Computers in Biology and Medicine 147 (2022) 105752

Contents lists available at ScienceDirect

Computers in Biology and Medicine

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

Boosted machine learning model for predicting intradialytic hypotension

using serum biomarkers of nutrition
Xiao Yang a, Dong Zhao b, **, Fanhua Yu c, Ali Asghar Heidari d, 1, Yasmeen Bano f,
Alisherjon Ibrohimov e, Yi Liu f, Zhennao Cai g, Huiling Chen g, *, Xumin Chen f, ***
a
School of Computer Science and Technology, Changchun University of Science and Technology, Changchun, Jilin, 130022, China
b
College of Computer Science and Technology, Changchun Normal University, Changchun, Jilin, 130032, China
c
College of Computer Science and Technology, Beihua University, Jilin, 132013, China
d
School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran, Tehran, Iran
e
Department of Urology, The First Affiliated Hospital of Wenzhou Medical University, Wenzhou, Zhejiang, 325000, China
f
Department of Nephrology, The First Affiliated Hospital of Wenzhou Medical University, Wenzhou, Zhejiang, 325000, China
g
College of Computer Science and Artificial Intelligence, Wenzhou University, Wenzhou, Zhejiang, 325035, China

A R T I C L E I N F O A B S T R A C T

Keywords: Intradialytic hypotension (IDH) is a serious complication of hemodialysis (HD), with an incidence of more than
Intradialytic hypotension 20%. IDH induces ischemic organ damage and even reduces the ultrafiltration and duration of HD sessions.
Hemodialysis Frequent attacks of IDH are a risk factor for death in HD patients. Malnutrition is common in HD patients and is
Optimization algorithm
also associated with mortality. Although the link between IDH episodes and malnutrition has been observed in
Feature selection
Kernel extreme learning machine
practice, it has not been supported by the data. To study the relationship, we propose a promising hybrid model
KELM called BSCWJAYA_KELM, which is a wrapper feature selection method based on a variant of the JAYA optimi­
ELM zation algorithm (SCWJAYA) and Kernel extreme learning machine (KELM). In this paper, we verify the opti­
Machine learning mization capability of the SCWJAYA algorithm in the model by comparing experiments with some state-of-the-
IEEE CEC art methods for IEEE CEC2014, IEEE CEC2017, and IEEE CEC2019 benchmark functions. The prediction accu­
racy of BSCWJAYA_KELM is validated by the public datasets and the HD dataset. In the experiments on the HD
dataset, 1940 HD sessions of 178 HD patients are analyzed by the developed BSCWJAYA_KELM model. The key
indicators selected from vast amounts of data are serum uric acid, dialysis vintage, age, diastolic pressure, and
albumin. The BSCWJAYA_KELM method is a stable and excellent prediction model that can achieve a more
accurate prediction of IDH.

1. Introduction basis [2]. With an end goal to accomplish some level of euvolemia, ul­
trafiltration objectives regularly include expulsion of what could be
Hemodialysis (HD) emerged as a procedure exhibiting the capacity compared with an entire plasma volume. Maintenance of adequate
to somewhat supplant renal function and hence was practiced as one of end-organ perfusion in this setting is dependent on the organization of a
the most widely adopted treatment modalities for patients with end- variety of complex compensatory components. Unfortunately, auxiliary
stage renal disease (ESRD) [1]. Although HD saves tens of thousands to a horde of patient-related and dialysis-related elements, this remu­
of lives, it is a partial and an intermittent therapy; thus, the complica­ neration regularly misses the mark and results in intradialytic hypo­
tions of HD cannot be ignored. People with maintenance HD face an tension (IDH).
extraordinary hemodynamic challenge, commonly on a thrice-weekly IDH is associated with impaired end-organ perfusion and decreased

* Corresponding author.
** Corresponding author.
*** Corresponding author.
E-mail addresses: yangxiao9799@163.com (X. Yang), zd-hy@163.com (D. Zhao), yufanhua@163.com (F. Yu), as_heidari@ut.ac.ir (A.A. Heidari), dissapear12@
yahoo.com (Y. Bano), A.i.s-96@mail.ru (A. Ibrohimov), liuyi651031@163.com (Y. Liu), cznao@wzu.edu.cn (Z. Cai), chenhuiling.jlu@gmail.com (H. Chen), cxm@
wzhospital.cn (X. Chen).
1
https://aliasgharheidari.com

https://doi.org/10.1016/j.compbiomed.2022.105752
Received 5 March 2022; Received in revised form 13 June 2022; Accepted 14 June 2022
Available online 24 June 2022
0010-4825/© 2022 Elsevier Ltd. All rights reserved.
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 1. Flowchart of SCWJAYA

regional oxygen saturation [3] during HD, resulting in myocardial observational study, patients with IDH had a 1.57-fold higher mortality
infarction, thrombosis of arteriovenous fistulas, ischemic stroke, and than those without IDH over a 5-year period [11].
rapid loss of residual kidney function. Stefansson et al. performed a IDH is a common complication during HD and is associated with a
retrospective study examining 39,497 hemodialysis records and showed higher incidence of organ damage and increased mortality. Certain
that patients with IDH had a 1.2-fold risk of myocardial infarction [4]. methods are available to prevent episodes of IDH during HD, including
Hekmat et al. showed that regional wall movement anomalies preceded higher dialysate sodium [12] or calcium concentration [13], cooling
a reduction in the left ventricular ejection fraction in the IDH group [5]. dialysate [14], decreasing ultrafiltration rates, and adjustment of the
A pilot study of 58 HD patients concluded an association of an approx­ patient’s routine hemodialysis schedule to frequent treatments and
imately 3% increment in cerebral ischemic events with every 10 mmHg longer sessions per week [14]. Although these methods can quickly raise
drop in mean arterial pressure (MAP) from baseline. These events were blood pressure (B_P) during IDH episodes, some adverse events still
correlated with decreased executive cognitive function [6]. In a ran­ exist. A previously observed association between higher all-cause mor­
domized controlled trial, the general pace of thrombosis of local arte­ tality and routine use of sodium profiling to limit or prevent IDH [15].
riovenous fistulas for most elevated quartiles of intradialytic Conversely, lower dialysate sodium is a global trend with the goal of
hypotension was approximately two times that of the lowest quartile, reducing fluid overload and a proven ability to effectively reduce
independent of predialysis systolic blood pressure (SBP) and other interdialytic weight gain (IDWG) without changing the left ventricular
covariates [7]. Jansen et al. performed a prospective study to investigate mass index [16]. There are controversies over dialysate calcium con­
residual renal function, which impacts the morbidity, mortality, and centrations. Higher calcium dialysate can prevent IDH without an
quality of life of dialysis patients. IDH was independently associated increased risk of vascular calcification [17]; however, it is inadequate
with an approximately 1 ml/min/1.73 m2 lower mean urea and creati­ for patients with hypercalcemia. In addition, calcium overload carries
nine clearance at 3 months, and the risk increased over time [8]. Other some potential risks, such as vascular and valvular calcification and
organs damaged were the mesentery [9] and limbs [10]. In an calcific uremic arteriolopathy [18]. There are no differences in all-cause

2
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 2. Flowchart of BSCWJAYA_KELM.

mortality or hospitalization rates between 2.5 mEq/L and lower levels

[19]. Cooling dialysate is a routine method to raise B_P during IDH
Table 2
episodes; however, low dialysate temperatures (35–35.5 ◦ C) are asso­
IEEE CEC2017 benchmark functions.
ciated with patient discomfort. Especially in those with gout, low dial­
ysate temperatures may promote UA deposition in the joints. Rates of No. Name Optimum

death, hospitalization, and missed hemodialysis treatments have shown F16 Shifted and Rotated Zakharov Function 300
no differences between 36 ◦ C and 37 ◦ C dialysate [20]. Decreased ul­ F17 Shifted and Rotated Rastrigin’s Function 500
F18 Shifted and Rotated Lunacek Bi_Rastrigin Function 700
trafiltration rates in a regular 4-h HD treatment are related to volume
F19 Shifted and Rotated Non-Continuous Rastrigin’s Function 800
overload and risk of heart failure. Physicians recommend longer hours F20 Shifted and Rotated Schwefel’s Function 1000
and/or more frequent treatments, but the cost increases the economic F21 Hybrid Function 1 (N = 3) 1100
burden. F22 Hybrid Function 3 (N = 3) 1300
Thus, predicting episodes of IDH is the key to a proper prescription of F23 Hybrid Function 5 (N = 4) 1500
F24 Hybrid Function 6 (N = 5) 1700
ultrafiltration to improve IDH-related adverse events. A few studies have F25 Composition Function 1 (N = 3) 2100
F26 Composition Function 2 (N = 3) 2200
F27 Composition Function 3 (N = 4) 2300
F28 Composition Function 4 (N = 4) 2400
F29 Composition Function 6 (N = 5) 2600
Table 1 F30 Composition Function 9 (N = 3) 2900
IEEE CEC2014 benchmark functions. Search Range: [− 100, 100]D
No. Name Optimum

F1 Shifted and Rotated Ackley’s Function 500

F2 Shifted and Rotated Weierstrass Function 600
F3 Shifted Rotated Rastrigin’s Function 900
F4 Shifted and Rotated Schwefel’s Function 1100
F5 Shifted and Rotated Katsuura Function 1200 Table 3
F6 Shifted and Rotated HappyCat Function 1300 IEEE CEC2019 benchmark functions.
F7 Shifted and Rotated Expandel Scaffer’s F6 Function 1600
No. Name Dimension Range Optimum
F8 Hybrid Function 2 (N = 3) 1800
F9 Hybrid Function 4 (N = 4) 2000 F31 Inverse Hilbert Matrix Problem 16 [-16834, 1
F10 Composition Function 1 (N = 5) 2300 16834]
F11 Composition Function 2 (N = 3) 2400 F32 Lennard–Jones Minimum 18 [-4, 4] 1
F12 Composition Function 3 (N = 3) 2500 Energy Cluster
F13 Composition Function 5 (N = 5) 2700 F33 Expanded Schaffer’s F6 10 [-100, 100] 1
F14 Composition Function 6 (N = 5) 2800 Function
F15 Composition Function 7 (N = 3) 2900 F34 Happy Cat Function 10 [-100, 100] 1
Search Range: [− 100, 100]D F35 Ackley Function 10 [-100, 100] 1

3
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 4 The rest of the paper is organized as follows. Section 2 summarizes

Public parameter settings. the evaluation of relevant studies. The proposed wrapper FS method
Name Value based on the metaheuristic optimization algorithm is detailed in Section
3. Section 4 shows the SCWJAYA algorithm’s capacity to solve global
Population size (N) 30
Individual dimension (D) 30 optimization tasks. Section 5 describes the feature selection experiments
Maximum number of function evaluations (MaxFEs) 300,000 using the BSCWJAYA_KELM method. The experimental findings are
Number of independent runs (Folds) 30 discussed in Section 6. The present and future works are summarized in
Section 7.
been conducted to predict episodes of IDH. Huang et al. used 5 models, a
2. Related works
linear regression model, least absolute shrinkage and selection operator
(LASSO), random forest, extreme gradient boosting, and support vector
Feature selection is the process of selecting a subset of relevant
regression model, to predict IDH episodes, and they achieved good ac­
features from a given set of features. In this section, we discuss previous
curacy [21]. However, these studies mainly focused on dialysis param­
work on wrapper FS. The purpose of FS is to reduce the difficulty of the
eters and demographic characteristics, while there are many risk factors
learning task by removing irrelevant and redundant features, saving
for IDH.
computational and storage costs, and reducing the risk of overfitting.
Hypoalbuminemia has been reported to be a risk factor for IDH [22]
Lyu et al. [31] presented a novel filter FS method based on the Maximal
and a criterion for diagnosing protein-energy wasting (PEW) [23]. Al­
Information Coefficient and Gram–Schmidt Orthogonalization
bumin infusion dialysis is effective in preventing IDH episodes [24].
(OMICFS). OMICFS demonstrated superior performance in
However, to date, no further research has been conducted to investigate
high-dimensional biomedical data mining. Xiong et al. [32] developed a
the correlation between albumin and IDH. PEW is a state of malnutri­
novel informative gene selection method based on cost-sensitive fast
tion, and nutritional status is exactly affected in HD patients [23]. To
correlation-based filter feature selection (CS–FCBF). CS-FCBF showed a
diagnose PEW, biochemical data, body composition, and dietary intake
better classification performance for tumors with fewer selected genes.
should be evaluated. Among these indices, serum albumin and choles­
Li et al. [33] proposed a new hybrid FS method using an enhanced
terol levels are serum biomarkers that are objective and do not need to
discrete artificial bee colony (ABC) optimizer. This wrapper FS method
be observed over months [23]. Beberashvili et al. confirmed that serum
ensured the classification accuracy of the Parkinson’s disease dataset.
uric acid (UA) levels parallel the nutritional status in HD patients [25].
Hu et al. [34] proposed a prediction framework based on an improved
UA is a new biomarker of nutrition. Thus, in our work, we developed a
binary Harris hawk optimization algorithm combined with the kernel
new prospective model to predict IDH utilizing serum biomarkers of
extreme learning machine. The model proposed by this method provides
nutrition, B_P, demographic characteristics, and parameters during
an effective strategy for the accurate early assessment of COVID-19 and
ultrafiltration.
the differentiation of disease severity. Kamkar et al. [35] investigated
In this paper, we propose a prediction model based on a meta­
the stability behavior of Tree-Lasso and validated the stability of the
heuristic optimization algorithm and a KELM called BSCWJAYA_KELM.
Tree-Lasso-based FS method using real datasets of cancer and acute
Here, SCWJAYA is used as the most important part of the model to
myocardial infarction. Niwas et al. [36] presented FS based on the
search for the optimal feature subsets. SCWJAYA is based on the JAYA
minimum redundancy maximum relevance and Laplacian score for the
algorithm introducing the sine chaos initialization strategy, the criss­
classification and diagnosis of angle-closure glaucoma. The effectiveness
cross search, and the wormhole strategy. SCWJAYA was tested with
of the redundant features of the Laplacian Score method was validated
well-known algorithms for 35 global optimization problems from IEEE
in the diagnosis of angle-closure glaucoma compared to the Minimum
CEC2014 [26], IEEE CEC2017 [27], and IEEE CEC2019 [28]. The
Redundancy Maximum Relevance method.
optimization performance of SCWJAYA was validated based on exper­
Common FS methods include filter, wrapper, and embedding. The
imental results such as the mean (Avg), standard deviation (Std),
filter method evaluates the predictive power of the selected features
convergence curve, Wilcoxon signed-rank test (WSRT) [29], and Freid­
with the help of some heuristic criteria based on information statistics.
man test (FT) [30]. In addition, we used public datasets and the HD
The selected subset of features varies according to the evaluation
dataset to evaluate the performance of the feature selection (FS) method
criteria. However, the filter method can quickly eliminate noisy fea­
in the five aspects of accuracy, specificity, precision, Mathews correla­
tures. It is computationally efficient and generalized. However, the
tion coefficient (MCC), and F-measure. In summary, the primary in­
evaluation criteria of the methods are independent of the particular
novations and contributions of this study are as follows:
learning method, and the classification accuracy of the selected feature
subset is often lower than that of the wrapper solver. The embedding FS
• A novel sine initialization, crisscross search, and wormhole search of
interconnects the FS process with the learner training process. Although
the JAYA algorithm is proposed, called SCWJAYA;
it is computationally efficient, it relies on a specific learning algorithm
•Based on SCWJAYA and the kernel extreme learning machine, an
and may suffer from overfitting. The wrapper method is dependent on
effective IDH prediction model named BSCWJAYA KELM is
the learning algorithm, and the performance of the learner is used as the
presented;
evaluation criterion for the feature subset. The performance of feature
•An IDH prediction model based on serum biomarkers of nutrition
subset classification is usually better and more efficient. Based on the
and HD parameters is developed for the first time;
above comparison, we have chosen the wrapper method to solve the
•SCWJAYA’s convergence accuracy, convergence speed and ability
problem in this paper. Although the wrapper method has higher
to jump out of local optima are all improved for 35 optimization
computational efficiency and accuracy, the computational cost of clas­
tasks;
sifier training is greater. In recent years, the wrapper method has used
•BSCWJAYA_KELM can effectively identify the most critical features
the metaheuristic optimization algorithm to find the optimal feature
on both the public datasets and the HD dataset.
subset, which makes up for the shortcomings of the wrapper method.

4
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 5
Avg and Std results of stability analysis.
D Methods F1 F2 F3

Avg Std Avg Std Avg Std

10 SCWJAYA 5.160Eþ02 8.137E+00 6.000Eþ02 5.224E-02 9.045Eþ02 1.449Eþ00

JAYA 5.202E+02 3.251E-01 6.030E+02 1.027E+00 9.312E+02 4.225E+00
30 SCWJAYA 5.201Eþ02 2.332E-02 6.055Eþ02 3.111Eþ00 9.609Eþ02 1.205Eþ01
JAYA 5.209E+02 5.282E-02 6.303E+02 3.700E+00 1.137E+03 1.297E+01
50 SCWJAYA 5.202Eþ02 5.441E-02 6.169Eþ02 5.753E+00 1.045Eþ03 2.701E+01
JAYA 5.211E+02 4.251E-02 6.588E+02 5.149Eþ00 1.403E+03 2.430Eþ01
100 SCWJAYA 5.206Eþ02 8.436E-02 6.680Eþ02 8.848Eþ00 1.418Eþ03 7.394E+01
JAYA 5.213E+02 2.864E-02 7.345E+02 9.601E+00 2.221E+03 5.074Eþ01
D Methods F4 F5 F6
Avg Std Avg Std Avg Std

10 SCWJAYA 1.155Eþ03 5.532Eþ01 1.200Eþ03 3.038E-02 1.300Eþ03 2.110E-02

JAYA 2.216E+03 1.844E+02 1.201E+03 1.100E-01 1.300Eþ03 4.643E-02
30 SCWJAYA 3.093Eþ03 4.041E+02 1.200Eþ03 6.894E-02 1.300Eþ03 5.543E-02
JAYA 7.929E+03 2.160Eþ02 1.202E+03 2.619E-01 1.301E+03 3.863E-01
50 SCWJAYA 5.525Eþ03 6.083E+02 1.200Eþ03 1.110E-01 1.300Eþ03 8.118E-02
JAYA 1.451E+04 3.435Eþ02 1.203E+03 3.339E-01 1.303E+03 3.441E-01
100 SCWJAYA 1.298Eþ04 1.097E+03 1.201Eþ03 2.042E-01 1.301Eþ03 8.261E-02
JAYA 3.176E+04 5.551Eþ02 1.204E+03 2.704E-01 1.305E+03 2.527E-01
D Methods F7 F8 F9
Avg Std Avg Std Avg Std

10 SCWJAYA 1.601Eþ03 4.678E-01 3.311Eþ03 1.293Eþ03 2.004Eþ03 1.594Eþ00

JAYA 1.603E+03 1.998E-01 4.416E+03 2.539E+03 2.216E+03 9.864E+01
30 SCWJAYA 1.610Eþ03 5.529E-01 2.516Eþ03 8.522Eþ02 2.085Eþ03 1.991Eþ01
JAYA 1.613E+03 1.864E-01 3.649E+07 2.081E+07 5.055E+03 1.181E+03
50 SCWJAYA 1.619Eþ03 3.646E-01 3.535Eþ03 1.184Eþ03 2.322Eþ03 6.889Eþ01
JAYA 1.622E+03 2.071E-01 1.903E+08 6.790E+07 2.217E+04 7.812E+03
100 SCWJAYA 1.642Eþ03 6.292E-01 3.805Eþ03 2.367Eþ03 5.585Eþ03 1.165Eþ03
JAYA 1.647E+03 2.553E-01 1.855E+09 3.501E+08 1.454E+05 3.558E+04
D Methods F10 F11 F12
Avg Std Avg Std Avg Std

10 SCWJAYA 2.500Eþ03 0.000Eþ00 2.510Eþ03 2.672Eþ00 2.615Eþ03 7.373Eþ00

JAYA 2.629E+03 6.604E-03 2.538E+03 5.353E+00 2.641E+03 1.252E+01
30 SCWJAYA 2.500Eþ03 7.948E-06 2.600Eþ03 3.544E-02 2.700Eþ03 1.689E-13
JAYA 2.639E+03 4.471E+00 2.624E+03 2.335E+01 2.718E+03 2.626E+00
50 SCWJAYA 2.500Eþ03 9.591E-13 2.600Eþ03 7.625E-02 2.700Eþ03 4.625E-13
JAYA 2.728E+03 1.307E+01 2.725E+03 9.251E+00 2.750E+03 8.990E+00
100 SCWJAYA 2.500Eþ03 1.167E-12 2.601Eþ03 3.027E-01 2.700Eþ03 1.568E-12
JAYA 3.094E+03 5.353E+01 3.001E+03 1.952E+01 2.889E+03 2.328E+01
D Methods F13 F14 F15
Avg Std Avg Std Avg Std

10 SCWJAYA 2.774Eþ03 9.734Eþ01 2.990Eþ03 3.037Eþ01 3.179Eþ03 4.657Eþ01

JAYA 2.851E+03 1.935E+02 3.305E+03 6.325E+01 3.875E+03 3.848E+02
30 SCWJAYA 2.900Eþ03 5.747E-02 3.000Eþ03 5.234E-04 3.560Eþ03 6.708Eþ02
JAYA 3.435E+03 2.218E+02 4.805E+03 4.890E+02 5.990E+06 3.091E+06
50 SCWJAYA 2.900Eþ03 2.385E-02 3.000Eþ03 1.297E-01 3.100Eþ03 1.153Eþ00
JAYA 4.554E+03 9.729E+01 8.521E+03 1.614E+03 1.676E+08 4.470E+07
100 SCWJAYA 2.901Eþ03 2.790Eþ00 3.002Eþ03 5.424Eþ00 3.100Eþ03 2.020E-03
JAYA 6.757E+03 1.777E+02 2.569E+04 1.527E+03 8.744E+08 1.390E+08
D Methods F16 F17 F18
Avg Std Avg Std Avg Std

10 SCWJAYA 3.000Eþ02 3.950E-14 5.047Eþ02 1.375Eþ00 7.157Eþ02 2.202Eþ00

JAYA 8.250E+02 1.565E+02 5.304E+02 4.047E+00 7.435E+02 7.238E+00
30 SCWJAYA 3.048Eþ02 1.871Eþ00 5.540Eþ02 1.107Eþ01 7.934Eþ02 1.379Eþ01
JAYA 4.147E+04 7.284E+03 7.360E+02 1.391E+01 1.029E+03 2.555E+01
50 SCWJAYA 4.910Eþ03 1.911Eþ03 6.516Eþ02 2.416Eþ01 9.146Eþ02 2.578Eþ01
JAYA 1.181E+05 1.354E+04 9.929E+02 2.459E+01 1.457E+03 5.866E+01
100 SCWJAYA 1.573Eþ05 2.003Eþ04 9.790Eþ02 5.183E+01 1.353Eþ03 7.825Eþ01
JAYA 4.002E+05 4.194E+04 1.756E+03 5.010Eþ01 3.144E+03 1.799E+02
D Methods F19 F20 F21
Avg Std Avg Std Avg Std

10 SCWJAYA 8.042Eþ02 1.453Eþ00 1.099Eþ03 1.066Eþ02 1.102Eþ03 9.622E-01

JAYA 8.295E+02 4.636E+00 2.079E+03 1.976E+02 1.123E+03 4.969E+00
30 SCWJAYA 8.531Eþ02 1.014Eþ01 3.355Eþ03 3.899E+02 1.136Eþ03 2.262Eþ01
JAYA 1.039E+03 1.804E+01 8.026E+03 2.553Eþ02 1.937E+03 1.408E+02
50 SCWJAYA 9.383Eþ02 2.271Eþ01 5.275Eþ03 6.994E+02 1.186Eþ03 2.254Eþ01
JAYA 1.338E+03 2.581E+01 1.430E+04 3.165Eþ02 3.416E+03 5.631E+02
100 SCWJAYA 1.216Eþ03 3.522Eþ01 1.354Eþ04 1.272Eþ03 2.020Eþ03 1.038Eþ02
JAYA 2.123E+03 6.330E+01 3.133E+04 6.335E+02 6.751E+04 1.453E+04
(continued on next page)

5
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 5 (continued )
D Methods F1 F2 F3

Avg Std Avg Std Avg Std

D Methods F22 F23 F24

Avg Std Avg Std Avg Std

10 SCWJAYA 2.925Eþ03 1.992Eþ03 1.503Eþ03 1.614Eþ00 1.702Eþ03 8.062E-01

JAYA 5.011E+03 3.547E+03 1.626E+03 3.925E+01 1.756E+03 8.423E+00
30 SCWJAYA 9.741Eþ03 9.566Eþ03 2.114Eþ03 5.153Eþ02 1.766Eþ03 3.571Eþ01
JAYA 5.000E+06 7.900E+06 5.102E+06 2.893E+06 2.306E+03 1.104E+02
50 SCWJAYA 4.997Eþ03 5.181Eþ03 1.105Eþ04 5.594Eþ03 2.324Eþ03 1.998E+02
JAYA 7.090E+08 1.680E+08 1.263E+08 5.081E+07 3.965E+03 1.568Eþ02
100 SCWJAYA 5.108Eþ03 3.407Eþ03 3.213Eþ03 1.473Eþ03 3.926Eþ03 3.569Eþ02
JAYA 3.287E+09 6.127E+08 1.236E+09 2.845E+08 1.081E+04 9.228E+02
D Methods F25 F26 F27
Avg Std Avg Std Avg Std

10 SCWJAYA 2.261Eþ03 5.411Eþ01 2.295Eþ03 2.017E+01 2.607Eþ03 1.544Eþ00

JAYA 2.272E+03 6.143E+01 2.313E+03 1.073Eþ00 2.645E+03 5.920E+00
30 SCWJAYA 2.349Eþ03 8.612Eþ00 2.300Eþ03 3.776E-13 2.703Eþ03 1.022Eþ01
JAYA 2.527E+03 1.393E+01 2.777E+03 6.772E+01 2.971E+03 3.584E+01
50 SCWJAYA 2.419Eþ03 1.896Eþ01 3.767Eþ03 2.322Eþ03 2.853Eþ03 2.422Eþ01
JAYA 2.766E+03 2.703E+01 1.443E+04 3.814E+03 3.496E+03 4.922E+01
100 SCWJAYA 2.678Eþ03 4.496Eþ01 1.441Eþ04 6.297E+03 3.122Eþ03 4.526Eþ01
JAYA 3.631E+03 6.681E+01 3.385E+04 6.245Eþ02 4.708E+03 9.047E+01
D Methods F28 F29 F30
Avg Std Avg Std Avg Std

10 SCWJAYA 2.642Eþ03 1.175Eþ02 2.868Eþ03 8.712E+01 3.140Eþ03 5.943Eþ00

JAYA 2.652E+03 1.239E+02 2.942E+03 8.092Eþ00 3.172E+03 1.126E+01
30 SCWJAYA 2.869Eþ03 8.745Eþ00 3.200Eþ03 5.476Eþ02 3.381Eþ03 4.158Eþ01
JAYA 3.132E+03 2.365E+01 6.497E+03 9.888E+02 4.557E+03 1.301E+02
50 SCWJAYA 3.019Eþ03 1.820Eþ01 3.116Eþ03 6.612Eþ02 3.592Eþ03 1.469Eþ02
JAYA 3.593E+03 5.190E+01 1.123E+04 1.166E+03 6.000E+03 3.522E+02
100 SCWJAYA 3.620Eþ03 4.050Eþ01 9.503Eþ03 4.769Eþ02 5.687Eþ03 4.394Eþ02
JAYA 6.153E+03 3.013E+02 3.144E+04 1.695E+03 1.324E+04 7.333E+02

Table 6
WSRT result of the stability analysis.
No. SCWJAYA vs. JAYA

D = 10 D = 30 D = 50 D = 100

P-value S P-value S P-value S P-value S

F1 7.7122E-04 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +

F2 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F3 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F4 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F5 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F6 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F7 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F8 2.8486E-02 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F9 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F10 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F11 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F12 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F13 6.0350E-03 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F14 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F15 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F16 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F17 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F18 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F19 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F20 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F21 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F22 1.4839E-03 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F23 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F24 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F25 1.1561E-01 = 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F26 1.7344E-06 + 1.7344E-06 + 2.1266E-06 + 1.7344E-06 +
F27 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F28 2.2888E-01 = 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F29 5.2165E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F30 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
+/− / = 28/0/2 30/0/0 30/0/0 30/0/0

6
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 3. WSRT and FT mean ranks of the stability analysis.

Metaheuristic optimization algorithms have a simple structure and [68] proposed a JAYA algorithm (JAYA_LF) based on the Levy flight
high optimization power and can be used to solve various optimization strategy. Zhang et al. [69] improved the local and JAYA algorithm
problems, e.g., plant disease recognition [37], medical diagnosis [38, search methods and proposed a new JAYA method (EJAYA). Yu et al.
39], economic emission dispatch problem [40], parameter tuning for [70] introduced the method of adaptive weights to adjust the tendency
machine learning models [41–43], expensive optimization problems of the algorithm to converge to the optimal solution and avoid the worst
[44,45], big data optimization problems [46], combination optimization solution in different search phases, which is called IJAYA. Iacca et al.
problems [47], among others. Common metaheuristic optimization al­ [71] improved the performance of the original JAYA algorithm (LJA)
gorithms include ant colony optimization (ACO) [48], particle swarm with the help of the Levy distribution. The abovementioned improved
optimization (PSO) [49], grey wolf optimizer (GWO) [50], bat-inspired methods of JAYA are effective in improving the global optimization
algorithm (BA) [51], JAYA optimization algorithm [52], differential capability. However, the no free lunch theorem states that no one so­
evolution algorithm (DE) [53], Harris hawks optimization algorithm lution can solve all optimization problems [72]. In other words, a
(HHO)2 [54], vortex search (VS) [55], weighted mean of vectors (INFO)3 particular method could be capable of considerable optimization in
[56], slime mould algorithm (SMA)4 [57], colony predation algorithm certain optimization domains but not in others [73]. Based on the study
(CPA) [58], Runge Kutta optimizer (RUN)5 [59], and Hunger Games of the above papers, we can conclude that JAYA is an excellent algo­
Search (HGS)6 [60]. These algorithms have been continuously applied to rithm and has many points for improvement. This paper has improved
different optimization problems since they were proposed. However, the the JAYA algorithm to make the classifier more efficient in obtaining the
complexity of optimization cases continues to rise, and the optimization optimal feature subset.
performance of these algorithms cannot keep up with demand. There­
fore, improvement of these algorithms has become a hot issue for 3. The proposed wrapper feature selection model
research. Guo et al. [61] proposed a self-optimization approach for
L-SHADE (SPS_L_SHADE_EIG). Yong et al. [62] proposed a novel bat 3.1. Basic JAYA
algorithm based on collaborative and dynamic learning of the opposite
population (CDLOBA). Cai et al. [63] proposed an improved GWO al­ JAYA is a metaheuristic optimization algorithm with a simple
gorithm that exploits the powerful exploratory power of the hierarchical structure and flexible optimization capability. In this paper, we are
mechanism (IGWO). Lynn et al. [64] proposed an ensemble of different middle-of-the-road about the novelty of this method. It iterates contin­
particle swarm optimization algorithms called the ensemble particle uously to keep the population close to the best individual and away from
swarm optimizer (EPSO). Liu et al. [65] proposed an ant colony opti­ the worst individual. The iterative formula of JAYA is shown in Eq. (1).
mization with Cauchy and greedy Levy mutations (CLACOR).
xnewi = xi + r1 × (xBest − |xi |) − r2 × (xWorst − |xi |), i = 1, 2, 3, …, N (1)
The JAYA algorithm is a recent metaheuristic algorithm. Wang et al.
[66] introduced the elite opposition learning strategy into the update where N denotes the number of search agents, xnewi denotes the pre­
phase of the algorithm (EO_JAYA), and it improved the diversity of so­ selected position of individual i in the next iteration, and xi denotes the
lutions. Zhang et al. [67] proposed a variant of the JAYA algorithm current position of individual i. r1 and r2 are random numbers between
(CLJAYA) that introduces three learning strategies. Motamarri et al. 0 and 1. xBest and xWorst denote the positions of the best and worst in­
dividuals of all search agents in the current iteration, respectively. In Eq.
(1), r1 × (xBest − |xi |) moves the current individual closer to the best in­
2
https://aliasgharheidari.com/HHO.html dividual, and − r2 × (xWorst − |xi |) moves the current individual away
3
https://aliasgharheidari.com/INFO.html from the worst individual. In addition, the fitness value of the current
4
https://aliasgharheidari.com/SMA.html position is compared with the fitness value of the preselected position by
5
https://aliasgharheidari.com/RUN.html greedy selection, and the individual with the best fitness value is
6
https://aliasgharheidari.com/HGS.html

7
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

retained. The pseudocode of the algorithm is shown in Algorithm 1. ⎧

⎨ LB + (UB − LB)⋅rand, if i = 1
⎪
Algorithm 1. The pseudocode of JAYA →
Xi = ( ̅̅→) (2)
⎪ 4
⎩ sin π⋅Xi− 1 , if i > 1
a1

where a1 is a random number, a1∈(0,4].

3.2. The proposed SCWJAYA 3.2.3. Crisscross search

Crisscross search is divided into horizontal crossover search (HCS)
3.2.1. Motivation and vertical crossover search (VCS). It has been shown to have great
The metaheuristic algorithm focuses on obtaining the global optimal potential in many optimizers, such as butterfly optimization [86], ABC
solution by both global exploration and local exploitation [74–76]. [87], ACO [88], and HHO [89] methods.
Global exploration searches for unknown regions by searching agents in The HCS updates the position between two different search agents.
the search space. Global exploration can increase the diversity of search Therefore, the HCS enables different search agents to exchange infor­
agents in the population to avoid premature convergence of the algo­ mation and learn from each other, which effectively improves the
rithm [77–79]. Local exploitation continues the search near the current exploration ability of search agents and thus the convergence speed of
search agent to find a better location [80]. Local exploitation can the algorithm. We assume that the jth position vector of the parent
improve the optimization accuracy of the algorithm [81,82]. Therefore, search agents xi1 and xi2 performs HCS, which can be expressed as Eqs.
improving the optimization performance of the algorithm also focuses (3) and (4).
on these two points [83–85]. The key point of improvement is how to ( )
balance global exploration and local exploitation. The original JAYA MSi1j = ε1 × xji1 + (1 − ε1 ) × xji2 + c1 × xji1 − xji2 (3)
algorithm mainly guides the current search agent’s search direction
( )
through the best and worst position of the search agent. Although this MSi2j = ε2 × xji2 + (1 − ε2 ) × xji1 + c2 × xji2 − xji1 (4)
approach can speed up the convergence of the algorithm, the inability to
obtain a more accurate solution is not the original intention of the where ε1 and ε2 are random numbers between 0 and 1; c1 and c2 are
optimization algorithm. In addition, the random generation of search j
random numbers between − 1 and 1; and MSi1 and MSi2 are the new
j

agents in the search space is not effective in improving the quality of the positions of xi1 and xi2 , respectively.
population. Based on the above problems, the crisscross search strategy In the late iteration, the search agent may not find a better position in
and wormhole search strategy are introduced to improve the conver­ a dimension, making the algorithm fall into a locally optimal solution.
gence accuracy and the ability to jump out of the local optimum of the The VCS updates the position between different dimensions for each
algorithm. Additionally, we use a sine chaos initialization method to search agent to avoid this phenomenon as much as possible. This process
enhance the quality of the initial population and the algorithm’s allows the algorithm to maintain a certain search capability in the late
convergence speed. The core method used by SCWJAYA is described iteration. xi performs VCS at the jth position, which can be expressed as
below. Eq. (5).

3.2.2. Sine chaos initialization MSij = ε × xji + (1 − ε) × xj1

i (5)
To speed up the convergence of the algorithm, SCWJAYA is imple­
j1
mented by sine chaos initialization. The search agents of the population where ε denotes a random number between 0 and 1, and xi denotes the
under sine chaos initialization are all generated from the previous search j1th dimension of xi .
agent. First, the first search agent in the search space is produced
randomly. The other search agents in the population are then generated 3.2.4. Wormhole search
according to the sine mapping and the previous individual. Eq. (2) il­ The wormhole search updates the population using WEP, TDR, and
lustrates the mathematical model for the sine initialization strategy. the current iteration of the optimal search agent. WEP is an adaptive
parameter that determines whether the current search agent is updated
or not. TDR is a weight parameter that controls the influence of random
search on the current search agent through different search periods. Eqs.

8
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 4. Convergence curve of the stability analysis (D is set to 10).

(6)–(8) illustrate the mathematical model for the wormhole search. 3.3. The implementation of SCWJAYA
⎧ { ̅→
⎨ Xα + TDR × ((UB − LB) × r5 + LB), r4 < 0.5 , WEP < r The SCWJAYA algorithm introduced the sine chaos initialization
→ strategy to improve the quality and convergence speed of the initial
3
Xl (t + 1) = ̅→
X − TDR × ((UB − LB) × r5 + LB), r4 ≥ 0.5
⎩→ α population and presented crisscross search and wormhole search stra­
Xl (t) , WEP ≥ r3 tegies to better balance the algorithm global and local searches. In this
(6) method, the sine chaos initialization, JAYA core update, crisscross
( ) search, and wormhole search strategies are performed in turn. Because
WEP = WEPmin + FEs ×
WEPmax − WEPmin
(7) this paper uses the function evaluation method to verify the method
MaxFEs performance, we can exclude the case of improving the algorithm
optimization performance by updating the policy stacking method in
TDR = 1 −
FEs1/p
(8) one iteration. The detailed implementation of SCWJAYA is shown in
MaxFEs1/p Fig. 1. According to the figure, SCWJAYA is described as follows.
where p controls the local exploitation capability of the algorithm,
Step 1: Initialize the parameters. The population size N, the indi­
which is set to 6. WEP is between WEPmin and WEPmax and is set to [0.2,
vidual dimension D, the upper boundary of the search space UB, the
1] in this paper. r3, r4, and r5 are random numbers between 0 and 1.
lower boundary of the search space LB, the maximum number of

9
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 5. Convergence curve of the stability analysis (D is set to 30).

function evaluations MaxFEs, the number of current function eval­ lected, as shown in many recent studies [90–93]. In subsection 3.2, we
uations FEs = 0, range of WEP, control parameter p. present a version of SCWJAYA for continuous domain numerical opti­
Step 2: Initialize the population X according to the sine chaos mization problems. In this subsection, conversion into a binary version
initialization strategy and calculate the fitness value Fit of X. is required for introduction into the prediction model, which is called
Step 3: Select the best individual xBest and the worst individual xWorst . BSCWJAYA. Each dimension of the search agent represents a feature and
Step 4: Update the population X according to Eq. (1), crisscross obtains an array of N consisting of ‘0′ and ‘1’; ‘0′ indicates that the
search strategy, and wormhole search strategy in turn. feature is not selected, whereas ‘1′ indicates that it is selected. Eqs. 9 and
Step 5: Update the number of function evaluations, FEs = FEs + 3 × 10 illustrate the process of converting a continuous domain to a discrete
N. domain.
Step 6: Check if the loop termination condition is satisfied. If FEs > {
1, sigmoid(Xd (t)) ≥ r
MaxFEs, output the optimal individual xBest . If FEs ≤ MaxFEs, skip to Xd (t + 1) = (9)
0, otherwise
Step 3.
1
sigmoid(x) = (10)
3.4. The proposed SCWJAYA-based feature selection method 1 + e− x/3

3.4.1. Discretization r is a random integer between 0 and 1, and Xd is the binary position of
The result obtained from FS in this paper is a binary classification the search agent.
problem, where each feature has only two cases, selected and unse­

10
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 6. Convergence curve of the stability analysis (D is set to 50).

3.4.2. Classifier ( )− 1
To improve the prediction accuracy of the model, the kernel extreme β = H T HH T +
I
L (12)
learning machine (KELM) is finally selected to perform the classification c
task in the model after a series of experiments. KELM is a more advanced The kernel function is introduced into the ELM, and the kernel matrix
technique that combines an extreme learning machine (ELM) with a is shown in Eqs. 13 and 14.
kernel function. Eq. (11) illustrates the learning goal function F(x) of ( ) ( )
ELM. ΩELM = HH T = h(xi )h xj = K xi , xj (13)

F(x) = h(x) × β = L (11) ( )− 1

I
F(x) = [K(x, x1 ); …; K(x, xn )] + ΩELM L (14)
where x is the input vector, h(x) denotes the output of the hidden node, β C
is the output weight, and L denotes the desired output.
where xi and xj are the trial input vectors, (x1, x2, …, xn) are the given
By converting the network training to a linear system solution, β =
training samples, n is the number of samples, and K is the kernel
H*⋅L, H and H* denote the output of the hidden layer node and its
function.
generalized inverse matrix, respectively. To improve the neural net­
work’s stability, the regularization factor C and the unit matrix I are
3.4.3. Fitness function
included, and the output weights’ least-squares solution is illustrated in
A fitness function is required to analyze feature subsets while
Eq. (12).
implementing the BSCWJAYA_KELM model. As a result, this paper

11
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 7. Convergence curve of the stability analysis (D is set to 100).

Table 7
where Error reflects KELM’s error rate, R is the length of the subset of
Strategy combination comparison parameter setting. features selected, D is the total length of all features, and α and β denote
the error rate and weight coefficient, respectively, of the feature subset.
Name Sine chaos Initialization Crisscross Search Wormhole Search
We place a higher premium on the error rate than on the effect of the
JAYA 0 0 0 chosen feature subset on the classification results. As a result, α is set to
SJAYA 1 0 0
0.99, whereas β is set to 0.01. The fitness function may be used to
CJAYA 0 1 0
WJAYA 0 0 1 determine the classification impact of a combined subset of character­
SCJAYA 1 1 0 istics, and the lower the fitness value, the more effective is the feature
SWJAYA 1 0 1 subset at classification.
CWJAYA 0 1 1
SCWJAYA 1 1 1
3.4.4. The proposed BSCWJAYA_KELM
To find more representative features in the dataset to help improve
employs a well-established approach for assessing classification systems the efficiency of medical diagnosis, in this paper, we utilized
on the basis of error rate and feature subsets. Eq. (15) illustrates the BSCWJAYA_KELM as a FS tool to select the optimal feature subset.
fitness function. BSCWJAYA is used as an optimizer to catch the optimal subset. After
finding the optimal features, KELM model is utilized as a classifier model
R
Fitness = α⋅Error + β⋅ (15) for the classification case. Fig. 2 shows a BSCWJAYA _KELM flowchart,
D

12
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 8
WSRT results of the strategy combination comparison experiment.
No. SJAYA CJAYA WJAYA SCJAYA

P-value S P-value S P-value S P-value S

F1 1.7344E-06 + 1.7344E-06 + 2.5967E-05 + 1.7344E-06 +

F2 1.7344E-06 + 2.3534E-06 + 2.9894E-01 = 1.1973E-03 +
F3 1.7344E-06 + 7.4987E-01 = 3.3269E-02 - 1.6394E-05 +
F4 1.7344E-06 + 2.3534E-06 + 1.4936E-05 + 1.2381E-05 +
F5 1.7344E-06 + 1.7344E-06 + 2.3704E-05 + 1.7344E-06 +
F6 1.7344E-06 + 5.6672E-03 + 1.7518E-02 + 2.2248E-04 +
F7 1.7344E-06 + 3.7243E-05 + 1.7988E-05 + 4.8603E-05 +
F8 1.7344E-06 + 5.7924E-05 + 1.7344E-06 + 5.3197E-03 +
F9 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F10 1.7344E-06 + 1.7127E-06 + 1.7344E-06 + 1.0000E+00 =
F11 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 4.4493E-05 -
F12 1.2207E-04 + 1.7344E-06 + 1.7344E-06 + 1.0000E+00 =
F13 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.3219E-03 +
F14 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 5.7502E-01 =
F15 1.7344E-06 + 1.3601E-05 + 1.7344E-06 + 5.5061E-01 =
F16 1.7344E-06 + 1.7344E-06 + 1.7344E-06 - 1.7344E-06 +
F17 1.7344E-06 + 1.0444E-02 + 5.0383E-01 = 6.3391E-06 +
F18 1.7344E-06 + 5.2165E-06 + 1.2044E-01 = 3.1817E-06 +
F19 1.7344E-06 + 1.6566E-02 + 4.2843E-01 = 1.9729E-05 +
F20 1.7344E-06 + 9.3157E-06 + 1.0357E-03 + 3.5152E-06 +
F21 1.7344E-06 + 4.1955E-04 + 1.7344E-06 + 3.5888E-04 +
F22 1.7344E-06 + 8.7740E-01 = 1.2506E-04 + 8.4508E-01 =
F23 1.7344E-06 + 2.9575E-03 + 3.5152E-06 + 7.1889E-01 =
F24 1.7344E-06 + 1.2506E-04 + 2.3704E-05 + 6.6392E-04 +
F25 1.7344E-06 + 1.9729E-05 + 2.0589E-01 = 5.3070E-05 +
F26 1.7344E-06 + 9.7656E-04 + 1.7344E-06 + 1.0000E+00 =
F27 1.7344E-06 + 3.4053E-05 + 5.7096E-02 = 5.7517E-06 +
F28 1.7344E-06 + 2.8434E-05 + 2.5364E-01 = 3.4053E-05 +
F29 1.9209E-06 + 2.3534E-06 + 9.3157E-06 + 6.9332E-02 =
F30 1.7344E-06 + 5.2165E-06 + 1.7344E-06 + 1.4936E-05 +
F31 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 8.2206E-02 =
F32 1.7344E-06 + 1.2866E-03 + 2.0515E-04 + 3.8723E-02 +
F33 1.7344E-06 + 1.7344E-06 + 1.2506E-04 + 6.3391E-06 +
F34 1.7344E-06 + 2.4147E-03 + 4.9498E-02 + 5.7096E-02 =
F35 3.8822E-06 + 1.1265E-05 + 2.1266E-06 + 5.7516E-02 =
+/− / = 35/0/0 33/0/2 26/2/7 23/1/11

No. SWJAYA CWJAYA JAYA

P-value S P-value S P-value S

F1 4.9916E-03 + 1.0444E-02 + 1.7344E-06 +

F2 1.7088E-03 - 3.1603E-02 + 1.7344E-06 +
F3 7.8647E-02 = 4.8603E-05 - 1.7344E-06 +
F4 8.1878E-05 + 9.5899E-01 = 1.7344E-06 +
F5 1.7988E-05 + 8.7740E-01 = 1.7344E-06 +
F6 1.3601E-05 + 7.6552E-01 = 1.7344E-06 +
F7 7.7122E-04 + 6.8836E-01 = 1.7344E-06 +
F8 6.3391E-06 + 1.5286E-01 = 1.7344E-06 +
F9 1.9209E-06 + 4.9498E-02 - 1.7344E-06 +
F10 1.7344E-06 + 1.7333E-06 + 1.7344E-06 +
F11 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F12 7.8044E-05 + 1.7344E-06 + 1.7344E-06 +
F13 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F14 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F15 1.7344E-06 + 5.2973E-05 + 1.7344E-06 +
F16 1.7344E-06 - 3.4935E-01 = 1.7344E-06 +
F17 1.5286E-01 = 5.4463E-02 = 1.7344E-06 +
F18 8.2206E-02 = 4.7795E-01 = 1.7344E-06 +
F19 7.6552E-01 = 3.6094E-03 - 1.7344E-06 +
F20 9.7110E-05 + 5.4463E-02 = 1.7344E-06 +
F21 1.7344E-06 + 4.0483E-01 = 1.7344E-06 +
F22 1.2506E-04 + 7.0356E-01 = 1.7344E-06 +
F23 4.4493E-05 + 6.8714E-02 = 1.7344E-06 +
F24 6.3391E-06 + 4.7795E-01 = 1.7344E-06 +
F25 3.0010E-02 + 7.4987E-01 = 1.7344E-06 +
F26 1.7344E-06 + 7.8125E-03 + 1.7344E-06 +
F27 1.7791E-01 = 4.0483E-01 = 1.7344E-06 +
F28 1.3591E-01 = 4.0702E-02 + 1.7344E-06 +
F29 2.8486E-02 + 8.4661E-06 + 1.9209E-06 +
F30 2.8786E-06 + 7.3433E-01 = 1.7344E-06 +
F31 6.9838E-06 + 1.7344E-06 + 1.7344E-06 +
F32 7.1570E-04 + 9.4261E-01 = 1.7344E-06 +
F33 1.3975E-02 + 6.5833E-01 = 1.7344E-06 +
(continued on next page)

13
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 8 (continued )
No. SWJAYA CWJAYA JAYA

P-value S P-value S P-value S

F34 2.2551E-03 + 5.4401E-01 = 1.7344E-06 +

F35 2.8308E-04 + 5.0274E-05 + 1.7344E-06 +
+/− / = 27/2/6 13/3/19 35/0/0

Fig. 8. WSRT and FT mean ranks of the strategy combination comparison experiment.

which illustrates the steps of obtaining the onset key factors of IDH. The state-of-the-art (SOTA) methods, and JAYA variant algorithms verify
main steps of the BSCWJAYA _KELM method are described below. SCWJAYA’s global optimization performance.

Step 1: Load the dataset and normalize the collected data to the range
4.1. Numerical optimization problems
[-1,1].
Step 2: The samples are divided using 20-fold cross-validation.
In this section, we verify the performance of SCWJAYA global
Step 3: Initialize the parameters of BSCWJAYA.
exploration, local exploitation, and jumping out of local optimums with
Step 4: Generate binary populations according to the sine chaos
35 benchmark functions selected experimentally from the IEEE
initialization strategy (Subsection 3.2.2) and the discrete method
CEC2014, IEEE CEC2017, and IEEE CEC2019 test components, which
(Subsection 3.4.1).
consist of unimodal, multimodal, hybrid, and composition functions.
Step 5: Calculate the fitness of each set of feature subsets based on the
The unimodal function has only one local optimum, which is the
fitness function (Subsection 3.4.3).
global optimal solution. The unimodal function model can better test the
Step 6: Update the individuals and calculate their fitness using
exploitation capability of SCWJAYA. However, most of the real prob­
BSCWJAYA. Finally, the subset of features with the smallest fitness
lems currently have multiple local optima, so simple multimodal,
value is selected as the optimal solution.
hybrid, and composition functions are used to verify the exploration
Step 7: Determine if the termination condition is met. If yes, run the
capability. The details of the function are shown in Table 1, Table 2 and
next step; otherwise, repeat Step 6 until thestop condition is satisfied.
Table 3.
Step 8: Obtain the best feature set.
Step 9: Utilize the obtained feature set for the initial parameters of
KELM model to find the last classification solutions. 4.2. Experiment setup
Step 10: Based on the classification solutions found in Step 9, find the
classification error accuracy, sensitivity, specificity, and other eval­ All experiments were performed on a computer with an Intel(R) Xeon
uation criteria. (R) CPU E5-2620 v4 @ 2.10 GHz processor and 16 GB RAM using
MATLAB 2018b. To verify its performance, we compared the proposed
4. Experiments for the global optimization problem SCWJAYA with some state-of-the-art methods, including JAYA, DE, TSA
[94], HHO, CDLOBA, IGWO, EPSO, CLACOR, VS, and SPS_L_SHA­
In this section, SCWJAYA tests the performance of 35 global opti­ DE_EIG. In addition, SCWJAYA is compared with other variants of the
mization problems. These problems come from the IEEE CEC2014, IEEE JAYA algorithms, including EO_JAYA, CLJAYA, EJAYA, IJAYA,
CEC2017, and IEEE CEC2019 benchmark function test components. The JAYA_LF, and LJA. The parameters of these comparison methods are
stability experiment, strategy combination comparison experiment, obtained from the relevant references. Notably, all the common pa­
qualitative analysis experiment, comparison experiment with some rameters in this experiment are shown in Table 4.

14
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 9. Convergence curves of the strategy combination comparison experiment.

4.3. Stability analysis experiment Table 6 shows the experimental results of SCWJAYA and JAYA in
terms of WSRT. From Table 6, when the problem dimension is 10, only
To verify that the SCWJAYA proposed in this paper has excellent the p values of the optimization results for F25 and F28 are greater than
optimization ability in different dimensions, this subsection tests the 0.05, which indicates that the variability between the SCWJAYA and
stability of SCWJAYA with JAYA in four different dimensions (dimen­ JAYA optimization results is small. The optimization results of
sion is set to 10, 30, 50, and 100). The dimensions of the benchmark SCWJAYA outperform those of the JAYA algorithm in all other di­
functions of IEEE CEC2019 have been explicitly stated (details are mensions, and these differences are statistically significant. In addition,
shown in Table 3). Therefore, the experiment in this subsection is only the last column of Table 6 shows the number of SCWJAYA and JAYA
validated for the thirty benchmark functions of IEEE CEC2014 and IEEE wins or losses. When the dimension is set to 10, SCWJAYA outperforms
CEC2017. JAYA in twenty-eight functions, and the differences in the other two
Table 5 shows the means and standard deviations of ten independent functions are insignificant. When the dimension is set to 30, 50, and 100,
runs of SCWJAYA and JAYA for thirty benchmark functions in different SCWJAYA outperforms JAYA in thirty benchmark functions.
dimensions. As seen from Table 5, SCWJAYA outperforms JAYA in all Fig. 3 shows the average ranking of the two algorithms in the sta­
dimensions in terms of means and standard deviations, which indicates bility experiments for WSRT and FT. From Fig. 3, it can be seen that
that SCWJAYA is easier to approach with the global optimal solution SCWJAYA is the best among these statistical methods.
than JAYA in these global optimization problems. The stability of Figs. 4–7 show the convergence curves of SCWJAYA and JAYA in
SCWJAYA is better. different dimensions, respectively. From the convergence curves of the

15
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 10. (a) Three-dimensional models of thirty-five benchmark functions. (b) Top view of the search process. (c) Search trajectories in the first dimension. (d) The
average fitness value of all search agents.

four dimensions, we can see that the SCWJAYA curve is below the JAYA strategy.
curve, which indicates that SCWJAYA has better convergence accuracy Table 8 shows the WSRT results of the strategy combination exper­
and convergence speed. iment. From Table 8, it can be seen that the p values for most of the
In summary, SCWJAYA is better than JAYA in the optimization functions are less than 0.05, except those for the CWJAYA method. This
performance tests of different dimensions, which indicates that finding indicates that the comparison of SCWJAYA with six other
SCWJAYA is more stable. methods for more than thirty-three functions provides significantly
different results, most of which support SCWJAYA as the winner. In
4.4. Strategy combination comparison experiment addition, the comparison between SCWJAYA and CWJAYA shows that
the optimization performance of SCWJAYA is stronger than that of
In this subsection, the experiments verify the effect of each of the CWJAYA for thirteen functions, and only three functions are weaker for
three strategies on the algorithm and the effect of the combination of the SCWJAYA than CWJAYA. The optimization results of the two algorithms
three strategies on the algorithm. The algorithm generates a total of for the other nineteen functions are similar.
eight algorithms based on different combinations of the three strategies. Fig. 8 shows the ranking of the two statistical methods for the eight
The experimental conditions are detailed in Table 7. Here, ‘0′ means that combination methods. In the WSRT statistical results, SCWJAYA has an
this strategy is not used, and ‘1′ means that the algorithm uses this average ranking of 1.66 for the thirty-five functions, and it has the best

16
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 11. (a) Balance curve of SCWJAYA. (b) Balance curve of JAYA. (c) Diversity curve of SCWJAYA and JAYA.

17
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 9
Avg and Std results of the comparison experiment with SOTA methods.
Methods F1 F2 F3

Avg Std Avg Std Avg Std

SCWJAYA 5.201E+02 2.968E-02 6.054E+02 3.154E+00 9.604E+02 1.226E+01

JAYA 5.209E+02 6.948E-02 6.293E+02 3.796E+00 1.132E+03 1.406E+01
DE 5.206E+02 4.655E-02 6.195E+02 2.066Eþ00 1.010E+03 1.082E+01
VS 5.200Eþ02 3.458E-05 6.025Eþ02 2.235E+00 9.768E+02 2.366E+01
HHO 5.201E+02 1.306E-01 6.302E+02 2.781E+00 1.085E+03 1.780E+01
CDLOBA 5.209E+02 9.070E-02 6.362E+02 3.459E+00 1.233E+03 6.011E+01
IGWO 5.205E+02 1.548E-01 6.184E+02 3.153E+00 1.015E+03 2.548E+01
EPSO 5.209E+02 5.962E-02 6.078E+02 2.450E+00 1.065E+03 2.081E+01
CLACOR 5.200E+02 4.977E-03 6.135E+02 5.483E+00 1.005E+03 3.924E+01
TSA 5.210E+02 4.558E-02 6.134E+02 6.612E+00 1.107E+03 1.470E+01
SPS_L_SHADE_EIG 5.200Eþ02 6.256E-06 6.078E+02 3.506E+00 9.544Eþ02 9.609Eþ00
Methods F4 F5 F6
Avg Std Avg Std Avg Std

SCWJAYA 2.991Eþ03 3.438E+02 1.200E+03 7.126E-02 1.300E+03 5.103E-02

JAYA 7.949E+03 2.712Eþ02 1.202E+03 2.123E-01 1.301E+03 2.960E-01
DE 5.733E+03 2.826E+02 1.201E+03 1.125E-01 1.300E+03 4.425E-02
VS 3.591E+03 6.362E+02 1.200Eþ03 2.900E-02 1.300Eþ03 4.361E-02
HHO 4.904E+03 6.575E+02 1.201E+03 4.676E-01 1.301E+03 1.241E-01
CDLOBA 5.651E+03 7.201E+02 1.200E+03 1.791E-01 1.300E+03 1.286E-01
IGWO 4.601E+03 5.839E+02 1.201E+03 4.138E-01 1.301E+03 1.127E-01
EPSO 5.919E+03 1.493E+03 1.202E+03 3.311E-01 1.300E+03 6.015E-02
CLACOR 4.039E+03 7.914E+02 1.200E+03 6.305E-02 1.300E+03 7.657E-02
TSA 8.005E+03 2.869E+02 1.203E+03 2.860E-01 1.300E+03 6.215E-02
SPS_L_SHADE_EIG 3.384E+03 5.181E+02 1.200E+03 5.125E-02 1.300E+03 7.023E-02
Methods F7 F8 F9
Avg Std Avg Std Avg Std

SCWJAYA 1.610Eþ03 4.609E-01 2.572E+03 1.116E+03 2.080Eþ03 2.256Eþ01

JAYA 1.613E+03 2.056E-01 3.687E+07 2.125E+07 4.520E+03 7.651E+02
DE 1.612E+03 3.072E-01 1.006E+04 7.927E+03 4.972E+03 1.311E+03
VS 1.610E+03 5.547E-01 6.983E+03 5.578E+03 2.176E+03 6.651E+01
HHO 1.612E+03 3.616E-01 5.815E+04 2.124E+04 6.084E+03 2.417E+03
CDLOBA 1.613E+03 2.807E-01 1.066E+04 7.540E+03 2.496E+04 1.682E+04
IGWO 1.612E+03 5.045E-01 2.707E+04 4.399E+04 3.624E+03 1.948E+03
EPSO 1.612E+03 3.559E-01 3.473E+03 2.419E+03 2.510E+03 1.674E+02
CLACOR 1.611E+03 4.949E-01 1.063E+04 8.287E+03 3.111E+03 9.813E+02
TSA 1.613E+03 2.184E-01 2.129Eþ03 3.225Eþ02 1.126E+04 4.287E+03
SPS_L_SHADE_EIG 1.610E+03 6.710E-01 1.878E+03 3.113E+01 2.048E+03 4.179E+01
Methods F10 F11 F12
Avg Std Avg Std Avg Std

SCWJAYA 2.500Eþ03 1.149E-12 2.600E+03 4.731E-02 2.700Eþ03 1.194E-13

JAYA 2.639E+03 5.332E+00 2.632E+03 2.220E+01 2.718E+03 2.994E+00
DE 2.615E+03 1.388E-12 2.625E+03 2.185E+00 2.708E+03 1.289E+00
VS 2.615E+03 7.995E-04 2.606E+03 1.079E+01 2.704E+03 7.510E-01
HHO 2.500Eþ03 0.000Eþ00 2.600Eþ03 1.437E-04 2.700Eþ03 0.000Eþ00
CDLOBA 2.616E+03 1.446E+00 2.709E+03 5.816E+01 2.726E+03 1.343E+01
IGWO 2.621E+03 3.266E+00 2.600E+03 5.373E-03 2.710E+03 2.135E+00
EPSO 2.615E+03 2.055E-12 2.627E+03 3.526E+00 2.711E+03 4.322E+00
CLACOR 2.614E+03 1.726E-01 2.622E+03 9.553E+00 2.701E+03 2.170E-01
TSA 2.615E+03 6.306E-05 2.626E+03 7.880E-01 2.719E+03 2.824E+00
SPS_L_SHADE_EIG 2.615E+03 1.458E-12 2.626E+03 1.273E+00 2.704E+03 1.850E+00
Methods F13 F14 F15
Avg Std Avg Std Avg Std

SCWJAYA 2.900E+03 7.276E-03 3.000Eþ03 4.039E-04 3.353E+03 5.307E+02

JAYA 3.376E+03 2.212E+02 4.687E+03 5.252E+02 5.954E+06 2.766E+06
DE 3.229E+03 9.919E+01 3.635E+03 2.453E+01 1.532E+05 8.139E+05
VS 3.101E+03 3.822E-01 3.698E+03 5.388E+01 3.908E+05 2.088E+06
HHO 2.900Eþ03 0.000Eþ00 3.000Eþ03 0.000Eþ00 4.253E+03 6.315E+03
CDLOBA 3.772E+03 4.449E+02 5.310E+03 6.821E+02 1.707E+07 1.590E+07
IGWO 3.109E+03 2.563E+00 3.837E+03 1.619E+02 1.238E+06 3.749E+06
EPSO 3.180E+03 8.461E+01 3.691E+03 7.261E+01 4.015E+03 2.872E+02
CLACOR 3.345E+03 1.768E+02 3.253E+03 5.749E+01 3.122Eþ03 3.148Eþ01
TSA 3.149E+03 8.083E+01 3.971E+03 7.897E+01 1.505E+04 1.293E+04
SPS_L_SHADE_EIG 3.122E+03 6.311E+01 3.734E+03 2.063E+02 3.663E+03 3.601E+01
Methods F16 F17 F18
Avg Std Avg Std Avg Std

SCWJAYA 3.060E+02 2.095E+00 5.513E+02 1.060E+01 7.918E+02 1.265E+01

JAYA 4.140E+04 8.621E+03 7.337E+02 1.564E+01 1.024E+03 1.890E+01
DE 2.004E+04 3.521E+03 6.066E+02 9.252Eþ00 8.447E+02 8.576Eþ00
VS 3.000Eþ02 7.685E-14 5.666E+02 1.854E+01 8.068E+02 2.255E+01
HHO 8.379E+02 2.815E+02 7.202E+02 2.856E+01 1.212E+03 7.264E+01
(continued on next page)

18
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 9 (continued )
Methods F1 F2 F3

Avg Std Avg Std Avg Std

CDLOBA 1.087E+03 1.407E+03 8.697E+02 5.616E+01 2.676E+03 2.501E+02

IGWO 1.208E+03 4.845E+02 6.116E+02 2.028E+01 8.887E+02 3.254E+01
EPSO 5.912E+03 1.516E+03 6.389E+02 4.330E+01 9.548E+02 1.497E+01
CLACOR 7.551E+02 8.079E+02 6.153E+02 3.093E+01 8.531E+02 4.191E+01
TSA 4.390E+04 6.287E+03 7.077E+02 1.255E+01 9.448E+02 1.239E+01
SPS_L_SHADE_EIG 3.000Eþ02 9.674E-14 5.512Eþ02 1.306E+01 7.868Eþ02 9.512E+00
Methods F19 F20 F21
Avg Std Avg Std Avg Std

SCWJAYA 8.560Eþ02 9.211E+00 3.164Eþ03 3.556E+02 1.133Eþ03 1.420Eþ01

JAYA 1.034E+03 1.303E+01 8.133E+03 2.340Eþ02 1.945E+03 1.579E+02
DE 9.095E+02 8.485Eþ00 5.822E+03 2.737E+02 1.154E+03 2.713E+01
VS 8.720E+02 2.408E+01 3.660E+03 6.007E+02 1.195E+03 3.807E+01
HHO 9.490E+02 2.462E+01 5.321E+03 5.229E+02 1.246E+03 3.443E+01
CDLOBA 1.113E+03 5.546E+01 5.498E+03 5.239E+02 1.318E+03 6.665E+01
IGWO 8.898E+02 2.324E+01 4.710E+03 7.030E+02 1.258E+03 3.479E+01
EPSO 9.564E+02 2.474E+01 6.548E+03 8.181E+02 1.166E+03 3.808E+01
CLACOR 9.139E+02 3.929E+01 4.238E+03 6.197E+02 1.163E+03 3.655E+01
TSA 1.000E+03 1.882E+01 8.122E+03 3.363E+02 1.244E+03 2.141E+01
SPS_L_SHADE_EIG 8.601E+02 1.156E+01 3.429E+03 4.471E+02 1.158E+03 3.279E+01
Methods F22 F23 F24
Avg Std Avg Std Avg Std

SCWJAYA 1.515E+04 1.249E+04 2.543E+03 2.039E+03 1.764Eþ03 3.467Eþ01

JAYA 6.371E+06 1.087E+07 4.449E+06 2.607E+06 2.317E+03 1.161E+02
DE 3.110E+04 1.801E+04 7.732E+03 4.541E+03 1.839E+03 4.420E+01
VS 7.917E+04 5.023E+04 5.169E+04 2.541E+04 1.846E+03 6.483E+01
HHO 2.847E+05 1.005E+05 5.471E+04 3.563E+04 2.448E+03 2.783E+02
CDLOBA 1.544E+05 7.890E+04 9.121E+04 6.281E+04 2.882E+03 3.416E+02
IGWO 2.425E+05 4.166E+05 5.178E+04 3.156E+04 2.055E+03 1.422E+02
EPSO 1.642E+04 1.414E+04 5.700E+03 5.096E+03 1.828E+03 9.381E+01
CLACOR 2.462E+04 2.449E+04 1.763E+04 1.359E+04 1.954E+03 1.525E+02
TSA 9.658E+03 6.060E+03 2.702E+03 1.602E+03 2.002E+03 1.157E+02
SPS_L_SHADE_EIG 1.428Eþ03 1.445Eþ02 1.558Eþ03 3.073Eþ01 1.885E+03 1.073E+02
Methods F25 F26 F27
Avg Std Avg Std Avg Std

SCWJAYA 2.349Eþ03 8.613Eþ00 2.300Eþ03 3.482E-13 2.705E+03 1.092E+01

JAYA 2.519E+03 1.482E+01 2.774E+03 7.973E+01 2.973E+03 2.237E+01
DE 2.406E+03 9.192E+00 3.581E+03 1.669E+03 2.758E+03 9.586Eþ00
VS 2.370E+03 1.747E+01 2.300E+03 1.134E+00 2.712E+03 1.714E+01
HHO 2.516E+03 7.273E+01 6.619E+03 1.867E+03 3.064E+03 8.001E+01
CDLOBA 2.630E+03 5.163E+01 6.888E+03 1.474E+03 3.185E+03 1.111E+02
IGWO 2.402E+03 2.499E+01 2.311E+03 1.735E+00 2.775E+03 3.320E+01
EPSO 2.431E+03 4.248E+01 2.301E+03 1.818E+00 2.692E+03 1.250E+01
CLACOR 2.407E+03 3.175E+01 5.146E+03 1.383E+03 2.760E+03 3.574E+01
TSA 2.496E+03 1.038E+01 2.300Eþ03 3.452E-06 2.855E+03 1.378E+01
SPS_L_SHADE_EIG 2.362E+03 1.345E+01 2.554E+03 7.913E+02 2.703Eþ03 1.403E+01
Methods F28 F29 F30
Avg Std Avg Std Avg Std

SCWJAYA 2.872Eþ03 1.257E+01 3.351Eþ03 6.512E+02 3.388Eþ03 3.973Eþ01

JAYA 3.128E+03 2.055E+01 6.496E+03 9.905E+02 4.501E+03 1.580E+02
DE 2.955E+03 1.395E+01 4.640E+03 9.565Eþ01 3.513E+03 6.624E+01
VS 2.885E+03 1.468E+01 4.161E+03 4.799E+02 3.482E+03 5.828E+01
HHO 3.407E+03 1.489E+02 6.452E+03 1.601E+03 4.270E+03 3.185E+02
CDLOBA 3.262E+03 9.695E+01 1.045E+04 1.434E+03 5.243E+03 6.654E+02
IGWO 2.935E+03 2.728E+01 4.743E+03 3.159E+02 3.809E+03 1.491E+02
EPSO 2.869E+03 2.380E+01 3.641E+03 6.077E+02 3.401E+03 5.693E+01
CLACOR 2.937E+03 2.989E+01 4.698E+03 3.787E+02 3.471E+03 1.345E+02
TSA 3.022E+03 1.025Eþ01 4.651E+03 1.242E+03 4.029E+03 1.398E+02
SPS_L_SHADE_EIG 2.884E+03 1.698E+01 4.073E+03 4.190E+02 3.439E+03 1.051E+02
Methods F31 F32 F33
Avg Std Avg Std Avg Std

SCWJAYA 5.354E+00 4.596E-01 1.245Eþ00 2.039E-01 2.045Eþ00 4.796E-01

JAYA 2.665E+03 5.291E+02 5.815E+00 7.169E-01 3.530E+00 2.548E-01
DE 7.352E+02 7.664E+01 3.001E+00 4.190E-01 2.503E+00 3.295E-01
VS 3.011E+02 9.385E+01 1.753E+00 1.149E+00 2.297E+00 7.107E-01
HHO 4.802E+00 2.198E-01 2.425E+00 1.119E+00 4.509E+00 3.384E-01
CDLOBA 1.948E+04 6.311E+03 8.657E+00 1.332E+00 4.808E+00 2.393E-01
IGWO 3.021E+02 1.766E+02 4.335E+00 2.950E+00 3.351E+00 4.648E-01
EPSO 5.128E+02 2.356E+02 4.586E+00 1.016E+00 2.845E+00 4.012E-01
CLACOR 5.018E+02 1.873E+02 2.001E+00 1.759E+00 3.012E+00 5.366E-01
TSA 2.143E+03 4.631E+02 5.900E+00 6.679E-01 3.342E+00 2.375E-01
SPS_L_SHADE_EIG 2.011Eþ00 1.028E+00 1.355E+00 1.415E-01 2.815E+00 4.591E-01

19
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Methods F34 F35

Avg Std Avg Std

SCWJAYA 1.089E+00 2.284E-02 1.176E+01 1.005E+01

JAYA 1.338E+00 5.929E-02 2.076E+01 1.069E+00
DE 1.119E+00 2.015E-02 2.034E+01 2.225E+00
VS 1.049Eþ00 2.015E-02 7.000Eþ00 9.322E+00
HHO 1.357E+00 1.483E-01 2.099E+01 1.492E-02
CDLOBA 1.383E+00 2.279E-01 2.129E+01 8.989E-02
IGWO 1.159E+00 6.943E-02 2.037E+01 3.654E+00
EPSO 1.177E+00 2.926E-02 9.867E+00 1.011E+01
CLACOR 1.064E+00 2.377E-02 2.100E+01 1.208E-03
TSA 1.159E+00 3.213E-02 1.743E+01 6.154E+00
SPS_L_SHADE_EIG 1.113E+00 4.248E-02 1.900E+01 6.102E+00

ranking among these combination methods. In the FT statistics results, during the iteration, and the blue line indicates the change in the local
SCWJAYA has an average ranking of 2.23 for thirty-five functions, and search during the iteration. The green line indicates the trend of the
SCWJAYA performs the best in the FT average ranking. balance status. The comparison of the first two columns in Fig. 11 shows
Fig. 9 shows the convergence curves of the strategy combination that SCWJAYA is more fully explored near the optimal solution. The
experiment. The curve of SCWJAYA is at the bottom of all the methods, balance status of SCWJAYA is postponed, and combined with the
as shown from the convergence curve. Notably, the four methods that experimental results of SCWJAYA and JAYA in thirty-five functions, we
introduce the sine chaos initialization strategy have a smaller average can assume that the balance status of SCWJAYA can achieve satisfactory
fitness value at the first iteration, which indicates that the sine chaos results. Fig. 11 c shows the process of population diversity change,
initialization strategy can improve the quality of the initial population. throughout which SCWJAYA is lower than JAYA in terms of population
By comparing CJAYA with JAYA, CWJAYA with WJAYA, SCJAYA with diversity. This result indicates that the distance between search agents of
SJAYA, and SCWJAYA with SWJAYA, it can be seen from the above four SCWJAYA is smaller, and the number of potential area search agents of
sets of convergence curves that the algorithm with the introduced the optimal solution is greater, which further improves the convergence
crisscross search strategy can improve the convergence accuracy and accuracy.
speed of the algorithm. From the convergence curves of four algorithms,
WJAYA, SWJAYA, CWJAYA, and SCWJAYA, it can be seen that the
4.6. Comparison experiment with SOTA methods
introduced wormhole search strategy can improve the convergence ac­
curacy and the ability to jump out of the local optimum of the algorithm.
To objectively and effectively verify the optimization capability of
In summary, the algorithm can maximize the optimization capability
the algorithm, this subsection compares SCWJAYA with several
of the algorithm when the three strategies are executed simultaneously.
advanced metaheuristic optimization methods, including JAYA, DE,
TSA, HHO, CDLOBA, IGWO, EPSO, CLACOR, VS, and SPS_L_SHADE_EIG.
4.5. Qualitative analysis of SCWJAYA Table 9 shows the mean and standard deviation of ten experiments
for eleven algorithms. The table shows that SCWJAYA performs best for
Fig. 10 shows the search trajectories of different functions. Fig. 10 a seventeen functions, and the optimization results for the other eighteen
shows the three-dimensional search space of the function. Fig. 10 b functions are all close to the best performing algorithms. These findings
shows the search trajectories of the search agents in two dimensions, indicate that SCWJAYA’s ability to optimize thirty-five global optimi­
where the red dot identifies the optimal position and the black dot in­ zation problems appears to be excellent from the objective point of view.
dicates the historical search position. Fig. 10 b shows that most of the Table 10 shows the p values, comparison results, and number of wins
search positions are locally searched at the global optimum point, which and losses between SCWJAYA and the ten optimization algorithms.
ensures that the algorithm can obtain a higher convergence accuracy. In Among all the comparison results, only a few have a p value greater than
addition, a small number of locations in the search space have been 0.05, which indicates that most of the comparison results are statistically
explored, indicating that SCWJAYA maintains global exploration while significant. According to the last column of Table 10, SCWJAYA out­
ensuring a local search. Fig. 10 c shows the search trajectory of the al­ performs JAYA, DE, and CDLOBA for all benchmark functions. Among
gorithm in the first dimension. The different function properties lead to the other ten comparison methods, SPS_L_SHADE_EIG performs the best.
different search trajectories. However, the convergence trajectory of the However, SCWJAYA still performs better than SPS_L_SHADE_EIG for
algorithm basically oscillates in a small magnitude at a certain iteration twenty functions and worse than SPS_L_SHADE_EIG for eight functions.
period. Fig. 10 d shows the average fitness value of SCWJAYA. The Fig. 12 shows the average ranking of WSRT and FT for the thirty-five
average fitness curve fluctuates because the global search is present functions. From these two rankings, it can be seen that SCWJAYA
throughout the iterations. ranks first in both statistics, with an average ranking of 1.97 and 2.41,
It is well known that global exploration and local exploration directly respectively, followed by SPS_L_SHADE_EIG. Surprisingly, the average
affect the optimization performance of an algorithm, and only explo­ ranking of both WSRT and FT of JAYA is 10th. From this figure, we can
ration and exploitation reach an optimal balance state for the algorithm see that the improvement strategy introduced in this paper effectively
to obtain the best results. In the ideal state, the algorithm requires more improves the optimization ability of the original JAYA algorithm.
global exploration in the early optimization process, and the algorithm Fig. 13 shows the convergence curves of the eleven algorithms. From
performs more local exploitation in the later optimization process. Since Fig. 13, it can be seen that SCWJAYA can find the solution faster, with
the optimization accuracy of the original JAYA algorithm cannot meet the exception of CLACOR and SPS_L_SHADE_EIG. It is worth noting that
the requirements, the balance status of the original algorithm is changed the convergence accuracy of SCWJAYA in the middle of the iteration is
by introducing improvement strategies. In this subsection, the experi­ not the best among these methods, but the wormhole search strategy
ment explores how SCWJAYA enhances the optimization ability through causes the algorithm to escape from the local optimal solution at the end
balance and population diversity analyses. Fig. 11 a and Fig. 11 b show of the iteration. Finally, SCWJAYA achieves the highest convergence
the balance experiment of SCWJAYA and JAYA, respectively. In Fig. 11 accuracy among all the comparison methods.
a and Fig. 11 b, the red line indicates the change in the global search The above experimental comparison shows that SCWJAYA exhibits

20
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 10
WSRT result of the comparison experiment with SOTA methods.
No. JAYA DE VS HHO CDLOBA

P-value S P-value S P-value S P-value S P-value S

F1 1.7344E-06 + 1.7344E-06 + 1.7344E-06 - 1.7791E-01 = 1.7344E-06 +

F2 1.7344E-06 + 1.7344E-06 + 1.2866E-03 - 1.7344E-06 + 1.7344E-06 +
F3 1.7344E-06 + 1.7344E-06 + 4.3896E-03 + 1.7344E-06 + 1.7344E-06 +
F4 1.7344E-06 + 1.7344E-06 + 2.6134E-04 + 1.7344E-06 + 1.7344E-06 +
F5 1.7344E-06 + 1.7344E-06 + 5.7517E-06 - 1.7344E-06 + 1.7988E-05 +
F6 1.7344E-06 + 2.8786E-06 + 6.1564E-04 - 1.7344E-06 + 2.3534E-06 +
F7 1.7344E-06 + 1.7344E-06 + 4.2767E-02 + 1.7344E-06 + 1.7344E-06 +
F8 1.7344E-06 + 1.7344E-06 + 5.7924E-05 + 1.7344E-06 + 1.7344E-06 +
F9 1.7344E-06 + 1.7344E-06 + 2.6033E-06 + 1.7344E-06 + 1.7344E-06 +
F10 1.7344E-06 + 4.3205E-08 + 1.7344E-06 + 9.7656E-04 - 1.7344E-06 +
F11 1.7344E-06 + 1.7344E-06 + 6.7328E-01 = 1.7344E-06 - 1.7344E-06 +
F12 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.0000E+00 = 1.7344E-06 +
F13 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.3733E-06 - 1.7344E-06 +
F14 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 2.6158E-07 - 1.7344E-06 +
F15 1.7344E-06 + 3.8822E-06 + 1.7344E-06 + 3.4180E-02 + 1.7344E-06 +
F16 1.7344E-06 + 1.7344E-06 + 1.7344E-06 - 1.7344E-06 + 2.7029E-02 +
F17 1.7344E-06 + 1.7344E-06 + 1.0357E-03 + 1.7344E-06 + 1.7344E-06 +
F18 1.7344E-06 + 1.7344E-06 + 9.8421E-03 + 1.7344E-06 + 1.7344E-06 +
F19 1.7344E-06 + 1.7344E-06 + 2.7653E-03 + 1.7344E-06 + 1.7344E-06 +
F20 1.7344E-06 + 1.7344E-06 + 1.0357E-03 + 1.7344E-06 + 1.7344E-06 +
F21 1.7344E-06 + 2.1053E-03 + 7.6909E-06 + 1.7344E-06 + 1.7344E-06 +
F22 1.7344E-06 + 4.8969E-04 + 2.3534E-06 + 1.7344E-06 + 1.7344E-06 +
F23 1.7344E-06 + 1.7988E-05 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F24 1.7344E-06 + 4.7292E-06 + 5.7517E-06 + 1.7344E-06 + 1.7344E-06 +
F25 1.7344E-06 + 1.7344E-06 + 1.2381E-05 + 2.6033E-06 + 1.7344E-06 +
F26 1.7344E-06 + 1.7344E-06 + 1.5625E-02 + 1.7344E-06 + 1.7344E-06 +
F27 1.7344E-06 + 1.7344E-06 + 7.1903E-02 = 1.7344E-06 + 1.7344E-06 +
F28 1.7344E-06 + 1.7344E-06 + 4.1955E-04 + 1.7344E-06 + 1.7344E-06 +
F29 2.3534E-06 + 1.7344E-06 + 5.5133E-05 + 2.8786E-06 + 1.7344E-06 +
F30 1.7344E-06 + 2.1266E-06 + 9.3157E-06 + 1.7344E-06 + 1.7344E-06 +
F31 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 4.7292E-06 - 1.7344E-06 +
F32 1.7344E-06 + 1.7344E-06 + 4.4052E-01 = 1.7344E-06 + 1.7344E-06 +
F33 1.7344E-06 + 8.3071E-04 + 1.2044E-01 = 1.7344E-06 + 1.7344E-06 +
F34 1.7344E-06 + 2.8434E-05 + 4.2857E-06 - 1.7344E-06 + 1.7344E-06 +
F35 1.1499E-04 + 1.1499E-04 + 8.6476E-02 = 1.5927E-03 + 1.7344E-06 +
+/− / = 35/0/0 35/0/0 24/6/5 28/5/2 35/0/0

No. IGWO EPSO CLACOR TSA SPS_L_SHADE_EIG

P-value S P-value S P-value S P-value S P-value S

F1 1.7344E-06 + 1.7344E-06 + 1.7344E-06 - 1.7344E-06 + 1.7344E-06 -

F2 1.7344E-06 + 1.7088E-03 + 9.3157E-06 + 3.4053E-05 + 6.4242E-03 +
F3 2.1266E-06 + 1.7344E-06 + 1.6394E-05 + 1.7344E-06 + 1.0639E-01 =
F4 1.7344E-06 + 2.8786E-06 + 1.2381E-05 + 1.7344E-06 + 3.3789E-03 +
F5 1.7344E-06 + 1.7344E-06 + 1.3059E-01 = 1.7344E-06 + 8.7297E-03 -
F6 1.7344E-06 + 1.7344E-06 + 9.6266E-04 + 1.7344E-06 + 6.4242E-03 +
F7 1.9209E-06 + 1.7344E-06 + 1.9209E-06 + 1.7344E-06 + 2.5637E-02 +
F8 2.1266E-06 + 4.0702E-02 + 2.8786E-06 + 8.2206E-02 = 8.1878E-05 -
F9 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 7.1570E-04 -
F10 1.7344E-06 + 4.3205E-08 + 1.7344E-06 + 1.7344E-06 + 4.3205E-08 +
F11 1.7344E-06 - 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F12 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 3.7896E-06 +
F13 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F14 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F15 1.7344E-06 + 2.5967E-05 + 3.2857E-01 = 1.9209E-06 + 1.5658E-02 +
F16 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 -
F17 1.7344E-06 + 3.1817E-06 + 1.7344E-06 + 1.7344E-06 + 8.4508E-01 =
F18 1.7344E-06 + 1.7344E-06 + 5.7517E-06 + 1.7344E-06 + 1.0639E-01 =
F19 2.3534E-06 + 1.9209E-06 + 5.2165E-06 + 1.7344E-06 + 7.1903E-02 =
F20 1.9209E-06 + 1.7344E-06 + 3.8822E-06 + 1.7344E-06 + 6.4242E-03 +
F21 1.9209E-06 + 1.3595E-04 + 6.8923E-05 + 1.7344E-06 + 4.3896E-03 +
F22 1.7344E-06 + 9.9179E-01 = 7.1903E-02 = 8.5896E-02 = 1.7344E-06 -
F23 1.7344E-06 + 2.1053E-03 + 5.2165E-06 + 1.1093E-01 = 2.8786E-06 -
F24 1.7344E-06 + 1.7088E-03 + 4.7292E-06 + 1.9209E-06 + 4.7292E-06 +
F25 1.7344E-06 + 3.8822E-06 + 2.1266E-06 + 1.7344E-06 + 2.0515E-04 +
F26 1.7344E-06 + 2.8288E-05 + 8.2981E-06 + 1.7344E-06 + 2.5000E-01 =
F27 1.9209E-06 + 1.3820E-03 - 1.9209E-06 + 1.7344E-06 + 6.7328E-01 =
F28 1.7344E-06 + 9.3676E-02 = 2.3534E-06 + 1.7344E-06 + 7.7309E-03 +
F29 1.7344E-06 + 2.9882E-01 = 1.7344E-06 + 9.7110E-05 + 1.1138E-03 +
F30 1.7344E-06 + 5.4401E-01 = 4.3896E-03 + 1.7344E-06 + 3.8723E-02 +
F31 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 + 1.7344E-06 -
F32 1.3601E-05 + 1.7344E-06 + 2.0589E-01 = 1.7344E-06 + 9.5899E-01 =
F33 1.7344E-06 + 1.4936E-05 + 3.1817E-06 + 1.7344E-06 + 4.2857E-06 +
(continued on next page)

21
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 10 (continued )
No. IGWO EPSO CLACOR TSA SPS_L_SHADE_EIG

P-value S P-value S P-value S P-value S P-value S

F34 7.6909E-06 + 2.3534E-06 + 1.1973E-03 - 1.7344E-06 + 1.9646E-03 +

F35 2.1266E-06 + 7.3749E-01 = 4.8603E-05 + 1.0444E-02 + 4.1691E-03 +
+/− / = 34/1/0 29/1/5 29/2/4 32/0/3 20/8/7

Fig. 12. WSRT and FT mean ranks of the comparison experiment with SOTA methods.

excellent optimization performance for global optimization problems, to evaluate the performance of the proposed BSCWJAYA_KELM FS
even when compared with many state-of-the-art algorithms. method. The comparison algorithms include BJAYA, BGWO, BHHO,
BMFO, BPSO, BSMA, BSSA, BCLJAYA, and BEJAYA. The algorithm’s
unique parameters remain intact, with the exception of the trans­
4.7. Comparison experiment with peer methods formation to a discrete version suitable for FS, as indicated in Table 13.

To further verify the performance of the algorithm, in this subsec­

tion, SCWJAYA is compared with six JAYA variants, including EO_JAYA, 5.2. Evaluation criteria
CLJAYA, EJAYA, IJAYA, JAYA_LF, and LJA. The experiment in this
subsection also examines the optimization performance of SCWJAYA In the FS experiments, accuracy, specificity, precision, MCC, and F-
using thirty-five global optimization problems. measure are used in this subsection to evaluate the performance of the
Table 11 shows the mean and standard deviation of the seven model. The evaluation metrics are defined as shown in Eqs. 16–20.
enhanced JAYA algorithms. SCWJAYA clearly has the best mean and TP + TN
Accuracy = (16)
standard deviation, which indicates that it has the best optimization TP + FP + FN + TN
performance and stability experimental results among the comparisons
in this subsection. Table 12 shows the WSRT comparison results of Specificity =
TN
(17)
SCWJAYA and peers. SCWJAYA performs the best for more than thirty- TN + FP
two benchmark functions. From the average ranking of WSRT in Fig. 14,
TP
it can be seen that SCWJAYA is the best performing algorithm, with an Precision = (18)
TP + FP
average ranking of 1.09. Moreover, SCWJAYA ranks No. 1 in the FT
experiment results with an average ranking of 1.35. Fig. 15 shows the TP × TN − FP × FN
convergence curves of the seven JAYA variants for different benchmark MCC = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (19)
(TP + FP) × (TP + FN) × (TN + FP) × (TN + FN)
functions. From Fig. 15, it can be seen that SCWJAYA has a great
advantage in terms of optimization quality and convergence speed. TP
According to the experiments and analysis in this subsection, F − measure = (20)
TP + FN+FP
SCWJAYA can beat the six excellent JAYA variants in terms of most
2

functions. where TP is the number of true positive instances, TN is the number of

true negative instances, FN is the number of false-negative instances,
5. Experiments for the feature selection problem and FP is the number of false-positive instances.
Accuracy indicates the correct rate of model classification for true-
5.1. Experimental settings positive and true-negative instances. A higher accuracy rate indicates
a larger proportion of correctly predicted samples. Specificity indicates
This section compares BSCWJAYA with other well-known algorithms the correct model classification rate of positive negatives in negative

22
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 13. Convergence curve of the comparison experiment with SOTA methods.

23
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 11
Avg and Std results of the comparison experiment with peer methods.
Methods F1 F2 F3

Avg Std Avg Std Avg Std

SCWJAYA 5.201Eþ02 1.944E-02 6.055Eþ02 2.809E+00 9.584Eþ02 1.264E+01

EO_JAYA 5.209E+02 5.026E-02 6.302E+02 5.146E+00 1.136E+03 1.439E+01
CLJAYA 5.209E+02 6.769E-02 6.118E+02 5.409E+00 1.019E+03 5.918E+01
EJAYA 5.206E+02 1.634E-01 6.230E+02 3.491E+00 9.998E+02 2.197E+01
IJAYA 5.209E+02 4.982E-02 6.333E+02 1.123Eþ00 1.140E+03 1.248Eþ01
JAYA_LF 5.210E+02 4.128E-02 6.064E+02 1.856E+00 1.106E+03 1.766E+01
LJA 5.209E+02 5.268E-02 6.326E+02 1.269E+00 1.150E+03 1.450E+01
Methods F4 F5 F6
Avg Std Avg Std Avg Std

SCWJAYA 3.034Eþ03 2.736Eþ02 1.200Eþ03 6.808E-02 1.300Eþ03 5.525E-02

EO_JAYA 8.077E+03 2.786E+02 1.202E+03 2.523E-01 1.300E+03 4.674E-02
CLJAYA 6.364E+03 1.439E+03 1.202E+03 4.030E-01 1.300E+03 1.108E-01
EJAYA 5.010E+03 5.191E+02 1.201E+03 4.873E-01 1.300E+03 8.399E-02
IJAYA 7.821E+03 3.440E+02 1.202E+03 3.116E-01 1.301E+03 7.214E-02
JAYA_LF 7.880E+03 3.729E+02 1.202E+03 2.822E-01 1.300E+03 8.386E-02
LJA 7.816E+03 3.461E+02 1.202E+03 2.928E-01 1.302E+03 3.571E-01
Methods F7 F8 F9
Avg Std Avg Std Avg Std

SCWJAYA 1.610Eþ03 5.783E-01 2.362Eþ03 7.618Eþ02 2.079Eþ03 2.114Eþ01

EO_JAYA 1.613E+03 1.874E-01 7.450E+06 1.374E+07 7.608E+03 2.316E+03
CLJAYA 1.612E+03 3.705E-01 7.189E+03 7.236E+03 2.377E+03 1.383E+02
EJAYA 1.612E+03 5.185E-01 6.518E+03 5.416E+03 2.338E+03 1.145E+02
IJAYA 1.613E+03 1.927E-01 4.284E+04 2.861E+04 3.683E+03 8.142E+02
JAYA_LF 1.612E+03 2.088E-01 3.455E+04 1.513E+05 2.970E+03 4.468E+02
LJA 1.613E+03 1.949E-01 2.837E+07 1.396E+07 7.149E+03 2.379E+03
Methods F10 F11 F12
Avg Std Avg Std Avg Std

SCWJAYA 2.500Eþ03 1.051E-12 2.600E+03 5.117E-02 2.700Eþ03 1.689E-13

EO_JAYA 2.638E+03 7.505E+00 2.600Eþ03 1.507E-03 2.702E+03 4.941E+00
CLJAYA 2.615E+03 2.442E-12 2.611E+03 1.337E+01 2.704E+03 2.348E+00
EJAYA 2.615E+03 1.216E-11 2.634E+03 9.480E+00 2.713E+03 3.475E+00
IJAYA 2.618E+03 4.840E-01 2.654E+03 2.332E+00 2.719E+03 2.329E+00
JAYA_LF 2.615E+03 1.971E-12 2.628E+03 6.490E+00 2.704E+03 1.404E+00
LJA 2.646E+03 4.318E+00 2.659E+03 4.783E+00 2.724E+03 3.512E+00
Methods F13 F14 F15
Avg Std Avg Std Avg Std

SCWJAYA 2.900Eþ03 2.805E-02 3.000Eþ03 3.272E-05 3.549Eþ03 7.224Eþ02

EO_JAYA 3.417E+03 2.164E+02 5.343E+03 8.807E+02 8.024E+06 5.379E+06
CLJAYA 3.391E+03 1.916E+02 4.554E+03 3.508E+02 3.612E+07 3.959E+07
EJAYA 3.502E+03 2.410E+02 4.124E+03 2.737E+02 3.886E+06 5.360E+06
IJAYA 3.332E+03 1.982E+02 4.279E+03 1.389E+02 1.964E+05 7.078E+04
JAYA_LF 3.174E+03 7.441E+01 3.735E+03 9.309E+01 9.927E+03 1.057E+04
LJA 3.317E+03 1.608E+02 4.900E+03 3.495E+02 6.162E+06 1.970E+06
Methods F16 F17 F18
Avg Std Avg Std Avg Std

SCWJAYA 3.053E+02 2.207E+00 5.522Eþ02 1.237Eþ01 7.913Eþ02 1.140Eþ01

EO_JAYA 3.163E+04 6.965E+03 7.277E+02 1.370E+01 9.743E+02 3.726E+01
CLJAYA 6.404E+02 2.576E+02 5.978E+02 5.416E+01 9.026E+02 5.001E+01
EJAYA 3.000Eþ02 1.569E-09 6.095E+02 2.705E+01 8.946E+02 3.855E+01
IJAYA 3.828E+04 7.216E+03 7.256E+02 1.335E+01 1.038E+03 1.927E+01
JAYA_LF 5.679E+03 2.165E+03 7.033E+02 1.798E+01 9.472E+02 1.581E+01
LJA 5.084E+04 6.165E+03 7.535E+02 1.464E+01 1.067E+03 2.216E+01
Methods F19 F20 F21
Avg Std Avg Std Avg Std

SCWJAYA 8.540Eþ02 9.974Eþ00 3.265Eþ03 2.595E+02 1.131Eþ03 1.483Eþ01

EO_JAYA 1.020E+03 1.198E+01 8.142E+03 2.579E+02 1.449E+03 1.056E+02
CLJAYA 8.998E+02 5.380E+01 6.263E+03 9.474E+02 1.192E+03 5.186E+01
EJAYA 8.918E+02 2.249E+01 5.271E+03 8.968E+02 1.227E+03 4.668E+01
IJAYA 1.039E+03 1.258E+01 7.915E+03 3.313E+02 1.339E+03 2.202E+01
JAYA_LF 1.003E+03 1.928E+01 7.922E+03 2.799E+02 1.194E+03 4.970E+01
LJA 1.050E+03 1.287E+01 8.015E+03 2.522Eþ02 2.106E+03 1.347E+02
Methods F22 F23 F24
Avg Std Avg Std Avg Std

SCWJAYA 1.453Eþ04 1.411Eþ04 2.670Eþ03 1.459Eþ03 1.772Eþ03 4.341Eþ01

EO_JAYA 2.737E+05 1.014E+06 4.128E+06 3.674E+06 2.316E+03 1.436E+02
CLJAYA 1.928E+04 2.084E+04 8.153E+03 9.829E+03 1.966E+03 1.226E+02
EJAYA 2.199E+04 1.781E+04 5.112E+03 4.897E+03 1.994E+03 1.280E+02
IJAYA 6.307E+04 3.124E+04 2.632E+04 1.070E+04 2.147E+03 8.967E+01
JAYA_LF 2.151E+04 1.918E+04 1.764E+05 2.570E+05 2.097E+03 1.321E+02
(continued on next page)

24
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 11 (continued )
Methods F1 F2 F3

Avg Std Avg Std Avg Std

LJA 8.651E+06 8.956E+06 2.350E+06 1.012E+06 2.306E+03 1.004E+02

Methods F25 F26 F27
Avg Std Avg Std Avg Std

SCWJAYA 2.347Eþ03 2.979E+01 2.300Eþ03 3.870E-13 2.700Eþ03 1.044Eþ01

EO_JAYA 2.506E+03 1.649E+01 2.579E+03 1.165E+02 2.963E+03 3.901E+01
CLJAYA 2.387E+03 4.925E+01 3.483E+03 2.198E+03 2.743E+03 4.048E+01
EJAYA 2.400E+03 2.430E+01 3.590E+03 1.886E+03 2.777E+03 3.867E+01
IJAYA 2.516E+03 1.508E+01 2.353E+03 1.550E+01 2.905E+03 1.803E+01
JAYA_LF 2.489E+03 2.017E+01 2.301E+03 1.391E+00 2.835E+03 4.907E+01
LJA 2.530E+03 1.204Eþ01 3.028E+03 3.693E+02 2.983E+03 1.556E+01
Methods F28 F29 F30
Avg Std Avg Std Avg Std

SCWJAYA 2.869Eþ03 1.106Eþ01 3.296Eþ03 6.154E+02 3.382Eþ03 5.081Eþ01

EO_JAYA 3.086E+03 4.403E+01 5.654E+03 9.840E+02 4.638E+03 1.820E+02
CLJAYA 2.921E+03 5.262E+01 4.308E+03 3.662E+02 3.566E+03 1.033E+02
EJAYA 2.948E+03 4.019E+01 5.015E+03 9.844E+02 3.815E+03 2.336E+02
IJAYA 3.065E+03 1.476E+01 6.132E+03 3.777E+02 4.235E+03 1.543E+02
JAYA_LF 3.010E+03 3.740E+01 5.067E+03 8.702E+02 3.860E+03 2.773E+02
LJA 3.137E+03 2.025E+01 6.608E+03 1.063Eþ03 4.560E+03 1.536E+02
Methods F31 F32 F33
Avg Std Avg Std Avg Std

SCWJAYA 5.422Eþ00 4.985E-01 1.314Eþ00 1.760E-01 1.957Eþ00 4.614E-01

EO_JAYA 5.182E+01 1.018E+02 4.613E+00 1.045E+00 3.756E+00 3.242E-01
CLJAYA 3.813E+02 1.037E+02 1.412E+00 1.982E-01 3.120E+00 4.327E-01
EJAYA 1.441E+02 7.575E+01 1.395E+00 7.470E-02 3.596E+00 5.639E-01
IJAYA 2.008E+03 4.519E+02 6.815E+00 7.965E-01 3.792E+00 2.029E-01
JAYA_LF 6.553E+02 1.996E+02 5.340E+00 6.339E-01 3.241E+00 3.376E-01
LJA 2.787E+03 5.027E+02 7.164E+00 5.662E-01 3.765E+00 2.089E-01

Methods F34 F35

Avg Std Avg Std

SCWJAYA 1.098Eþ00 1.872E-02 1.020Eþ01 9.770E+00

EO_JAYA 1.142E+00 4.152E-02 2.059E+01 1.139E+00
CLJAYA 1.128E+00 4.054E-02 1.958E+01 5.192E+00
EJAYA 1.087E+00 4.825E-02 2.116E+01 6.993E-02
IJAYA 1.192E+00 3.416E-02 2.087E+01 8.292E-01
JAYA_LF 1.131E+00 2.759E-02 1.882E+01 6.148E+00
LJA 1.343E+00 7.509E-02 2.003E+01 1.431E+00

instances. A higher specificity indicates a lower classification error. smallest. The experimental results of the BSCWJAYA-based method are
Precision is the probability that a sample is positive among all the more stable in other datasets. According to the above analysis, the
samples that are predicted to be positive. A higher precision indicates a BSCWJAYA_KELM FS method outperforms all other methods in terms of
more accurate prediction of positive cases. MCC indicates the reliability prediction accuracy and stability.
of the model. A closer MCC to 1 indicates a more perfect prediction of Table 16 shows the specificity results of the ten algorithms. It is
the subject. The F-measure can comprehensively evaluate a classifier. A obvious from the table that the mean values of the BSCWJAYA-based
higher F-measure indicates that the classification results follow method are all No 1. The average specificity results of BSCWJAYA can
expectations. reach 98.14%, 100.00%, 98.33%, 100.00%, 94.41%, 100.00%, 98.93%,
100.00% and 92.26%, respectively. The standard deviation results are
not the best performance only for the Ionosphere and Wielaw datasets,
5.3. Feature selection for public datasets
but the prediction results of BSCWJAYA are equally stable. Therefore,
the classification error of BSCWJAYA_KELM is lower in these compar­
In this subsection, we demonstrate the overall predictive capability
ative methods species.
of BSCWJAYA_KELM. The BSCWJAYA_KELM FS framework is compared
Table 17 shows the mean and standard deviation of the precision
with the common nine methods for the University of California Irvine
results. The average results of the BSCWJAYA-based method on the nine
datasets (UCI) (https://archive.ics.uci.edu/ml/datasets.php) and the
datasets reaches 98.93%, 100.00%, 98.89%, 100.00%, 97.27%,
Wielaw dataset [95,96]. The proposed method aims to improve the
100.00%, 100.00%, 97.27%, and 92.48%, respectively. The average
prediction accuracy of the HD dataset. Therefore, the datasets we use in
results according to Table 17 indicate that the BSCWJAYA-based
this subsection are all dichotomous problems. The details of the public
method has the best and most stable Precision results. Therefore,
dataset are shown in Table 14.
BSCWJAYA_KELM is more accurate in predicting positive instances in
Table 15 shows the accuracy results of ten algorithms for the UCI
comparison to other prediction methods.
dataset, and the BSCWJAYA-based method has the best average results
Table 18 shows the mean and standard deviation of MCC. In the
over 20 independently run experiments. The mean accuracy of the
mean results of MCC, the proposed method in this paper reaches 0.9853,
BSCWJAYA-based method for the nine datasets is 99.30%, 99.09%,
0.9819, 0.9669, 0.9655, 0.9578, 1.0000, 0.9873, 0.9737 and 0.8302.
98.21%, 98.75%, 98.04%, 100.00%, 99.33%, 98.76%, and 90.86%,
According to the comparison results presented in Table 18, the
respectively. In the Heart dataset, Ionosphere dataset, and Wielaw
BSCWJAYA-based method can show better performance and more stable
dataset, respectively, the standard deviation of the of the BPSO-based
results. Therefore, the proposed BSCWJAYA_KELM is better for
method, BJAYA-based method, and BEJAYA-based method is the

25
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 12
WSRT result of the comparison experiment with peer methods.
No. EO_JAYA CLJAYA EJAYA

P-value S P-value S P-value S

F1 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +

F2 1.7344E-06 + 1.1265E-05 + 1.7344E-06 +
F3 1.7344E-06 + 9.7110E-05 + 3.8822E-06 +
F4 1.7344E-06 + 2.3534E-06 + 1.7344E-06 +
F5 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F6 2.1266E-06 + 5.7517E-06 + 1.4936E-05 +
F7 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F8 3.1817E-06 + 5.7924E-05 + 6.6392E-04 +
F9 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F10 1.7344E-06 + 4.3205E-08 + 1.0144E-07 +
F11 1.7344E-06 - 1.7088E-03 + 1.7344E-06 +
F12 2.5000E-01 = 5.5627E-06 + 1.7344E-06 +
F13 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F14 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F15 1.7344E-06 + 2.3534E-06 + 1.1265E-05 +
F16 1.7344E-06 + 2.1266E-06 + 1.7344E-06 -
F17 1.7344E-06 + 2.2248E-04 + 1.9209E-06 +
F18 1.7344E-06 + 2.8786E-06 + 1.7344E-06 +
F19 1.7344E-06 + 4.1955E-04 + 1.7344E-06 +
F20 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F21 1.7344E-06 + 6.9838E-06 + 2.3534E-06 +
F22 2.0515E-04 + 3.8203E-01 = 3.8723E-02 +
F23 1.7344E-06 + 2.2248E-04 + 2.5637E-02 +
F24 1.7344E-06 + 3.1817E-06 + 1.9209E-06 +
F25 1.7344E-06 + 1.2866E-03 + 1.7344E-06 +
F26 1.7344E-06 + 2.7343E-04 + 2.1073E-06 +
F27 1.7344E-06 + 2.1630E-05 + 1.9209E-06 +
F28 1.7344E-06 + 1.1265E-05 + 1.9209E-06 +
F29 1.7344E-06 + 2.1266E-06 + 1.2381E-05 +
F30 1.7344E-06 + 2.3534E-06 + 1.7344E-06 +
F31 3.8822E-06 + 1.7344E-06 + 1.7344E-06 +
F32 1.7344E-06 + 4.9080E-01 = 2.9894E-01 =
F33 1.7344E-06 + 2.6033E-06 + 2.1266E-06 +
F34 2.8434E-05 + 1.0357E-03 + 1.4139E-01 =
F35 1.2506E-04 + 6.8923E-05 + 1.7344E-06 +
+/− / = 33/1/1 33/0/2 32/1/2

No. IJAYA JAYA_LF LJA

P-value S P-value S P-value S

F1 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +

F2 1.7344E-06 + 6.8714E-02 = 1.7344E-06 +
F3 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F4 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F5 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F6 1.7344E-06 + 8.4661E-06 + 1.7344E-06 +
F7 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F8 1.7344E-06 + 7.1570E-04 + 1.7344E-06 +
F9 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F10 1.7344E-06 + 4.3205E-08 + 1.7344E-06 +
F11 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F12 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F13 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F14 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F15 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F16 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F17 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F18 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F19 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F20 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F21 1.7344E-06 + 3.7243E-05 + 1.7344E-06 +
F22 3.1817E-06 + 1.5286E-01 = 1.7344E-06 +
F23 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F24 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F25 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F26 1.7344E-06 + 1.9531E-03 + 1.7344E-06 +
F27 1.7344E-06 + 1.9209E-06 + 1.7344E-06 +
F28 1.7344E-06 + 1.9209E-06 + 1.7344E-06 +
F29 1.7344E-06 + 2.3534E-06 + 1.9209E-06 +
F30 1.7344E-06 + 2.6033E-06 + 1.7344E-06 +
F31 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F32 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
F33 1.7344E-06 + 1.7344E-06 + 1.7344E-06 +
(continued on next page)

26
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 12 (continued )
No. IJAYA JAYA_LF LJA

P-value S P-value S P-value S

F34 1.7344E-06 + 6.3198E-05 + 1.7344E-06 +

F35 5.3070E-05 + 9.7536E-04 + 3.0650E-04 +
+/− / = 35/0/0 33/0/2 35/0/0

Fig. 14. WSRT and FT mean ranks of the comparison experiment with peer methods.

prediction of the target dataset. 5.4. Feature selection for the HD dataset
Table 19 shows the mean and standard deviation of the F-measure.
According to the data in the table, the mean standards of the 5.4.1. Dataset description
BSCWJAYA-based method reach 99.45%, 99.24%, 98.36%, 96.67%, Patients were recruited in a single center from May 5 to May 31,
98.56%, 100.00%, 99.23%, 98.21% and 90.13%, respectively. The 2020 (outpatient HD unit of First Affiliated Hospital of Wenzhou Med­
BSCWJAYA-based method is the most stable among the experimental ical University). The inclusion criteria were as follows: (1) age ≥18
results of the Breast, Congress, Hepatitis, Sonar, and Vote datasets. years; (2) three times per week for HD, with each treatment lasting 4 h.
Based on the analysis of the experimental data shown in Table 19, we The exclusion criteria were as follows: (1) treatment sessions with
can conclude that the BSCWJAYA_KELM method proposed in this paper missing data; (2) an unexpected death during a session of treatment; (3)
is a more effective classification method. alanine transaminase or aspartate transaminase or bilirubin above the
The comprehensive prediction performance of BSCWJAYA_KELM upper limits of normal. There were 178 patients with a total of 1940 HD
was validated on the UCI dataset and the Wielaw dataset. However, due sessions included in the study. Nobody uses febuxostat or allopurinol for
to the small amount of data and the characteristics of the dataset, hypouricemic treatment since our institution does not accept off-label
overfitting would lead to experimental results that are not realistic and applications. The initial dialysate temperature was 37 ◦ C, the sodium
objective. Therefore, in this subsection, we also use two medical datasets concentration was 140 mmol/L, and the calcium concentration was 1.5
with a larger data size and with validated objectivity and efficiency, mmol/L. Patient blood was tested once a month. The samples were
which are described in Ref. [97] and Ref. [98]. The sizes of these two collected before dialysis. Patient B_P was measured 5 times during each
datasets are 1940*15 and 3071*69. Fig. 16 shows the results of 20 treatment session: at 0, 1, 2, and 3 h after drawing blood and at rein­
independently run experiments on the two medical datasets. The figure fusion. IDH was defined as a reduction in SBP of 20 mmHg or a reduction
shows that BSCWJAYA_KELM also outperforms other similar methods in MAP of 10 mmHg from predialysis to nadir intradialytic levels plus
on both medical datasets. In other words, the predictive performance of ≥2 responsive measures [99]. The attributes of the dataset are shown in
the model proposed in this paper is equally applicable to other medical Table 20.
datasets and not just the data used in this paper.
Based on the comparative experiments for the UCI dataset in this 5.4.2. Experiment and analysis
subsection, it can be seen that BSCWJAYA_KELM has reliable and The performance of the classifier can also affect the prediction per­
excellent prediction performance. From the experimental results of ac­ formance. Therefore, the optimal subset of features obtained from
curacy, specificity, precision, MCC, and F-measure, it seems that SCWJAYA that are input to which classifier performs the classification
BSCWJAYA_KELM achieves the expected goal. task is also the focus of this paper. Therefore, BSCWJAYA is combined
with K-nearest neighbor (KNN), fuzzy K-nearest neighbor (FKNN),
support vector machine (SVM), and KELM for predicting IDH. The
reliability of the experimental results is verified based on the accuracy,
specificity, precision, MCC, F-measure, and time cost. The parameter

27
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 15. Convergence curve of the comparison experiment with peer methods.

Table 13
Parameter settings of the binary version algorithms. Table 14
Methods Public Other Parameters Description of public datasets.
parameters Datasets Samples Features Classes
BSCWJAYA T = 100; ε1 , ε2 , a1 random ​ in ​ [0, 1]; a1∈(0,4]; p = 0.1 Breast 569 31 2
BJAYA N = 20 rbest = random ​ in ​ [0, 1]; rworst = random ​ in ​ [0, 1] Congress 435 17 2
BGWO a = [20] Heart 270 14 2
BHHO ωa = 0.7; ωb = 0.2; |E| = [20]; β = 1.5 hepatitis 155 20 2
BMFO b =1 Ionosphere 351 35 2
BPSO c1 = 2; c2 = 2; Vmax = 6 Sonar 208 61 2
BSMA z = 0.03; vc = [10]; r = random ​ in ​ [0, 1] Vote 101 17 2
BSSA c1 = random ​ in ​ [0, 1]; c2 = random ​ in ​ [0, 1]; β = wdbc 569 31 2
1.5 Wielaw 240 31 2
BCLJAYA φ1 , φ2 , φ3 ,
φ4 ​ subject ​ to ​ standard ​ normal ​ distribution
BEJAYA λ1 = random ​ in ​ [0, 1]; λ2 = random ​ in ​ [0, 1];
k subject ​ to ​ standard ​ normal ​ distribution

28
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 15
Accuracy values on the UCI and Wielaw datasets.
Methods Breast Congress Heart

Avg Std Avg Std Avg Std

BSCWJAYA 99.30% 1.442E-02 99.09% 1.869E-02 98.21% 3.929E-02

BJAYA 99.13% 1.554E-02 98.60% 2.189E-02 96.62% 3.837E-02
BGWO 99.12% 1.565E-02 98.20% 3.047E-02 96.68% 4.526E-02
BHHO 99.13% 1.556E-02 98.19% 3.056E-02 96.70% 3.744E-02
BMFO 99.13% 1.906E-02 97.92% 3.476E-02 96.68% 4.430E-02
BPSO 98.95% 2.299E-02 98.18% 3.095E-02 97.03% 3.733E-02
BSMA 98.95% 1.986E-02 98.87% 2.005E-02 96.62% 5.213E-02
BSSA 99.11% 1.576E-02 98.83% 2.078E-02 96.68% 3.775E-02
BCLJAYA 98.08% 4.277E-02 97.94% 2.335E-02 93.96% 1.104E-01
BEJAYA 98.78% 1.714E-02 96.49% 6.673E-02 96.68% 4.526E-02
Methods hepatitis Ionosphere Sonar
Avg Std Avg Std Avg Std

BSCWJAYA 98.75% 3.847E-02 98.04% 3.282E-02 100.00% 0.000Eþ00

BJAYA 98.26% 4.248E-02 97.73% 2.860E-02 100.00% 0.000Eþ00
BGWO 97.50% 5.130E-02 97.12% 4.388E-02 99.05% 2.942E-02
BHHO 96.30% 5.830E-02 97.73% 3.377E-02 99.05% 2.942E-02
BMFO 97.57% 4.995E-02 96.83% 3.511E-02 99.09% 2.798E-02
BPSO 97.50% 5.130E-02 97.71% 3.897E-02 99.05% 2.942E-02
BSMA 98.13% 4.579E-02 97.12% 3.516E-02 99.55% 2.033E-02
BSSA 97.39% 5.389E-02 97.17% 3.825E-02 99.00% 4.472E-02
BCLJAYA 96.23% 5.931E-02 97.14% 3.502E-02 99.55% 2.033E-02
BEJAYA 97.95% 5.027E-02 97.40% 3.513E-02 97.27% 8.394E-02
Methods Vote wdbc Wielaw
Avg Std Avg Std Avg Std

BSCWJAYA 99.33% 2.052E-02 98.76% 1.731E-02 90.86% 7.616E-02

BJAYA 99.00% 2.442E-02 98.43% 2.125E-02 90.30% 9.743E-02
BGWO 98.69% 2.695E-02 98.60% 1.765E-02 88.33% 8.281E-02
BHHO 98.33% 2.979E-02 98.43% 2.104E-02 90.05% 6.894E-02
BMFO 98.67% 3.488E-02 98.40% 2.179E-02 90.25% 6.519E-02
BPSO 99.33% 2.052E-02 98.43% 2.383E-02 90.25% 8.387E-02
BSMA 98.31% 3.006E-02 98.58% 2.965E-02 89.20% 8.827E-02
BSSA 98.64% 3.632E-02 98.58% 1.789E-02 90.21% 6.515E-02
BCLJAYA 96.69% 9.268E-02 98.42% 1.790E-02 88.10% 6.599E-02
BEJAYA 97.31% 7.630E-02 98.42% 2.412E-02 89.69% 6.442E-02

29
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 16
Specificity value on UCI datasets and the Wielaw dataset.
Methods Breast Congress Heart

Avg Std Avg Std Avg Std

BSCWJAYA 98.14% 3.828E-02 100.00% 0.000Eþ00 98.33% 5.130E-02

BJAYA 97.68% 4.124E-02 98.26% 4.248E-02 96.67% 6.840E-02
BGWO 97.68% 4.124E-02 99.44% 2.485E-02 95.00% 7.836E-02
BHHO 97.68% 4.124E-02 98.82% 3.641E-02 94.17% 8.156E-02
BMFO 97.73% 5.001E-02 98.89% 3.420E-02 95.00% 7.836E-02
BPSO 97.18% 6.228E-02 98.75% 3.847E-02 94.17% 8.156E-02
BSMA 97.18% 5.316E-02 99.44% 2.485E-02 94.17% 8.156E-02
BSSA 97.64% 4.207E-02 100.00% 0.000Eþ00 96.67% 6.840E-02
BCLJAYA 94.86% 1.136E-01 97.71% 4.711E-02 90.00% 1.658E-01
BEJAYA 96.73% 4.583E-02 96.39% 9.721E-02 95.83% 7.404E-02
Methods hepatitis Ionosphere Sonar
Avg Std Avg Std Avg Std

BSCWJAYA 100.00% 0.000Eþ00 94.41% 9.532E-02 100.00% 0.000Eþ00

BJAYA 100.00% 0.000Eþ00 93.57% 8.106E-02 100.00% 0.000Eþ00
BGWO 100.00% 0.000Eþ00 92.14% 1.207E-01 99.17% 3.727E-02
BHHO 100.00% 0.000Eþ00 94.41% 7.850E-02 99.00% 4.472E-02
BMFO 100.00% 0.000Eþ00 91.07% 9.919E-02 100.00% 0.000Eþ00
BPSO 100.00% 0.000Eþ00 93.45% 1.124E-01 100.00% 0.000Eþ00
BSMA 100.00% 0.000Eþ00 91.91% 9.935E-02 100.00% 0.000Eþ00
BSSA 100.00% 0.000Eþ00 92.14% 1.042E-01 100.00% 0.000Eþ00
BCLJAYA 100.00% 0.000Eþ00 91.91% 9.935E-02 99.17% 3.727E-02
BEJAYA 100.00% 0.000Eþ00 92.86% 9.771E-02 96.67% 1.160E-01
Methods Vote wdbc Wielaw
Avg Std Avg Std Avg Std

BSCWJAYA 98.89% 3.420E-02 100.00% 0.000Eþ00 92.26% 1.255E-01

BJAYA 98.89% 3.420E-02 100.00% 0.000Eþ00 88.21% 1.139E-01
BGWO 97.83% 4.451E-02 100.00% 0.000Eþ00 88.21% 1.261E-01
BHHO 97.33% 4.747E-02 100.00% 0.000Eþ00 89.17% 8.687E-02
BMFO 97.89% 5.418E-02 100.00% 0.000Eþ00 90.24% 1.358E-01
BPSO 98.89% 3.420E-02 100.00% 0.000Eþ00 89.52% 1.612E-01
BSMA 98.89% 3.420E-02 100.00% 0.000Eþ00 91.31% 1.080E-01
BSSA 98.39% 3.940E-02 100.00% 0.000Eþ00 89.05% 1.029E-01
BCLJAYA 95.06% 1.505E-01 100.00% 0.000Eþ00 88.45% 9.931E-02
BEJAYA 96.89% 9.328E-02 100.00% 0.000Eþ00 91.07% 1.253E-01

30
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 17
Precision value on UCI datasets and Wielaw dataset.
Methods Breast Congress Heart

Avg Std Avg Std Avg Std

BSCWJAYA 98.93% 2.191E-02 100.00% 0.000Eþ00 98.89% 3.420E-02

BJAYA 98.67% 2.365E-02 98.93% 2.617E-02 97.57% 4.995E-02
BGWO 98.66% 2.391E-02 99.62% 1.720E-02 96.44% 5.622E-02
BHHO 98.67% 2.365E-02 99.28% 2.216E-02 95.90% 5.744E-02
BMFO 98.70% 2.831E-02 99.23% 2.390E-02 96.39% 5.671E-02
BPSO 98.47% 3.315E-02 99.29% 2.199E-02 95.83% 5.841E-02
BSMA 98.45% 2.925E-02 99.67% 1.491E-02 95.66% 6.128E-02
BSSA 98.66% 2.391E-02 100.00% 0.000E+00 97.71% 4.711E-02
BCLJAYA 97.37% 5.587E-02 98.62% 2.836E-02 93.06% 1.134E-01
BEJAYA 98.14% 2.597E-02 97.86% 5.843E-02 96.99% 5.380E-02
Methods hepatitis Ionosphere Sonar
Avg Std Avg Std Avg Std

BSCWJAYA 100.00% 0.000Eþ00 97.27% 4.480E-02 100.00% 0.000Eþ00

BJAYA 100.00% 0.000E+00 96.76% 4.073E-02 100.00% 0.000E+00
BGWO 100.00% 0.000E+00 96.03% 5.947E-02 99.17% 3.727E-02
BHHO 95.00% 2.236E-01 97.12% 4.036E-02 99.17% 3.727E-02
BMFO 100.00% 0.000E+00 95.55% 4.835E-02 100.00% 0.000E+00
BPSO 100.00% 0.000E+00 96.88% 5.200E-02 100.00% 0.000E+00
BSMA 95.00% 2.236E-01 95.96% 4.842E-02 100.00% 0.000E+00
BSSA 90.00% 3.078E-01 96.05% 5.306E-02 100.00% 0.000E+00
BCLJAYA 95.00% 2.236E-01 95.99% 4.814E-02 99.17% 3.727E-02
BEJAYA 90.00% 3.078E-01 96.31% 4.873E-02 97.02% 1.010E-01
Methods Vote wdbc Wielaw
Avg Std Avg Std Avg Std

BSCWJAYA 98.57% 4.397E-02 100.00% 0.000Eþ00 92.48% 1.125E-01

BJAYA 98.57% 4.397E-02 100.00% 0.000Eþ00 88.30% 1.149E-01
BGWO 97.14% 5.863E-02 100.00% 0.000Eþ00 88.71% 1.242E-01
BHHO 96.19% 6.796E-02 100.00% 0.000Eþ00 88.75% 9.169E-02
BMFO 97.14% 7.474E-02 100.00% 0.000Eþ00 91.23% 1.110E-01
BPSO 98.57% 4.397E-02 100.00% 0.000Eþ00 90.91% 1.234E-01
BSMA 98.45% 4.779E-02 100.00% 0.000Eþ00 91.06% 1.059E-01
BSSA 97.57% 6.027E-02 100.00% 0.000Eþ00 88.68% 1.006E-01
BCLJAYA 95.12% 1.210E-01 100.00% 0.000Eþ00 88.07% 9.401E-02
BEJAYA 96.07% 1.170E-01 100.00% 0.000Eþ00 92.27% 1.019E-01

31
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 18
MCC value on the UCI datasets and the Wielaw dataset.
Methods Breast Congress Heart

Avg Std Avg Std Avg Std

BSCWJAYA 0.9853 3.018E-02 0.9819 3.727E-02 0.9669 7.157E-02

BJAYA 0.9817 3.253E-02 0.9721 4.380E-02 0.9363 7.233E-02
BGWO 0.9816 3.265E-02 0.9646 6.011E-02 0.9355 8.885E-02
BHHO 0.9817 3.254E-02 0.9639 6.076E-02 0.9368 7.173E-02
BMFO 0.9820 3.935E-02 0.9584 6.972E-02 0.9357 8.696E-02
BPSO 0.9781 4.805E-02 0.9635 6.252E-02 0.9432 7.145E-02
BSMA 0.9781 4.142E-02 0.9775 3.994E-02 0.9335 1.037E-01
BSSA 0.9814 3.307E-02 0.9769 4.112E-02 0.9370 7.154E-02
BCLJAYA 0.9605 8.685E-02 0.9586 4.700E-02 0.8772 2.355E-01
BEJAYA 0.9743 3.599E-02 0.9288 1.398E-01 0.9358 8.862E-02
Methods hepatitis Ionosphere Sonar
Avg Std Avg Std Avg Std

BSCWJAYA 0.9655 1.063E-01 0.9578 7.106E-02 1.0000 0.000Eþ00

BJAYA 0.9489 1.249E-01 0.9513 6.131E-02 1.0000 0.000Eþ00
BGWO 0.9309 1.417E-01 0.9400 9.150E-02 0.9825 5.396E-02
BHHO 0.8643 2.533E-01 0.9511 7.370E-02 0.9822 5.473E-02
BMFO 0.9313 1.410E-01 0.9324 7.471E-02 0.9828 5.292E-02
BPSO 0.9309 1.417E-01 0.9509 8.373E-02 0.9815 5.712E-02
BSMA 0.9155 2.401E-01 0.9387 7.482E-02 0.9914 3.844E-02
BSSA 0.8658 3.140E-01 0.9404 7.953E-02 0.9827 7.722E-02
BCLJAYA 0.8640 2.536E-01 0.9389 7.468E-02 0.9917 3.727E-02
BEJAYA 0.8827 3.116E-01 0.9453 7.412E-02 0.9486 1.589E-01
Methods Vote wdbc Wielaw
Avg Std Avg Std Avg Std

BSCWJAYA 0.9873 3.913E-02 0.9737 3.687E-02 0.8302 1.406E-01

BJAYA 0.9806 4.743E-02 0.9670 4.479E-02 0.8154 1.922E-01
BGWO 0.9749 5.163E-02 0.9704 3.717E-02 0.7820 1.618E-01
BHHO 0.9675 5.783E-02 0.9670 4.395E-02 0.8092 1.379E-01
BMFO 0.9751 6.468E-02 0.9662 4.580E-02 0.8197 1.141E-01
BPSO 0.9873 3.913E-02 0.9671 4.953E-02 0.8206 1.508E-01
BSMA 0.9666 5.942E-02 0.9705 6.120E-02 0.7920 1.712E-01
BSSA 0.9720 7.681E-02 0.9701 3.764E-02 0.8128 1.285E-01
BCLJAYA 0.9438 1.439E-01 0.9672 3.726E-02 0.7726 1.311E-01
BEJAYA 0.9486 1.432E-01 0.9666 5.068E-02 0.8107 1.135E-01

32
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 19
F-measure value on the UCI datasets and the Wielaw dataset.
Methods Breast Congress Heart

Avg Std Avg Std Avg Std

BSCWJAYA 99.45% 1.125E-02 99.24% 1.552E-02 98.36% 3.729E-02

BJAYA 99.32% 1.215E-02 98.84% 1.813E-02 96.89% 3.557E-02
BGWO 99.31% 1.229E-02 98.45% 2.633E-02 97.06% 4.088E-02
BHHO 99.32% 1.215E-02 98.47% 2.609E-02 97.16% 3.236E-02
BMFO 99.32% 1.475E-02 98.24% 3.005E-02 97.07% 3.911E-02
BPSO 99.20% 1.737E-02 98.49% 2.554E-02 97.45% 3.213E-02
BSMA 99.20% 1.522E-02 99.06% 1.679E-02 97.02% 4.721E-02
BSSA 99.31% 1.229E-02 99.02% 1.752E-02 96.96% 3.481E-02
BCLJAYA 98.59% 3.075E-02 98.31% 1.917E-02 94.90% 9.096E-02
BEJAYA 99.05% 1.334E-02 97.13% 5.413E-02 97.02% 4.121E-02
Methods hepatitis Ionosphere Sonar
Avg Std Avg Std Avg Std

BSCWJAYA 96.67% 1.026E-01 98.56% 2.369E-02 100.00% 0.000Eþ00

BJAYA 95.00% 1.221E-01 98.31% 2.123E-02 100.00% 0.000Eþ00
BGWO 93.33% 1.368E-01 97.88% 3.196E-02 98.99% 3.126E-02
BHHO 86.67% 2.513E-01 98.30% 2.533E-02 98.99% 3.126E-02
BMFO 93.33% 1.368E-01 97.66% 2.559E-02 98.89% 3.420E-02
BPSO 93.33% 1.368E-01 98.35% 2.778E-02 98.73% 3.942E-02
BSMA 91.67% 2.388E-01 97.88% 2.564E-02 99.44% 2.485E-02
BSSA 86.67% 3.134E-01 97.91% 2.828E-02 98.75% 5.590E-02
BCLJAYA 86.67% 2.513E-01 97.90% 2.549E-02 99.55% 2.033E-02
BEJAYA 88.33% 3.111E-01 98.06% 2.582E-02 97.32% 7.860E-02
Methods Vote wdbc Wielaw
Avg Std Avg Std Avg Std

BSCWJAYA 99.23% 2.368E-02 98.21% 2.509E-02 90.13% 8.140E-02

BJAYA 98.78% 3.000E-02 97.75% 3.131E-02 89.24% 1.220E-01
BGWO 98.46% 3.157E-02 98.02% 2.493E-02 87.14% 9.958E-02
BHHO 97.94% 3.683E-02 97.76% 3.020E-02 89.22% 8.225E-02
BMFO 98.40% 4.260E-02 97.68% 3.199E-02 89.65% 6.310E-02
BPSO 99.23% 2.368E-02 97.73% 3.447E-02 89.97% 8.158E-02
BSMA 97.80% 3.925E-02 97.88% 4.556E-02 87.68% 1.053E-01
BSSA 98.23% 4.898E-02 98.00% 2.525E-02 89.64% 6.981E-02
BCLJAYA 96.66% 8.296E-02 97.83% 2.461E-02 87.19% 7.349E-02
BEJAYA 96.75% 8.915E-02 97.68% 3.585E-02 88.45% 7.864E-02

Fig. 16. Comparative experiments with other medical datasets.

settings of these four wrapper FS methods are shown in Table 21. significant improvement in classification accuracy, and we can
Fig. 17 shows the mean-error histograms of the four BSCWJAYA compensate for this deficiency by using parallel computing techniques
wrapper FS-based methods. We can easily conclude that the perfor­ or increasing the computing power of computing devices. Fig. 18 shows
mance of BSCWJAYA_KELM hinges on the other methods. Therefore, the twenty runs of the boxplot for ten algorithms. The maximum, me­
this paper uses the wrapper FS method based on BSCWJAYA and KELM dian, and minimum values in the figure show that the experimental
to predict IDH. results of BSCWJAYA_KELM are excellent and stable. The excellent
To verify that BSCWJAYA_KELM is highly competitive among similar classification results of BSCWJAYA_KELM are not obtained by a few
methods, BSCWJAYA_KELM is compared with nine similar methods. chance experiments but due to its stability and excellent classification
Table 22 shows the average results for the six evaluation criteria. We can performance. Fig. 19 shows the convergence trends of the ten algorithms
see that the accuracy, specificity, precision, MCC, and F-measure of (the fitness function in subsection 3.4.3 is used as the objective func­
BSCWJAYA_KELM are 94.27%, 96.13%, 93.83%, 0.8796, and 92.51%, tion). The search ability of ten algorithms for optimal feature subsets is
respectively. However, the time consumption of BSCWJAYA_KELM is mainly evaluated in terms of convergence speed and convergence ac­
relatively large. Certain time consumption is inevitable along with the curacy. BSCWJAYA KELM outperforms the other nine models in terms of

33
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Table 20
Description of 18 attributes.
No. Feature Detailed description

A1 Dialysis vintage (month) Non-IDH group (Median, IQR) = 69, 81

IDH group (Median, IQR) = 69, 91
A2 Diabetes mellitus (DM) No = 0; Yes = 1
A3 Ultrafiltration volume (kg) Non-IDH group (Median, IQR) = 2.0, 1.2
IDH group (Median, IQR) = 2.3, 1.1
A4 Age (years old) Non-IDH group (Median, IQR) = 66, 21
IDH group (Median, IQR) = 61, 19
A5 Post-dialysis weight of last session (kg) Non-IDH group (Median, IQR) = 54.2, 14.0
IDH group (Median, IQR) = 57.2, 14.2
A6 Predialysis weight (kg) Non-IDH group (Median, IQR) = 56.1, 14.6
IDH group (Median, IQR) = 59.4, 19.4
A7 Interdialytic weight gain (kg) Non-IDH group (Median, IQR) = 2.0, 1.1
IDH group (Median, IQR) = 2.2, 1.2
A8 Percentage of interdialytic weight gain (%) Non-IDH group (Median, IQR) = 3.71, 2.10
IDH group (Median, IQR) = 3.86, 1.76
A9 Systolic pressure (mmHg) Non-IDH group (Median, IQR) = 134, 27
IDH group (Median, IQR) = 151, 31
A10 Diastolic pressure (mmHg) Non-IDH group (Median, IQR) = 72, 16
IDH group (Median, IQR) = 81, 18
A11 Mean arterial pressure (mmHg) Non-IDH group (Median, IQR) = 93.00, 17.67
IDH group (Median, IQR) = 104.67, 20.33
A12 Heart rates (bpm) Non-IDH group (Median, IQR) = 76, 15
IDH group (Median, IQR) = 80, 17
A13 Gender Male = 1; Female = 0
A14 Albumin (g/L) Non-IDH group (Median, IQR) = 39.2, 5.0
IDH group (Median, IQR) = 38.6, 4.8
A15 Serum uric acid (μmol/L) Non-IDH group (Median, IQR) = 391, 90
IDH group (Median, IQR) = 417, 99
A16 Total cholesterol (mmol/L) Non-IDH group (Median, IQR) = 4.59, 1.35
IDH group (Median, IQR) = 4.94, 2.08
A17 Triglyceride (mmol/L) Non-IDH group (Median, IQR) = 1.80, 1.55
IDH group (Median, IQR) = 2.12, 1.42
A18 HDL-c (mmol/L) Non-IDH group (Median, IQR) = 1.01, 0.43
IDH group (Median, IQR) = 0.98, 0.40
A19 LDL-c (mmol/L) Non-IDH group (Median, IQR) = 2.22, 0.97
IDH group (Median, IQR) = 2.38, 1.29

IDH: intradialytic hypotension; IQR: interquartile range; kg: kilogram; HDL-c: high-density lipoprotein cholesterol; LDL-c: low-density lipoprotein cholesterol.

convergence speed and accuracy, which suggests that it can locate the dialysis vintage (A1) is selected 93 times, age (A4) is selected 72 times,
best feature subset more quickly and precisely, hence boosting the diastolic pressure (A10) is selected 51 times, and albumin (A14) is
classification performance. selected 48 times. These features are not negligible in the prediction of
In this subsection, BSCWJAYA_KELM is also compared with some IDH.
well-known classifiers. Fig. 20 shows that BSCWJAYA_KELM performs The features selected by classifiers to separate the albumin-decreased
the best among the six classification methods, followed by the back­ group from the albumin-nondeclined group are shown in Table 23.
propagation (BP) neural network. Extreme learning machine (ELM) is
unsatisfactory in predicting IDH, which indicates that the classification 6. Discussion
models based on BSCWJAYA and KELM can compensate for the short­
comings of a single classifier in classification and achieve more accurate 6.1. Summary and comparison with previous studies
classification results.
From the above experimental analysis, we can conclude that IDH is an acute complication of HD, while malnutrition is a chronic
BSCWJAYA_KELM can obtain a feature subset with better results for the complication. To prevent IDH episodes, previous studies have developed
HD dataset. To determine whether the selected optimal feature subset is various models to predict IDH. Because the pathogenesis of IDH is linked
significant for medical diagnosis, we expanded the number of runs to to vessels, volume and heart function, most models analyze predialysis
100. The importance of each feature for clinical diagnosis was verified B_P, IDWG [100], heart rates, and echocardiographic parameters [101].
by the number of occurrences of each feature. Fig. 21 shows the number Lin et al. [102] proposed a prediction model based on time-dependent
of occurrences of all features. The important features affecting IDH, in logistic regression analysis and the least absolute shrinkage and selec­
order of importance, are serum uric acid, dialysis vintage, age, diastolic tion operator and applied the method to IDH. However, the prediction
pressure, and albumin, where serum uric acid (A15) is selected 98 times, results for IDH cannot meet the current requirements for the IDH
auxiliary prediction model. Passos et al. [103] used logistic regression
with both categorical and continuous independent variables to predict
Table 21 IDH. However, the performance of logistic regression is suboptimal
Parameter settings for the four methods. when the feature space is large, which leads to an inability of the model
Methods Parameter values to handle larger and more complex HD datasets. Although there are
similar methods for HD management, the method proposed in this paper
BSCWJAYA_KNN K=1
BSCWJAYA_FKNN K = 1, m = 2 differs from them. For example, Hu et al. [98] predicted the nutritional
BSCWJAYA_SVM C = 850, γ = 0.17 status of HD patients. However, the classifier of this method is the
BSCWJAYA_KELM M=1 k-nearest neighbor. We also proposed an IDH prediction model based on
an improved GWO with KELM [104]. In this study, we improved the

34
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 17. Results of four SCWJAYA-based classifiers.

malnutrition comprehensively, they contain subjective scoring items,

and some items need to be observed over months. To develop a more
Table 22 accessible and precise prediction model, a FS method utilizing serum
Average results of ten methods in six metrics. nutrition biomarkers is conducted in the present study.
Methods Accuracy Specificity Precision

Value Rank Value Rank Value Rank 6.2. Findings

BSCWJAYA 94.27% 1 96.13% 1 93.83% 1
BJAYA 93.65% 6 95.04% 7 92.25% 7 Based on the discussion of the experimental results of similar
BGWO 93.61% 7 95.21% 5 92.44% 5 methods, it can be concluded that optimization algorithms with
BHHO 93.56% 8 94.87% 9 92.01% 8
powerful search capabilities and excellent classifiers are still important
BMFO 93.97% 2 95.63% 3 93.09% 3
BPSO 93.82% 4 94.79% 10 92.01% 9
for our future research. Khurana et al. [108] proposed an integrated
BSMA 93.76% 5 95.88% 2 93.37% 2 classifier model based on biogeography-based optimization. The effect
BSSA 93.45% 9 95.20% 6 92.44% 6 of the optimization algorithm on the accuracy of the classification model
BCLJAYA 93.87% 3 95.46% 4 92.85% 4 was demonstrated by comparing the integrated classifier model with six
BEJAYA 92.37% 10 94.87% 8 90.84% 10
individual classification models. The experimental results showed that
Methods MCC F-measure Time cost
Value Rank Value Rank Value Rank the classification model optimized by the optimization algorithm could
achieve the best text classification and improve the accuracy. Hu et al.
BSCWJAYA 0.8796 1 92.51% 1 6622 10
BJAYA 0.8672 6 91.77% 6 1658 7
[98] verified that the optimization algorithm combined with different
BGWO 0.8662 7 91.61% 8 2809 9 classifiers had different accuracies and stabilities of classification. Silva
BHHO 0.8653 8 91.69% 7 852 1 et al. [109] proposed two asynchronous parallel strategies to improve
BMFO 0.8738 2 92.12% 2 1701 8 the execution efficiency of FS models. As a result, we have suggested a
BPSO 0.8709 4 92.06% 3 1642 6
model that is both robust and efficient, namely, a wrapper FS method
BSMA 0.8688 5 91.80% 5 1628 5
BSSA 0.8626 9 91.49% 9 1610 3 based on a metaheuristic algorithm and KELM. The method does not
BCLJAYA 0.8715 3 92.02% 4 1610 2 require training data and can still show advantages in dealing with
BEJAYA 0.8367 10 89.19% 10 1617 4 larger-volume datasets, making it a simple and effective prediction
method. Second, our proposed prediction model combined with KELM,
an efficient classifier, helps to improve the performance of the predic­
JAYA algorithm by KELM. Compared with our previous study focusing
tion model. Third, the computational efficiency of SCWJAYA can be
on chronic kidney disease-mineral and bone disease, the present study
further improved by parallel techniques.
concentrated on the relationship between IDH and nutritional status.
In the present study, dialysis vintage, age, diastolic blood pressure
Combining the classification accuracy, specificity, and other indicators,
(DBP), serum UA and albumin levels are the high-frequency features
it can be seen that the classification accuracy proposed in this paper is
selected by the classifier.
higher than that in previous studies.
UA is the end product of purine metabolism and is eliminated mainly
Malnutrition is associated with orthostatic hypotension in the elderly
by the renal system. Purine originates from food and senile cell meta­
[105] and low B_P in patients with CKD [106]. Serum biomarkers of
bolism. UA is a standard nutritional marker for low protein intake in HD
nutrition, such as albumin [22] and cholesterol [107], are risk factors for
patients [110]. Jiang et al. found that low UA levels were characterized
IDH calculated by Pearson’s correlation or logistic regression. Chiang’s
by malnutrition and an increased risk of long-term mortality in main­
study [106], in which malnutrition was assessed using a scale, also
tenance HD patients [111]. On the one hand, hypouricemia reduces B_P
found that malnutrition is a risk factor for IDH. Although scales evaluate
by reducing heat-induced endothelium-dependent microvascular

35
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 18. Boxplot of the performance of ten methods in six metrics.

36
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

Fig. 19. Convergence evolution trends of ten methods.

Fig. 20. Comparison of BSCWJAYA_KELM with well-known classifiers.

vasodilation, renin-angiotensin system activity, and myeloperoxidase nutritional status and higher antioxidant capacity [115]. However, hy­
activity and increasing lipid peroxidation [112]. On the other hand, peruricemia is also harmful. It is closely related to other risk factors for
hypouricemia is related to malnutrition, which is associated with hy­ IDH, including electrocardiographic left ventricular hypertrophy [116],
potension in HD patients [106]. Although studies ascribe this phenom­ higher carotid-femoral pulse wave velocity (PWV) [117] and coronary
enon to malnutrition, what cannot be ignored is the maintenance of artery disease [118]. UA affects vascular stiffness and B_P, possibly by
peripheral vascular resistance by UA. Baroreceptor sensitivity and inducing endothelial cell injury and stimulating vascular smooth muscle
sympathetic activity are essential factors in maintaining peripheral cell proliferation [119]. Damaged vascular resistance results in hyper­
resistance. A statistically significant correlation can be appreciated be­ tension, B_P fluctuation and increased pulse pressure in the general
tween parameters of UA level and sympathetic domain, such as heart population, while its clinical manifestation is predialysis hypertension
rate variability [113]. The heart rate turbulence slope is one of the pa­ and hypotension during ultrafiltration in HD patients. Impaired cardiac
rameters of heart rate turbulence, which is modulated by the barore­ function is due to coronary artery stenosis and cardiomyocyte injury. UA
ceptor reflex. There is a significant negative correlation between leads to cardiomyocyte hypertrophy independent of ischemic injury
turbulence slope and serum UA levels [114]. [120]. Moreover, UA is associated with metabolic syndrome and insulin
HD patients with hyperuricemia consistently have a better resistance, which also affect B_P and cardiac function.

37
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

peripheral resistance. Even if patients with IDH have an increased pe­

ripheral resistance predialysis, they may not be able to mount an in­
crease in their peripheral resistance to compensate for further blood
volume ultrafiltration to maintain their B_P during HD sessions.
Owing to the excellent search capability of the proposed SCWJAYA
method, the method may also be used to tackle other problems, such as
identifying molecular markers for cancer diagnosis [127–129], disease
module identification [130,131], Kayak cycle phase segmentation [132]
and location-based services [133,134]. We can also combine the pro­
posed method with machine learning methods to tackle other practical
problems, such as urban road planning [135], molecular signature
identification [136,137], image augmentation [138,139], pharma­
coinformatic data mining [140,141], drug discovery [142–144], essay
recommendation [145], corn leaf disease diagnosis [146], fault detec­
tion [147], and human motion capture [148].
Fig. 21. Number of times each feature was selected for one hundred
experiments. 6.3. Limitations

Serum albumin is also a marker of malnutrition, and intravenous To our knowledge, our study is the first to illuminate the relationship
supplementation with albumin is an acute management of IDH. Naka­ of serum UA and other nutritional biomarkers with IDH. There are still
moto et al. conducted a small sample retrospective study showing that some limitations in the present study. First, other indices of nutritional
hypoalbuminemia is a risk factor for IDH [22]. Albumin is responsible status were not included, such as triceps skinfold thickness, upper arm
for maintaining a stable plasma colloid osmotic pressure (generating a circumference and body component, which require observation for 2–3
proportion of approximately 70%–80%) and shifting water from tissue months. Second, other serum nutritional biomarkers were not screened,
into regional venous beds [121]. In the HD sessions, plasma refill for example, hemoglobin, prealbumin and total iron binding capacity
compensates for volume ultrafiltration to avoid IDH. (TIBC). Hemoglobin, transferrin and TIBC are affected not only by
Older age is a risk factor for IDH [122]. In the present study, age is nutritional status but also by bleeding, inflammation, and the activity of
also a high-frequency feature selected by the classifier. The mean age of hypoxia-inducible factor 1, among others. The half-life of prealbumin is
the IDH group is younger than that of the non-IDH group. Because the only 2 days, which is much lower than that of albumin. Although pre­
key benefit of the approach we utilize is that it better covers the search albumin is also a serum biomarker of nutrition, its use to predict a one-
space, the wrapper method has superior classification accuracy and month IDH risk is inappropriate. Third, IDH was divided into two
computational efficiency, as well as greater flexibility in FS. In addition, groups, not according to the severity. Fourth, KELM utilizes matrix op­
it employs a metaheuristic search approach to discover the ideal feature erations in the prediction and classification process, which undoubtedly
subset and evaluates the selected feature subset using the classification increases the time complexity of the model. Fifth, although the sine
performance of machine learning algorithms, which compensates for the chaos initialization strategy improves the convergence speed in the early
drawbacks of existing mining techniques. stage, some shortcomings persist compared with other methods.
Patients with long HD vintages are more prone to IDH episodes. The
possible reasons are left ventricular hypertrophy, cardiac function with 7. Conclusions and future works
systolic and/or diastolic dysfunction and decompensation of peripheral
resistance. Dialysis-related amyloidosis is characterized by β2-micro­ This study is a validated IDH prediction method that utilizes the
globulin amyloid assembly with HD vintage [123]. Amyloid may be clinical dataset provided by the HD Center of the First Affiliated Hospital
accumulated in the different vital parts including the heart, nerves and of Wenzhou Medical University. The method enables simpler and more
blood vessels, instigating diverse clinical conditions, such as the car­ accurate prediction of IDH. The key innovation of the model is to use the
diomyopathy, autonomic dysfunction, chronic hypotension and IDH. In optimization capability of SCWJAYA to obtain the optimal feature
addition, the severity of valve calcification and vascular calcification subset and then combine it with KELM to achieve IDH prediction. By
(VC) are associated with HD vintage [124]. Cardiac valve calcification introducing sine chaos initialization, crisscross search and wormhole
affects the left ventricular mass and ejection fraction and induces heart search, the convergence speed and convergence accuracy of the original
failure [125]. VC not only induces coronary artery stenosis and accel­ JAYA are improved, and the probability of the algorithm falling into
erates myocardial ischemia but also increases peripheral arterial local optimum is reduced. The superiority of SCWJAYA’s performance is
stiffness. verified by a series of experiments on 35 benchmark functions. Subse­
DBP is regulated by peripheral vascular resistance. As mentioned quently, BSCWJAYA_KELM is analyzed by FS experiments with some
above, peripheral resistance is a compensatory mechanism of B_P during traditional prediction methods and common similar prediction methods.
dialysis. It is regulated by baroreceptor sensitivity, sympathetic over­ The experimental results show that BSCWJAYA_KELM is a more accu­
activity and vascular stiffness. Park et al. [126] illustrated a U-shaped rate and stable prediction model. Modification and upgradation of the
association of predialysis systolic pressure and mortality, while the precision and efficiency of this prediction model will be our future focus.
targets of DBP have not been defined in previous studies. The normal Conversely, extension of the proposed SCWJAYA to the multiobjective
range of DBP should be individualized. Abnormal DBP reveals impaired version is another potential future area of study [149–151]. We intend to

Table 23
Serum nutritional indices and dialysis parameters in the IDH and non-IDH groups.
Index Non-IDH (n = 1189) IDH (n = 751) p value

Serum uric acid (Median, IQR, μmol/L) 391, 90 417, 99 <0.001

Albumin (Median, IQR, g/L) 39.2, 5.0 38.6, 4.8 0.107
Age (Median, IQR, years old) 66, 21 61, 19 <0.001
Diastolic pressure (Median, IQR, mmHg) 72, 16 81, 18 <0.001
Dialysis vintage (Median, IQR, month) 69, 81 69, 91 0.597

IDH: intradialytic hypotension; IQR: interquartile range.

38
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

develop a model to provide a proper ultrafiltration volume prescription [23] D. Fouque, et al., A proposed nomenclature and diagnostic criteria for protein-
energy wasting in acute and chronic kidney disease, Kidney Int. 73 (4) (2008)
to help HD patients achieve optimum B_P and avoid volume overload.
391–398.
[24] E. Macedo, et al., A randomized trial of albumin infusion to prevent intradialytic
hypotension in hospitalized hypoalbuminemic patients, Crit. Care 25 (1) (2021)
Declaration of competing interest 18.
[25] I. Beberashvili, et al., Longitudinal study of serum uric acid, nutritional status,
and mortality in maintenance hemodialysis patients, Clin. J. Am. Soc. Nephrol. 11
The authors declare that there is no conflict of interests regarding the (6) (2016) 1015–1023.
publication of article. [26] J.J. Liang, B.Y. Qu, P.N. Suganthan, Problem Definitions and Evaluation Criteria
for the CEC 2014 Special Session and Competition on Single Objective Real-
Parameter Numerical Optimization, 2013.
Acknowledgments [27] G. Wu, R. Mallipeddi, P. Suganthan, Problem Definitions and Evaluation Criteria
for the CEC 2017 Competition and Special Session on Constrained Single
Objective Real-Parameter Optimization, 2016.
This work was supported in part by the Natural Science Foundation
[28] K. Price, et al., Problem definitions and evaluation criteria for the 100-digit
of Zhejiang Province (LZ22F020005, LQ21H050008) and the National challenge special session and competition on single objective numerical
Natural Science Foundation of China (62076185 and U1809209). It was optimization, in: Technical Report, Nanyang Technological University, 2018.
[29] S. Garcia, et al., Advanced nonparametric tests for multiple comparisons in the
also supported by the New Technologies and Products Projects of Zhe­
design of experiments in computational intelligence and data mining:
jiang Health Committee (2021PY054) and the Basic Scientific Research experimental analysis of power, Inf. Sci. 180 (10) (2010) 2044–2064.
Projects of Wenzhou Science and Technology Bureau (Y2020026). We [30] J.D. a, et al., A practical tutorial on the use of nonparametric statistical tests as a
acknowledge the comments of the reviewers. methodology for comparing evolutionary and swarm intelligence algorithms,
Swarm and Evolutionary Computation 1 (1) (2011) 3–18.
[31] H.Q. Lyu, et al., A filter feature selection method based on the Maximal
References Information Coefficient and Gram-Schmidt Orthogonalization for biomedical data
mining, Comput. Biol. Med. 89 (2017) 264–274.
[32] Y.L. Xiong, et al., Informative gene selection based on cost-sensitive fast
[1] A. Liew, Perspectives in renal replacement therapy: Haemodialysis, Nephrology
correlation-based filter feature selection, Curr. Bioinf. 16 (8) (2021) 1060–1068.
23 (Suppl 4) (2018) 95–99.
[33] H.L. Li, et al., A hybrid feature selection algorithm based on a discrete artificial
[2] C.W. McIntyre, Effects of hemodialysis on cardiac function, Kidney Int. 76 (4)
bee colony for Parkinson’s diagnosis, ACM Trans. Internet Technol. 21 (3) (2021).
(2009) 371–375.
[34] J. Hu, et al., Detection of COVID-19 severity using blood gas analysis parameters
[3] T. Kitano, et al., Changes in tissue oxygenation in response to sudden intradialytic
and Harris hawks optimized extreme learning machine, Comput. Biol. Med. 142
hypotension, J. Artif. Organs 23 (2) (2020) 187–190.
(2022), 105166.
[4] B.V. Stefánsson, et al., Intradialytic hypotension and risk of cardiovascular
[35] I. Kamkar, et al., Stable feature selection for clinical prediction: exploiting ICD
disease, Clin. J. Am. Soc. Nephrol. 9 (12) (2014) 2124–2132.
tree structure using Tree-Lasso, J. Biomed. Inf. 53 (2015) 277–290.
[5] R. Hekmat, et al., Correlation between asymptomatic intradialytic hypotension
[36] S.I. Niwas, et al., Cross-examination for angle-closure glaucoma feature detection,
and regional left ventricular dysfunction in hemodialysis patients, Iran J Kidney
Ieee Journal of Biomedical and Health Informatics 20 (1) (2016) 343–354.
Dis 5 (2) (2011) 97–102.
[37] H. Yu, et al., Apple Leaf Disease Recognition Method with Improved Residual
[6] C. MacEwen, et al., Relationship between hypotension and cerebral ischemia
Network, Multimedia Tools and Applications, 2022.
during hemodialysis, J. Am. Soc. Nephrol. 28 (8) (2017) 2511–2520.
[38] J. Xia, et al., Performance optimization of support vector machine with
[7] T.I. Chang, et al., Intradialytic hypotension and vascular access thrombosis,
oppositional grasshopper optimization for acute appendicitis diagnosis, Comput.
J. Am. Soc. Nephrol. 22 (8) (2011) 1526–1533.
Biol. Med. (2022), 105206.
[8] M.A. Jansen, et al., Predictors of the rate of decline of residual renal function in
[39] J. Xia, et al., Evolving kernel extreme learning machine for medical diagnosis via
incident dialysis patients, Kidney Int. 62 (3) (2002) 1046–1053.
a disperse foraging sine cosine algorithm, Comput. Biol. Med. 141 (2022),
[9] E.Y. Seong, et al., The relationship between intradialytic hypotension and
105137.
hospitalized mesenteric ischemia: a case-control study, Clin. J. Am. Soc. Nephrol.
[40] R. Dong, et al., Boosted kernel search: framework, analysis and case studies on the
13 (10) (2018) 1517–1525.
economic emission dispatch problem, Knowl. Base Syst. 233 (2021), 107529.
[10] R. Matsuura, et al., Intradialytic hypotension is an important risk factor for
[41] B. Shi, et al., Evolutionary warning system for COVID-19 severity: colony
critical limb ischemia in patients on hemodialysis, BMC Nephrol. 20 (1) (2019)
predation algorithm enhanced extreme learning machine, Comput. Biol. Med.
473.
136 (2021), 104698.
[11] J.A. Chou, et al., Intradialytic hypotension, blood pressure changes and mortality
[42] S. Wu, et al., Evolving fuzzy k-nearest neighbors using an enhanced sine cosine
risk in incident hemodialysis patients, Nephrol. Dial. Transplant. 33 (1) (2018)
algorithm: case study of lupus nephritis, Comput. Biol. Med. 135 (2021), 104582.
149–159.
[43] Y. Sun, et al., Evolving deep convolutional neural networks for image
[12] E.M. Ettema, et al., Effect of plasma sodium concentration on blood pressure
classification, IEEE Trans. Evol. Comput. 24 (2) (2019) 394–407.
regulators during hemodialysis: a randomized crossover study, BMC Nephrol. 19
[44] S.-H. Wu, Z.-H. Zhan, J. Zhang, SAFE: scale-adaptive fitness evaluation method
(1) (2018) 214.
for expensive optimization problems, IEEE Trans. Evol. Comput. 25 (3) (2021)
[13] L. Gabutti, et al., Haemodynamic consequences of changing bicarbonate and
478–491.
calcium concentrations in haemodialysis fluids, Nephrol. Dial. Transplant. 24 (3)
[45] J.-Y. Li, et al., Boosting data-driven evolutionary algorithm with localized data
(2009) 973–981.
generation, IEEE Trans. Evol. Comput. 24 (5) (2020) 923–937.
[14] A. Ayoub, M. Finlayson, Effect of cool temperature dialysate on the quality and
[46] J.-H. Yi, et al., An improved NSGA-III algorithm with adaptive mutation operator
patients’ perception of haemodialysis, Nephrol. Dial. Transplant. 19 (1) (2004)
for Big Data optimization problems, Future Generat. Comput. Syst. 88 (2018)
190–194.
571–585.
[15] I. Dasgupta, et al., Associations between hemodialysis facility practices to manage
[47] F. Zhao, et al., A novel cooperative multi-stage hyper-heuristic for combination
fluid volume and intradialytic hypotension and patient outcomes, Clin. J. Am.
optimization problems, Complex System Modeling and Simulation 1 (2) (2021)
Soc. Nephrol. 14 (3) (2019) 385–393.
91–108.
[16] M.R. Marshall, et al., Effect of low-sodium versus conventional sodium dialysate
[48] K. Socha, Christian Blum, An ant colony optimization algorithm for continuous
on left ventricular mass in home and self-care satellite facility hemodialysis
optimization: application to feed-forward neural network training, Neural
patients: a randomized clinical trial, J. Am. Soc. Nephrol. 31 (5) (2020)
Computing and Applications 16 (3) (2007) 235–247.
1078–1091.
[49] J. Kennedy, R. Eberhart, Particle swarm optimization, in: Icnn95-international
[17] D.L. Zimmerman, et al., Dialysate calcium concentration and mineral metabolism
Conference on Neural Networks, 2002.
in long and long-frequent hemodialysis: a systematic review and meta-analysis for
[50] S. Mirjalili, et al., Grey Wolf Optimizer. Advances in engineering software 69,
a Canadian Society of Nephrology clinical practice guideline, Am. J. Kidney Dis.
Elsevier, 2014, pp. 46–61.
62 (1) (2013) 97–111.
[51] X.S. Yang, A New Metaheuristic Bat-Inspired Algorithm, Nature inspired
[18] A.P. Guérin, et al., Arterial stiffening and vascular calcifications in end-stage renal
cooperative strategies for optimization (NICSO 2010) (2010) 65–74.
disease, Nephrol. Dial. Transplant. 15 (7) (2000) 1014–1021.
[52] Venkata Rao, R. Jaya, A Simple and New Optimization Algorithm for Solving
[19] S.M. Brunelli, et al., Facility dialysate calcium practices and clinical outcomes
Constrained and Unconstrained Optimization Problems, vol. 7, 2016, pp. 19–34.
among patients receiving hemodialysis: a retrospective observational study, Am.
[53] R. Storn, Differential Evolution-A Simple and Efficient Heuristic for Global
J. Kidney Dis. 66 (4) (2015) 655–665.
Optimization over Continuous Space, 1997, p. 11.
[20] K.S. Gray, D.E. Cohen, S.M. Brunelli, Dialysate temperature of 36 ◦ C: association
[54] Ali Asghar Heidari, et al., Harris hawks optimization: Algorithm and applications
with clinical outcomes, J. Nephrol. 31 (1) (2018) 129–136.
97 (2019) 849–872.
[21] J.C. Huang, et al., Predictive modeling of blood pressure during hemodialysis: a
[55] B. Dogan, T. Olmez, A new metaheuristic for numerical function optimization:
comparison of linear model, random forest, support vector regression, XGBoost,
Vortex Search algorithm, Information Sciences 293 (2015) 125–145.
LASSO regression and ensemble method, Comput. Methods Progr. Biomed. 195
[56] I. Ahmadianfar, et al., INFO: an efficient optimization algorithm based on
(2020), 105536.
weighted mean of vectors, Expert Syst. Appl. (2022), 116516.
[22] H. Nakamoto, et al., Hypoalbuminemia is an important risk factor of hypotension
during hemodialysis, Hemodial. Int. 10 (Suppl 2) (2006) S10–S15.

39
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

[57] S. Li, et al., Slime mould algorithm: a new method for stochastic optimization, [90] J. Hu, et al., Orthogonal learning covariance matrix for defects of grey wolf
Future Generat. Comput. Syst. 111 (2020) 300–323. optimizer: insights, balance, diversity, and feature selection, Knowl. Base Syst.
[58] J. Tu, et al., The colony predation algorithm, Journal of Bionic Engineering 18 (3) 213 (2021), 106684.
(2021) 674–710. [91] J. Hu, et al., Dispersed foraging slime mould algorithm: continuous and binary
[59] I. Ahmadianfar, et al., RUN beyond the Metaphor: an Efficient Optimization variants for global optimization and wrapper-based feature selection, Knowl. Base
Algorithm Based on Runge Kutta Method, Expert Systems with Applications, Syst. 237 (2022), 107761.
2021, 115079. [92] J. Too, G. Liang, H. Chen, Memory-based Harris hawk optimization with learning
[60] Y. Yang, et al., Hunger games search: visions, conception, implementation, deep agents: a feature selection approach, Eng. Comput. (2021).
analysis, perspectives, and towards performance shifts, Expert Syst. Appl. 177 [93] J. Hu, et al., Chaotic diffusion-limited aggregation enhanced grey wolf optimizer:
(2021), 114864. insights, analysis, binarization, and feature selection, Int. J. Intell. Syst. 1 (64)
[61] S.M. Guo, et al., A self-optimization approach for L-SHADE incorporated with (2021).
eigenvector-based crossover and successful-parent-selecting framework on CEC [94] Kiran, M. Servet Tsa, Tree-seed algorithm for continuous optimization 42 (19)
2015 benchmark set, in: Evolutionary Computation, 2015. (2015) 6686–6698.
[62] J. Yong, et al., A novel bat algorithm based on collaborative and dynamic [95] W. Pietruszkiewicz, Dynamical systems and nonlinear Kalman filtering applied in
learning of opposite population, in: 2018 IEEE 22nd International Conference on classification, in: 2008 7th IEEE International Conference on Cybernetic
Computer Supported Cooperative Work in Design, CSCWD), 2018. Intelligent Systems, IEEE, 2008.
[63] Z. Cai, et al., Evolving an optimal kernel extreme learning machine by using an [96] Y. Zhang, et al., Towards augmented kernel extreme learning models for
enhanced grey wolf optimization strategy, Expert Systems with Applications 138 bankruptcy prediction: algorithmic behavior and comprehensive analysis,
(2019), 112814. Neurocomputing 430 (2021) 185–212.
[64] N. Lynn, P. Suganthan, Ensemble particle swarm optimizer, Applied Soft [97] X. Yang, et al., An optimized machine learning framework for predicting
Computing 55 (2017) 533–548. intradialytic hypotension using indexes of chronic kidney disease-mineral and
[65] L. Liu, et al., Ant colony optimization with Cauchy and greedy Levy mutations for bone disorders, Comput. Biol. Med. 145 (2022), 105510.
multilevel COVID 19 X-ray image segmentation, Comput. Biol. Med. 136 (2021), [98] J. Hu, et al., An effective model for predicting serum albumin level in
104609. hemodialysis patients, Comput. Biol. Med. 140 (2022), 105054.
[66] L. Wang, C. Huang, A novel Elite Opposition-based Jaya algorithm for parameter [99] K.D. Workgroup, K/DOQI clinical practice guidelines for cardiovascular disease
estimation of photovoltaic cell models, Optik 155 (2018) 351–356. in dialysis patients, Am. J. Kidney Dis. 45 (4 Suppl 3) (2005) S1–S153.
[67] Zhang, Y.Y. and Z.G. Jin, Comprehensive learning Jaya algorithm for engineering [100] F. Deng, et al., The relationship between prescription of ultrafiltration and
design optimization problems. J. Intell. Manuf.. intradialytic hypotension in Chinese hemodialysis patients, Ann. Palliat. Med. 10
[68] R. Motamarri, N. Bhookya, JAYA algorithm based on Levy flight for global MPPT (5) (2021) 5316–5321.
under partial shading in photovoltaic system, Ieee Journal of Emerging and [101] H. Poorzand, et al., Echocardiographic parameters in patients with and without
Selected Topics in Power Electronics 9 (4) (2021) 4979–4991. hypotension during dialysis, J. Cardiovasc. Thorac. Res. 13 (3) (2021) 228–233.
[69] Y.Y. Zhang, A.N. Chi, S. Mirjalili, Enhanced Jaya Algorithm: A Simple but [102] C.-J. Lin, et al., Intelligent system to predict intradialytic hypotension in chronic
Efficient Optimization Method for Constrained Engineering Design Problems, hemodialysis, J. Formos. Med. Assoc. 117 (10) (2018) 888–893.
Knowledge-Based Systems, 2021, p. 233. [103] R. da Hora Passos, et al., Ultrasound-based clinical profiles for predicting the risk
[70] K.J. Yu, et al., Parameters identification of photovoltaic models using an of intradialytic hypotension in critically ill patients on intermittent dialysis: a
improved JAYA optimization algorithm, Energy Convers. Manag. 150 (2017) prospective observational study, Crit. Care 23 (1) (2019) 389.
742–753. [104] X. Yang, et al., An optimized machine learning framework for predicting
[71] G. Iacca, V.C. dos Santos, V.V. de Melo, An improved Jaya optimization algorithm intradialytic hypotension using indexes of chronic kidney disease-mineral and
with Levy flight, Expert Syst. Appl. (2021) 165. bone disorders, Comput. Biol. Med. (2022), 105510.
[72] D.H. Wolpert, W. Macready, No free lunch theorems for optimization, IEEE [105] S.E. Kocyigit, et al., Improvement of nutritional status enhances cognitive and
Transactions on Evolutionary Computation 1 (1) (1997) 67–82. physical functions in older adults with orthostatic hypotension, Nutrition 90
[73] X. Xu, C. Wang, P. Zhou, GVRP considered oil-gas recovery in refined oil (2021), 111261.
distribution: from an environmental perspective, Int. J. Prod. Econ. 235 (2021), [106] H.P. Chiang, et al., Blood pressure modifies outcomes in patients with stage 3 to 5
108078. chronic kidney disease, Kidney Int. 97 (2) (2020) 402–413.
[74] H. Zhang, et al., Differential evolution-assisted salp swarm algorithm with chaotic [107] M. Matsuo, et al., Hypocholesterolemia is a risk factor for reduced systemic
structure for real-world problems, Eng. Comput. (2022). vascular resistance reactivity during hemodialysis, Hypertens. Res. 44 (8) (2021)
[75] W. Shan, et al., Double adaptive weights for stabilization of moth flame 988–995.
optimizer: balance analysis, engineering cases, and medical diagnosis, Knowl. [108] A. Khurana, O.P. Verma, Novel approach with nature-inspired and ensemble
Base Syst. 214 (2021), 106728. techniques for optimal text classification, Multimed. Tool. Appl. 79 (33–34)
[76] J. Tu, et al., Evolutionary biogeography-based whale optimization methods with (2020) 23821–23848.
communication structure: towards measuring the balance, Knowl. Base Syst. 212 [109] J. Silva, A. Aguiar, F. Silva, Parallel asynchronous strategies for the execution of
(2021), 106642. feature selection algorithms, Int. J. Parallel Program. 46 (2) (2018) 252–283.
[77] Y. Zhang, et al., Towards augmented kernel extreme learning models for [110] C. Park, et al., Serum uric acid, protein intake and mortality in hemodialysis
bankruptcy prediction: algorithmic behavior and comprehensive analysis, patients, Nephrol. Dial. Transplant. 32 (10) (2017) 1750–1757.
Neurocomputing 430 (2021) 185–212. [111] M.Y. Jiang, et al., Clinical implications and outcome prediction in chronic
[78] Z. Cai, et al., Evolving an optimal kernel extreme learning machine by using an hemodialysis patients with lower serum potassium×uric acid product, Eur. J.
enhanced grey wolf optimization strategy, Expert Syst. Appl. 138 (2019), 112814. Intern. Med. 26 (8) (2015) 646–651.
[79] Y. Xu, et al., An efficient chaotic mutative moth-flame-inspired optimizer for [112] B. De Becker, et al., Severe hypouricemia impairs endothelium-dependent
global optimization tasks, Expert Syst. Appl. 129 (2019) 135–155. vasodilatation and reduces blood pressure in healthy young men: a randomized,
[80] G.-Q. Zeng, et al., Binary-coded extremal optimization for the design of PID placebo-controlled, and crossover study, J. Am. Heart Assoc. 8 (23) (2019),
controllers, Neurocomputing 138 (2014) 180–188. e013130.
[81] G.-Q. Zeng, et al., Design of fractional order PID controller for automatic [113] K.U. Kunikullaya, et al., Correlation of serum uric acid with heart rate variability
regulator voltage system based on multi-objective extremal optimization, in hypertension, Hipertens. Riesgo Vasc. 32 (4) (2015) 133–141.
Neurocomputing 160 (2015) 173–184. [114] K. Szymanowska, et al., Clinical significance of heart rate turbulence assessment
[82] G.-Q. Zeng, et al., Adaptive population extremal optimization-based PID neural in patients with chronic heart failure, Kardiol. Pol. 66 (12) (2008) 1289–1295.
network for multivariable nonlinear control systems, Swarm Evol. Comput. 44 [115] E. Domínguez-Zambrano, et al., Association between serum uric acid levels,
(2019) 320–334. nutritional and antioxidant status in patients on hemodialysis, Nutrients 12 (9)
[83] H. Yu, et al., Image segmentation of Leaf Spot Diseases on Maize using multi-stage (2020).
Cauchy-enabled grey wolf algorithm, Eng. Appl. Artif. Intell. 109 (2022), 104653. [116] C.W. Liu, et al., The dose-response effects of uric acid on the prevalence of
[84] Y. Chen, et al., Multi-threshold image segmentation using a multi-strategy metabolic syndrome and electrocardiographic left ventricular hypertrophy in
shuffled frog leaping algorithm, Expert Syst. Appl. (2022), 116511. healthy individuals, Nutr. Metabol. Cardiovasc. Dis. 29 (1) (2019) 30–38.
[85] C. Chen, et al., Multi-Threshold image segmentation of maize diseases based on [117] M. Kanbay, et al., Impact of uric acid on hypertension occurrence and target
elite comprehensive particle swarm optimization and otsu, Front. Plant Sci. 12 organ damage: insights from the STANISLAS cohort with a 20-year follow-up, Am.
(2021), 789911, https://doi.org/10.3389/fpls.2021.789911. J. Hypertens. 33 (9) (2020) 869–878.
[86] H. Bao, et al., Random replacement crisscross butterfly optimization algorithm for [118] F.S. Yen, et al., Urate-lowering therapy may prevent the development of coronary
standard evaluation of overseas Chinese associations, Electronics 11 (7) (2022) artery disease in patients with gout, Front. Med. 7 (2020) 63.
1080. [119] D.B. Corry, et al., Uric acid stimulates vascular smooth muscle cell proliferation
[87] H. Su, et al., Horizontal and vertical search artificial bee colony for image and oxidative stress via the vascular renin-angiotensin system, J. Hypertens. 26
segmentation of COVID-19 X-ray images, Comput. Biol. Med. 142 (2022), (2) (2008) 269–275.
105181. [120] X.J. Zhang, et al., Potential risk of hyperuricemia: leading cardiomyocyte
[88] D. Zhao, et al., Ant colony optimization with horizontal and vertical crossover hypertrophy by inducing autophagy, Am J Transl Res 12 (5) (2020) 1894–1903.
search: fundamental visions for multi-threshold image segmentation, Expert Syst. [121] J.P. Doweiko, D.J. Nompleggi, Role of albumin in human physiology and
Appl. 167 (2021), 114122. pathophysiology, JPEN - J. Parenter. Enter. Nutr. 15 (2) (1991) 207–211.
[89] Y. Liu, et al., Horizontal and vertical crossover of Harris hawk optimizer with [122] J.J. Sands, et al., Intradialytic hypotension: frequency, sources of variation and
Nelder-Mead simplex for parameter estimation of photovoltaic models, Energy correlation with clinical outcome, Hemodial. Int. 18 (2) (2014) 415–422.
Convers. Manag. 223 (2020), 113211.

40
X. Yang et al. Computers in Biology and Medicine 147 (2022) 105752

[123] G. Santagati, et al., Positron emission tomography can support the diagnosis of [137] B. Li, et al., NOREVA: normalization and evaluation of MS-based metabolomics
dialysis-related amyloidosis, J. Clin. Med. 8 (9) (2019). data, Nucleic Acids Res. 45 (W1) (2017) W162–W170.
[124] H.C. Chen, et al., Abdominal aortic calcification score can predict future coronary [138] Q. Guan, et al., Medical image augmentation for lesion detection using a texture-
artery disease in hemodialysis patients: a 5-year prospective cohort study, BMC constrained multichannel progressive GAN, Comput. Biol. Med. 145 (2022),
Nephrol. 19 (1) (2018) 313. 105444.
[125] P. Ureña-Torres, et al., Valvular Heart Disease and Calcification in CKD: More [139] Y. Chen, et al., Generative adversarial networks in medical image augmentation: a
Common than Appreciated, Nephrol Dial Transplant, 2019. review, Comput. Biol. Med. 144 (2022), 105382.
[126] J. Park, et al., A comparative effectiveness research study of the change in blood [140] J. Yin, et al., Varidt 1.0: variability of drug transporter database, Nucleic Acids
pressure during hemodialysis treatment and survival, Kidney Int. 84 (4) (2013) Res. 48 (D1) (2020) D1042–D1050.
795–802. [141] F. Zhu, et al., Therapeutic target database update 2012: a resource for facilitating
[127] Y. Su, et al., A heuristic algorithm for identifying molecular signatures in cancer, target-oriented drug discovery, Nucleic Acids Res. 40 (D1) (2012) D1128–D1136.
IEEE Trans. NanoBioscience 19 (1) (2019) 132–141. [142] Y. Li, et al., Clinical trials, progression-speed differentiating features and swiftness
[128] L. Li, et al., SCMFMDA: predicting microRNA-disease associations based on rule of the innovative targets of first-in-class drugs, Briefings Bioinf. 21 (2) (2020)
similarity constrained matrix factorization, PLoS Comput. Biol. 17 (7) (2021) 649–662.
e1009165. [143] F. Zhu, et al., Clinical success of drug targets prospectively predicted by in silico
[129] Q.-W. Wu, et al., Extra trees method for predicting LncRNA-disease association study, Trends Pharmacol. Sci. 39 (3) (2018) 229–231.
based on multi-layer graph embedding aggregation, IEEE ACM Trans. Comput. [144] F. Zhu, et al., Clustered patterns of species origins of nature-derived drugs and
Biol. Bioinf (2021), https://doi.org/10.1109/TCBB.2021.3113122. clues for future bioprospecting, Proc. Natl. Acad. Sci. U. S. A 108 (31) (2011)
[130] Y. Su, et al., A community structure enhancement-based community detection 12943–12948.
algorithm for complex networks, IEEE Transactions on Systems, Man, and [145] G. Liang, et al., A text GAN framework for creative essay recommendation,
Cybernetics: Systems 51 (5) (2019) 2833–2846. Knowl. Base Syst. 232 (2021), 107501.
[131] Y. Tian, et al., EMODMI: a multi-objective optimization based method to identify [146] H. Yu, et al., Corn leaf diseases diagnosis based on K-means clustering and deep
disease modules, IEEE Transactions on Emerging Topics in Computational learning, IEEE Access 9 (2021) 143824–143835.
Intelligence 5 (4) (2020) 570–582. [147] H. Cui, et al., A novel advancing signal processing method based on coupled
[132] S. Qiu, et al., Sensor Combination Selection Strategy for Kayak Cycle Phase multi-stable stochastic resonance for fault detection, Appl. Sci. 11 (12) (2021)
Segmentation Based on Body Sensor Networks, IEEE Internet of Things Journal, 5385.
2021, https://doi.org/10.1109/JIOT.2021.3102856. [148] S. Qiu, et al., Sensor network oriented human motion capture via wearable
[133] Z. Wu, et al., Constructing dummy query sequences to protect location privacy intelligent system, Int. J. Intell. Syst. (2021), https://doi.org/10.1002/int.22689.
and query privacy in location-based services, World Wide Web 24 (1) (2021) [149] K. Yu, et al., Dynamic selection preference-assisted constrained multiobjective
25–49. differential evolution, IEEE Transactions on Systems, Man, and Cybernetics:
[134] Z. Wu, et al., A location privacy-preserving system based on query range cover-up Systems (2021), https://doi.org/10.1109/TSMC.2021.3061698.
for location-based services, IEEE Trans. Veh. Technol. 69 (2020). [150] K. Yu, et al., Purpose-directed two-phase multiobjective differential evolution for
[135] X. Ran, et al., A novel k-means clustering algorithm with a noise algorithm for constrained multiobjective optimization, Swarm Evol. Comput. 60 (2021),
capturing urban hotspots, Appl. Sci. 11 (23) (2021), 11202. 100799.
[136] J. Fu, et al., Optimization of metabolomic data processing using NOREVA, Nat. [151] J. Liang, et al., A Survey on Evolutionary Constrained Multi-Objective
Protoc. 17 (1) (2022) 129–151. Optimization, IEEE Transactions on Evolutionary Computation, 2022, https://
doi.org/10.1109/TEVC.2022.3155533.

41