the number of
clinical EPO-Hb data points to be used for identification, Nt ,
stability bound, r, and the initial reduced model order, O.
As this is not a fixed model order, the algorithm will select
the appropriate model order based on model (In)validation
at step 4. This value is to start the algorithm. Steps 1-3 in
Algorithm-1 produce models with model orders equal to the
number of data points used in the identification step. Based
on these parameters, the parametric and non-parametric por-
tions of the models are identified using a semi-blind robust
identification technique, and a reduced-order model is gener-
ated at steps 4-5 as shown. The next step is to determine the
number of experimental (clinical patient) data points, which
have not previously been used in identification for model
(In)validation. The algorithm uses the same number of data
points for the model (In)validation process as the numbers
of coefficients. Then model (In)validation is performed as
explained in Section-III.
If the conditions for model (In)validation are satisfied, then
the algorithm displays the validated individualized patient
model transfer function for the patient. Alternatively, if the
invalidated, then the reduced model order is increased for
the same r. If the reduced-order model is greater than 6,
then r is increased. However, if r > 2 then the number
of data points Nt used for model identification is increased
and the algorithm returns to step 1. In the last case, the
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3106856, IEEE Access

A. Affan et al.: Adaptive Individualized Modeling from Limited Clinical Data for Precise Anemia Management

user must increase the number of clinical data points used

in the identification process and re-run the algorithm. The
limit on the maximum low order model, which is 6 in our
experiments, can be increased but low order models are
preferred for the controller design due to a smaller number
of parameters to tune in controller design. Low order models
have also been preferred in clinical applications as well since
the explanation of the resulting models is easier for low order
models.
Algorithm 1 Semi-blind robust identification with model
(In)validation.
Input: Patient-ID, Patient-Specific EPO-Hb data, Nt , r, O.
Output: Discrete-time individualized patient model, Gr
1: Set Nt , r = 1.01, O = 3
2: function Semiblind (Nt , r, EPO-Hb data, ID);
3: return(G)
4: function ReduceModelorder (G, O);
5: return(Gr ) FIGURE 4. Prediction results for patient-1 model obtained by semi-blind
robust identification with model (In)validation.
6: Select number of new data points, Mt used for model
(In)validation equal to the order of Gr .
7: function ModelInvalidation(Gr ,Mt );
mance of patient models obtained by the semi-blind robust
8: if k∆k∞ < 1 then identification technique is analyzed by calculating minimum
9: print(model validated.) mean squared error (MMSE) between measured clinical Hb
10: return(Gr ) data and predicted Hb level using reduced-order models. The
11: else performance of semi-blind robust identification with model
12: print(model (In)validated.) (In)validation is compared semi-blind robust identification
13: If O < 7 then without model (In)validation by computing the MMSE val-
14: O = O + 1; repeat step 4. ues. The model is checked for (In)validation at each time
15: else step (weekly), n, and updated if the model is not validated.
16: if r < 2 The mathematical expressions for patient-1 models using
17: Set O = 3; Increase r; semi-blind robust identification with model (In)validation are
18: repeat step 2. given below:
19: else
 0.9z4 +0.07z3 +0.02z2 −0.1z+0.45
20: Set O = 3; r = 1.01; Increase Nt ;  z 4 −0.56z 3 −0.26z 2 −0.15z+0.01 10≤ n≤24

21: repeat step 2.  3.3z4 +1.26z3 −1.29z2 −1.6z+0.27

n=25
z 4 −0.58z 3 −0.77z 2 −0.05z+0.41 (9)
22: endif G1 (z)=
2.2z 4 +4z 3 −1.63z 2 −0.98z+2.3
 z 4 −0.99z 3 −0.46z 2 +z−0.5 n=26
23: endif 

 5 4 3 2
−0.1z −0.1z +0.6z −0.2z −0.3z+0.2
24: endif z 5 −0.9z 4 −0.4z 3 +0.4z 2 −0.4z+0.3 27≤ n≤50

25: exit
V. SIMULATION RESULTS
For the simulation purposes, the clinical data of 50 patients
has been collected from the University of Louisville, Kid-
ney Disease Program. This study is IRB approved by the
Human Subjects Protection Program Office at University of
Louisville. Due to the space limitations results for only four
patients with the most challenging cases from the available
database are shown in the paper. These patients represent
the good and bad responder to the medication. Each patient
received three EPO dosages per week and the Hb level
was tested once a week. To match the input-output data
dimension, the average of three EPO dosages is used as v,
while input u for the identification is the difference between
v values. One-step-ahead prediction is used to show the
predicting capabilities of the identified models. The perfor-
VOLUME 4, 2016
The model prediction results of the above model for patient-
1 are shown in Fig. 4. In the following figures of semi-
blind robust identification simulation results, the red line
with square markers shows the actual clinical Hb values of
the patient, the solid blue line shows the model prediction
results of the full order model obtained using semi-blind
robust identification technique, the green line with diamond
markers shows the prediction results of the reduced-order
model, the magenta vertical dashed line shows the number
of data points, Nt , used in the identification process for the
first time and the cyan vertical dashed line represents the
points where the model is (In)validated and then updated
(i.e., the model could not be validated with the recent pa-
tient data through the model (In)validation algorithm and
therefore model identification algorithm is rerun as shown
in Algorithm-1.) Finally, the vertical blue bars in Figures
4 through 11 show the weekly EPO dosages. It is impor-
5

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3106856, IEEE Access

A. Affan et al.: Adaptive Individualized Modeling from Limited Clinical Data for Precise Anemia Management

tant to mention that whenever the model is (In)validated,

all previously available data points are used to update the
model. The full order model is equal to the Nt + 2, model
consist of parametric and non-parametric portion, and the
reduced order model range from 3rd order to 6th order model
(not limited). The selection of reduced-order model is based

TABLE 1. Model (In)Validation Results of Patient-1

Data Range (Weekly) k∆k∞

10-24 0.93
25 0.99
26 0.84
27-50 0.86

on model (In)validation conditions, the model order which

satisfies the model (In)validation condition of k∆k∞ < 1
is selected as the final reduced-order model. The identified
model is checked for model (In)validation at each time step FIGURE 6. Prediction results for patient-13 model obtained by semi-blind
(week) and updated as the model is (In)validated. The model robust identification with model (In)validation.
(In)validation results for patient-1 are given in TABLE 1.
All values of k∆k∞ are less than one for patient-1, which
satisfies the condition for model (In)validation. It is important presence of three cyan vertical lines. It is worth mentioning
to mention that in clinical applications personalized drug that all the previously available data points are used to update
dose-response models from as few numbers of clinical data the model whenever a model is (In)validated e.g., at week
as possible is desired. 25, the previous 24 data points are used in the identification
In this application, only 5-10 data points corresponding process to update the model. For comparison, the mathemat-
to 5–10-week data collection time from anemia patients are ical model of patient-1 using identification process without
used in the identification algorithm. Therefore, more data model (In)validation is shown in (10) and prediction results
points are not used in the identification algorithm since it are shown in Fig. 5.
5 − 0.3 z 4 + 0.1 z 3 − 0.1 z 2 − 0.01 z + 0.002
would cause a delay in the treatment (through the proposed G1 (z)= 0.5zz5 (10)
− 1.7 z 4 + 0.8 z 3 − 0.3 z 2 + 0.14 z − 0.01
algorithm) of patients. The identified model by the semi-
blind identification for the patient-1 does (In)validate, i.e., Model of patient-1 shown in (10) is identified using 10
the resulting individualized model is updated three times at data points. As this model is not processed through model
weeks 25, 26, and 27. This is represented in Fig. 4 with the (In)validation, it is therefore, not possible to confirm that the
model is suitable for controller design as time passes. This
is an alarming situation because the patient’s condition may
change over time due to the spread of disease or the patient
suffers from another disease. However, model (In)validation
can provide an alert for situations where a model is not
suitable for prediction and controller design anymore. The
benefit of model (In)validation can be seen in the sense of
prediction error as shown in Fig. 12. The prediction error
for models obtained using model (In)validation is lower
than the model obtained without a model (In)validation. The
mathematical expressions for patient-13 models obtained by
semi-blind robust identification with model (In)validation are
as follows:
0.6s3 +0.4s2 +0.1s+0.2




 s3 −0.7s2 −0.1s−0.1 n=5
5 4 3 2


 0.6s +0.2s −0.2s +0.1s +0.1s−0.03

5 4 −0.2s3 +0.004s2 +0.01s+0.004 6≤n≤7
 s −0.8s



G13 (z)= 0.7s3 −0.2s2 −0.1s+0.2 (11)
 s3 −1.3s2 +0.4s−0.1 8≤n≤9
0.1s4 +0.5s3 +0.4s2 +0.5s+0.2





 s4 −0.9s3 +0.3s2 −0.3s−0.03 10≤n≤16

4 3 2
 −0.2s −0.1s −0.03s −0.2s+0.1



s4 −0.8s3 +0.5s2 −0.03s−0.6 17≤n≤51

FIGURE 5. Prediction results for patient-1 model obtained by semi-blind Patent-13 is a challenging patient model because the values
robust identification without model (In)validation. of Hb are varying 8g/dl to 14g/dl and EPO dosages are
6 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3106856, IEEE Access

A. Affan et al.: Adaptive Individualized Modeling from Limited Clinical Data for Precise Anemia Management

FIGURE 7. Prediction results for patient-13 model obtained by semi-blind FIGURE 8. Prediction results for patient-19 model obtained by semi-blind
robust identification without model (In)validation. robust identification with model (In)validation.

also varying a lot. It can indicate that the patient is a poor
responder to the medication and finding an optimal dosage
is a difficult task for this patient. The prediction results of
patient-13 in Fig. 6 show that it was difficult to find an
appropriate model which satisfies all the conditions even at
earlier time steps, the model is updated three times within
the first 10 weeks. The interesting part is between the week
17 and 23, where the EPO dosage is zero. This phenomenon
is challenging for the identification algorithm as the model
identified is for non-zero EPO dosages, therefore, it will
be challenging for the model to be validated for zero EPO
dosages. However, the semi-blind robust identification tech-
nique updated the model to capture the dynamics of this por-
tion as soon as the identification algorithm sensed this trend
in the data. The model (In)validation results for the patient-
13 are shown in TABLE 2. All values of k∆k∞ are less than
one which satisfies the conditions of model (In)validation.
To compare the performance of models shown in (11), the
mathematical model of patient 13 using identification process
without model (In)validation is shown in (12) and prediction
results are shown in Fig. 7.
0.7z 5 − 0.5z 4 + 0.1z 3 − 0.1z 2 − 0.03z + 0.003
G13 (z) =
z 5 − 2z 4 + 1.3z 3 − 0.5z 2 + 0.2z − 0.02
The prediction error in models identified by semi-blind
robust identification with model (In)validation is smaller than
of models obtained by semi-blind robust identification with-
out a model (In)validation. It can be seen by the prediction
results and error table that semi-blind robust identification
with the model (In)validation performed better by incorporat-
ing the fluctuations in the patient characteristics which is not
possible without a model (In)validation. The mathematical
expressions for patient-19 models obtained by semi-blind
robust identification with model (In)validation are as follows:
0.9z 5 −0.5z 4 −0.02z 3 +0.04z 2 −0.1z+0.01



 z 5 −2.1z 4 +1.6z 3 −0.7z 2 +0.2z−0.02
n=5
 0.5z5 +0.2z4 −0.6z3 +0.3z2 −0.1z+0.004 6≤ n≤10


G19 (z)= z 5 −2.1z 4 +1.6z 3 −0.6z 2 +0.1z−0.005 (13)
0.6z 4 −0.3z 3 +0.4z 2 −0.6z+1.3
 11≤ n≤ 12

 z 4 −1.3z 3 +0.5z 2 −0.3z+0.04
 0.7z5 −0.01z4 +0.5z3 +0.9z2 −0.7z+0.6


13≤ n≤50
z 5 −1.2z 4 +0.5z 3 −0.04z 2 −0.5z+0.2
The patient-19 is a very interesting case as the EPO dosage
has frequent changes and Hb values are ranging between
8g/dl and 14g/dl as shown in Fig. 8. This patient does not
qualify as a poor responder to medication because the change
between EPO dosages is low even though Hb values are
varying a lot. Therefore, it becomes an interesting scenario
where the patient response is not ordinary. However, the iden-
tification process with the model (In)validation can identify
(12)
the model and timely update the model to represent the true
dose-response characteristics of the patient. TABLE 3 shows
the model (In)validation results for the patient-19. The values
of k∆k∞ are less than one which satisfies the condition

TABLE 2. Model (In)Validation Results of Patient-13

TABLE 3. Model (In)Validation Results of Patient-19

Data Range (Weekly) k∆k∞

5 0.96 Data Range (Weekly) k∆k∞
6-7 0.81 5 0.93
8-9 0.96 6-10 0.96
10-16 0.98 11-12 0.96
17-51 0.99 13-50 0.88

VOLUME 4, 2016 7

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3106856, IEEE Access

A. Affan et al.: Adaptive Individualized Modeling from Limited Clinical Data for Precise Anemia Management

FIGURE 9. Prediction results for patient-19 model obtained by semi-blind FIGURE 10. Prediction results for patient-21 model obtained by semi-blind
robust identification without model (In)validation. robust identification with model (In)validation.

for the model (In)validation. For the sake of comparison,  0.9z5 −0.3z4 −0.1z3 +0.002z2 −0.2z+0.1
the model of patient-19 without model (In)validation is as 
 z 5 −1.3z 4 +0.2z 3 +0.1z 2 +0.03z+0.03 5≤ n≤28
follows:  −0.3z4 +0.3z3 +3z2 +0.02z−2

29≤ n≤38
G21 (z)= z 4 −0.3z 3 −1.2z 2 −0.03z+0.4 (15)
0.6z 4 +0.1z 3 −1.5z 2 +1.4z+2.4
 z 4 +0.2z 3 −1.2z 2 −0.2z+0.3 39≤ n≤ 41
1.1z 5 − 0.9z 4 0.2z 3
− 0.08z 2
− 0.1z − 0.0004

+ 
−1.7z 3 +4.9z 2 −5.6z+3
G19 (z) = (14)

42≤ n≤51
z 5 − 2z 4 + 1.5z 3 − 0.6z 2 + 0.2z + 0.00014 z 3 −1.9z 2 +1.2z−0.24

0.9z 5 − 0.3z 4 − 0.1z 3 − 0.01z 2 − 0.2z + 0.1

G21 (z) = (16)
The dose-response for the above model for patient 19 is z 5 − 1.3z 4 + 0.2z 3 + 0.1z 2 + 0.02z + 0.03
shown in Fig. 9. The model shown in (14) is identified using The model (In)validation results shown in TABLE 4 and
5 initial data points and the reduced model order is 5th order. prediction error is shown in Fig. 12. As documented in these
The model orders in (13) and (14) are almost similar however figures and tables, the semi-blind robust identification tech-
the difference is the model (In)validation and advantage of it nique identified individualized patient models and updated
can be seen in Fig. 12 in the sense of prediction error values. the model to capture the time-varying patient dynamics with
The error between actual clinical data of Hb and predicted Hb
by model is 1.6 and 2.65 given by (13) and (14), respectively.
Patient-21 models obtained by semi-blind robust identi-
fication with model (In)validation are shown in (15) and
without the model (In)validation are shown in (16). The pre-
diction results for models (15) and (16) are shown in Fig. 10
and Fig. 11, respectively. For the initial model identification,
only five data points are used as shown in Fig. 10 and Fig. 11
with the magenta color vertical lines. In Fig. 10, the model
has been updated three times. It is interesting to discuss that
between weeks of 14 and 17, the model is validated for zero
EPO dosages even though non-zero EPO dosages are used
for initial model identification but between weeks 38 and
44, the model is updated two times. This can be explained
by observing the change in EPO dosages. As the change in
EPO dosages is high between weeks of 38 and 44, the iden-
tification process therefore, has to go through an extensive
search to find the model which can be used to accommodate
such fluctuations in the EPO dosages. However, the change
in EPO dosages between weeks 14 and 17 is small, therefore, FIGURE 11. Prediction results for patient-21 model obtained by semi-blind
the model is easily validated during this period. robust identification without model (In)validation.

8 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3106856, IEEE Access
A. Affan et al.: Adaptive Individualized Modeling from Limited Clinical Data for Precise Anemia Management
TABLE 4. Model (In)Validation Results of Patient-21
Data Range (Weekly) k∆k∞
5-28 0.87
29-38 0.92
39-41 0.94
42-51 0.34
low error within an appropriate time frame for each patient.
This is especially important for efficient controller design
and hence for finding an individualized drug-dose regimen
for the patients. The MMSE values for the identified models
using model (In)validation are considerably smaller than of
models identified without a model (In)validation and accept-
able. This shows that the semi-blind robust identification
performs better with the model (In)validation to identify the
individualized models by using a considerably lower number
of clinical patient-specific data points and by updating the
models as well using time-domain clinical patient-specific
data.
FIGURE 12. Error Comparison between semi-blind robust identification with
and without model (In)validation.
VI. CONCLUSION
Patients with CKD suffer from the deficiency of red blood
cells due to the damaged kidneys and require a robust and
stable drug management system to maintain the Hb level
between 11-12 g/dL in response to the EPO dosage. In
this research work, the individualized patient models have
developed using a semi-blind robust identification technique
considering non-zero initial conditions. The measurement
data of 50 patients from the University of Louisville, Kidney
Disease Program has been used to verify the performance of
the approach. Due to the space limitations results for only
four patients were shown in the paper. The ones shown in
the paper show the results for the most challenging cases
of the available data, especially for patients 13, 19, and 21.
To design the controller based on the identified models, the
model (In)validation framework is also described and applied
to the identified models. This is because as time progresses
during the treatment of a patient there will be some changes
VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
in the response of the patient due to the progress of the
disease, life changes of the patient and at one point the
model may not any more accurately represent the specific
patient and thus it will need to be updated, i.e., adapted
accordingly to those changes. In this paper, we investigate
the model (In)validation technique as a model adaptation
strategy for the patient-specific individualized models. It
is shown by the simulations using time-domain patient-
specific clinical data that the model (In)validation algorithm
leads to efficient adaptive models that capture the varying
model dynamics. Performance of these identified models
is compared with models identified by semi-blind robust
identification without the model (In)validation by calculating
MMSE. The MMSE results show that the models obtained
via model (In)validation performed very well, and these
models are shown to be stable and robust based on the
model (In)validation results. The future aim is to design the
controller based on the individualized models identified by
the semi-blind robust identification technique.
ACKNOWLEDGMENT
The authors would like to thank Drs. Michael E. Brier and
Adam E. Gaweda at Department of Medicine of the Univer-
sity of Louisville for their valuable discussions and providing
the clinical data. Funding from NSF under grant 1722825 is
gratefully acknowledged.
REFERENCES
[1] A.S. Levey, and J. Coresh, "Chronic kidney disease". The lancet, vol.379,
no. 9811, pp.165-180, 2012.
[2] E. Moore and R. Bellomo, "Erythropoietin (EPO) in acute kidney injury,"
Annals of Intensive Care, vol. 1, no. 1, pp. 1-10, 2011.
[3] M. Panjeta, I. Tahirovic, J. Karamehic, E. Sofic, O. Ridic and J. Coric,
"The Relation of Erythropoietin Towards Hemoglobin and Hematocrit in
Varying Degrees of Renal Insufficiency," Materia socio-medica, vol. 27,
no. 3, pp. 144-148, 2015.
[4] F. Valderrábano, "Erythropoietin in chronic renal failure," Kidney Interna-
tional, vol. 50, no.4, pp. 1373-1391, 1996.
[5] J. D. Martín-Guerrero, G. Camps-Valls, E. Soria-Olivas, A. J. Serrano-
López, J. J. Pérez-Ruixo and N. V. Jiménez-Torres, "Dosage individualiza-
tion of erythropoietin using a profile-dependent support vector regression,"
IEEE Trasaction on Biomedical Engineering, vol. 50, no.10, pp. 1136-
1142, 2003.
[6] R. S. Sánchez-Peña and M. Sznaier, "Robust Systems Theory and Appli-
cations", New York: Wiley, 1998.
[7] L. Ljung, "System identification-theory for the user, 2nd ed., NJ: prentice-
hall, Upper Saddle River, 1999.
[8] M. C. Mazzaro, P. A. Parrilo and R. S. Sánchez Peña, "Robust Identifi-
cation: An approach to select the class of candidate models," Internation
Journal of Control, vol. 74, no. 12, pp. 1210-1218, 2001.
[9] J. Schoukens, Y. Rolain and R. Pintelon, "On the use of parametric and
non-parametric noise-models in time- and frequency domain system iden-
tification," in 49th IEEE Conference on Decision and Control, Atlanta,
GA, USA, 2010, pp. 316-321.
[10] T. Inanc, M. Sznaier, P. A. Parrilo and R. S. Sánchez Peña, "Robust identifi-
cation with mixed parametric/nonparametric models and time/frequency-
domain experiments: theory and an application," IEEE Transactions on
Control Systems Technology, vol. 9, no.4, pp. 608-617, 2001.
[11] M. Wenjing, "Semi-blind Robust Identification and Model (In)Validation,"
Ph.D. thesis, Pennsylvania State University, Pennsylvania, USA, 2006.
[Online]. Available: https://etda.libraries.psu.edu/catalog/8026
[12] Y. Muhittin, "Robust Systems Theory Applications to Macroeconomic
Stabilization Problems," Ph.D. thesis, Pennsylvania State
University, Pennsylvania, USA, 2005. [Online]. Available:
https://etda.libraries.psu.edu/catalog/6870
9
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3106856, IEEE Access

A. Affan et al.: Adaptive Individualized Modeling from Limited Clinical Data for Precise Anemia Management

[13] E. Akabua, T. Inanc, A. Gaweda, M. E. Brier, S. Kim and J. M. Zurada,
"Individualized model discovery: The case of anemia patients," computer
methods and programs in biomedicine, vol. 118, no. 1, pp. 23-33, 2014.
[14] H. Mirinejad, A. E. Gaweda, M. E. Brier, J. M. Zurada and T. Inanc ,
"Individualized drug dosing using RBF-Galerkin method: Case of anemia
management in chronic kidney disease," Computer Methods and Programs
Biomedicine, vol. 148, pp. 45-53, 2017.
[15] S. Tøffner-Clausen, J. M. Grimble and A. M. Johnson, "System Identifica-
tion and Robust Control", Springer, 1996.
[16] R. Bellazzi, "Drug Delivery Optimization through Bayesian Networks: An
Application to Erythropoietin Therapy in Uremic Anemia," Computers
and Biomedical Research, vol. 26, no. 3, pp. 274-293, 1996.
[17] S. M. Öztaner, T. Taşkaya Temizel, S. R. Erdem and M. Özer, "A
Bayesian Estimation Framework for Pharmacogenomics Driven Warfarin
Dosing: A Comparative Study," in IEEE Journal of Biomedical and Health
Informatics, vol. 19, no. 5, pp. 1724-1733, Sept. 2015.
[18] M. Moscoso-Vasquez, P. Colmegna, N. Rosales, F. Garelli and R. Sanchez-
Pena, "Control-Oriented Model With Intra-Patient Variations for an Arti-
ficial Pancreas," in IEEE Journal of Biomedical and Health Informatics,
vol. 24, no. 9, pp. 2681-2689, Sept. 2020.
JACEK M. ZURADA (LF’15) received the Ph.D.
degree in electrical engineering from Gdansk Uni-
versity of Technology, Poland, in 1975. He is a
professor of electrical and computer engineering at
the University of Louisville, USA. He authored or
coauthored several books and over 380 papers in
computational intelligence, neural networks, ma-
chine learning, logic rule extraction, and bioin-
formatics, and delivered more than 100 presenta-
tions throughout the world. Prof. Zurada has been
awarded numerous distinctions, including the 2013 Joe Desch Innovation
Award, the 2015 Distinguished Service Award, and five honorary professor-
ships. In 2014, he served as the IEEE V-President and Technical Activities
(TAB Chair). He also chaired the IEEE TAB Periodicals Committee, and
the TAB Periodicals Review and Advisory Committee. He was the Editor-
in-Chief of the IEEE Transactions on Neural Networks from 1997 to 2003
and an Associate Editor of the IEEE Transactions on Circuits and Systems,
Neural Networks, the Proceedings of the IEEE, Neurocomputing, and sev-
eral other journals. From 2004 to 2005, he was the President of the IEEE
Computational Intelligence Society. He has been a Board Member of IEEE,
IEEE CIS, and IJCNN. He has been elected as a Member of Polish Academy
of Sciences since 2005.

[19] A. E. Gaweda, A. A. Jacobs, M. E. Brier and J. M. Zurada, "Pharmaco-

dynamic population analysis in chronic renal failure using artificial neural
networks—a comparative study," Neural Networks, vol. 16, no. 5-6, pp. TAMER INANC (SM’16) received the B.S. de-
841-845, 2003. gree in electrical engineering from Dokuz Eylul
University, Turkey in 1991. He received the M.S.
[20] J. D. Martin Guerrero, E. S. Olivas, G. C. Valls, A. J. Serrano López, J. J. and Ph.D. degrees in electrical engineering from
Pérez Ruixo and N. V. Jiménez Torres, "Use of neural networks for dosage Penn State University in 1994 and 2002, respec-
individualisation of erythropoietin in patients with secondary anemia to tively. He was a Postdoctoral Scholar at Control
chronic renal failure," Computer in Biology and Medicine, vol. 33, no. 4, and Dynamical Systems Department at Caltech,
pp. 361-373, 2003. USA, between 2002 and 2004. He is currently an
Associate Professor at Department of Electrical
[21] A. E. Gaweda, "Improving management of Anemia in End Stage Renal and Computer Engineering at the University of
Disease using Reinforcement Learning," in 2009 International Joint Con- Louisville, USA. His current research interests include Control Systems
ference on Neural Networks, Atlanta, GA, USA, 2009, pp. 953-958. and System Identification, Control Systems Applications to Biomedical
Problems, Personalized Optimal Drug Dosing, 3D Fingerprint Identification
[22] M. K. Muezzinoglu , "Approximate dynamic programming for anemia
and Biometrics, and Autonomous Robotics.
management," Ph.D thesis, University of Louisville , Louisville, USA
2006, [Online]. Available: https://ir.library.louisville.edu/etd/1024/

[23] A. A. Mangoni and S. H. D. Jackson, “Age-related changes in pharmacoki-

netics and pharmacodynamics: basic principles and practical applications,”
British Journal of Clinical Pharmacology, vol. 57, no. 1, pp. 6–14, 2003.

[24] A. Hammerlein, H. Derendorf, and D. T. Lowenthal, “Pharmacokinetic

and Pharmacodynamic Changes in the Elderly,” Clinical Pharmacokinet-
ics, vol. 35, no. 1, pp. 49–64, Jul. 1998.

[25] J. Wallin, “Dose adaptation based on pharmacomet-

ric models,” Ph.D thesis, Acta Universitatis, Up-
saliensis, 2009. [Online]. Available:http://www.diva-
portal.org/smash/get/diva2:210545/FULLTEXT02

AFFAN AFFAN (Student Member, IEEE) is a

Ph.D student at Electrical and Computer Engineer-
ing Department, University of Louisville, USA.
He received his BS degree in Electrical Engi-
neering in 2017 from COMSATS University Is-
lamabad (CUI), Pakistan and MS degree in 2019
from Lahore University of Management Sciences
(LUMS), Pakistan. He was summer research stu-
dent in 2018 and 2019 at Technische Universität
Kaiserslautern, Germany.

10 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/