=Paper=
{{Paper
|id=Vol-3896/paper23
|storemode=property
|title=Comparative analysis of models for assessing asthma severity based on paraclinical studies
|pdfUrl=https://ceur-ws.org/Vol-3896/paper23.pdf
|volume=Vol-3896
|authors=Oleh Pihnastyi,Olha Kozhyna,Iuliia Karpushenko
|dblpUrl=https://dblp.org/rec/conf/ittap/PihnastyiKK24
}}
==Comparative analysis of models for assessing asthma severity based on paraclinical studies==
https://ceur-ws.org/Vol-3896/paper23.pdf
Comparative analysis of models for
assessing asthma severity based on
paraclinical studies
Oleh Pihnastyi1,∗,†, Olha Kozhyna2,† and Iuliia Karpushenko3,†
1
National Technical University "Kharkiv Polytechnic Institute", 2 Kyrpychova, Kharkiv, 61002, Ukraine
2
Kharkiv National Medical University, 4 Nauky Avenue, Kharkiv, 61022, Ukraine
3
Kharkiv National Medical University, 4 Nauky Avenue, Kharkiv, 61022, Ukraine
Abstract
Severe course of bronchial asthma remains a human problem that can lead to death of the patient.
To improve diagnostic and prognostic processes in asthma management, this study is conducted
using neural network-based models. This paper presents a comparative analysis of models for
assessing the severity of bronchial asthma in children based on the paraclinical data. The study
included 90 children aged 6 to 18 years including 70 children diagnosed with bronchial asthma and
20 healthy children as a control group. The main biomarker, that is researched in the paper, is
serum thymic stromal lymphopoietin. Its concentration is measured by enzyme-linked
immunosorbent assay and used as the main parameter in the construction of regression models.
Two predictive models using neural networks are developed in this paper. The first model is
focused on thymus stromal lymphopoietin levels as the main predictor of asthma severity, whereas
the second model is taken into account a wider range of laboratory parameters including total
clinical blood counts, immunoglobulin E levels and measures of immune status. The models were
trained and tested on the same paraclinical study dataset. The neural network architecture was
standardized to ensure comparability of the models. The results suggest that integrating multiple
biomarkers and laboratory examination measures into predictive models may offer more reliable
and cost-effective tools for assessing asthma severity, especially in resource-limited settings. The
study emphasizes the importance of developing alternative diagnostic methods that are accessible
and affordable, especially in countries where the availability of biomarkers such as thymic stromal
lymphopoietin may be limited.
Keywords
TSLP, bronchial asthma, child, neural network, multilayer perceptron 1
1. Introduction
The severe and uncontrolled course of asthma is responsible for several thousand deaths per
day worldwide [1]. Asthma is a chronic disease of the respiratory system occurring at all
1
ITTAP’2024: 4th International Workshop on Information Technologies: Theoretical and Applied Problems,
November 20–22, 2024, Ternopil, Ukraine, Opole, Poland
∗
Corresponding author.
†
These authors contributed equally.
pihnastyi@gmail.com (O. Pihnastyi); olga.kozhyna.s@gmail.com (O. Kozhyna); yv.karpushenko@knmu.edu.ua
(I. Karpushenko)
0000-0002-5424-9843 (O. Pihnastyi); 0000-0002-4549-6105 (O. Kozhyna); 0000-0002-2196-8817 (I. Karpushenko
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
ages. The study of bronchial asthma has increased the understanding of the molecular
mechanisms underlying the heterogeneity of airway inflammation [2]. One of the key
substances, that is involved in the formation of allergic diseases, is thymic stromal
lymphopoietin (TSLP). TSLP was first isolated in thymus stromal cells and belongs to the
group of cytokines [3]. When respiratory tract epithelium is damaged by viruses, bacteria,
allergens, and chemical irritants, TSLP is released, which triggers a pronounced inflammatory
response [4]. The paper of A. Berraies et. al. is devoted to the determination of an increased
level of TSLP in induced sputum in children with asthma [5]. A similar problem is explored in
[6], where the study of biopsy samples of bronchial epithelium of patients with asthma
revealed an increased level of TSLP. A correlation between allergen exposure to bronchi in
patients with asthma and higher expression of TSLP+???? cells in bronchial epithelium and
submucosa is proved [7]. The key role of cytokines IL-25, IL-33 and TSLP, as the main
regulators of airway inflammation, in the formation of allergic rhinitis, chronic rhinosinusitis
and asthma has been shown [8]. Negative effect of thymic stromal lymphopoietin cytokine on
the course of disease was proved in [9, 10, 11]. The drug Tezepelumab, a human monoclonal
antibody that is designed to inhibit TSLP, is currently being investigated. The use of the drug
provided clinically significant, rapid and sustained relief of asthma exacerbations regardless of
asthma phenotype, including patients with severe and uncontrolled asthma [12]. The use of
the inhaled drug Ecleralimab, a potent neutralizing antibody fragment against human TSLP, in
patients with mild atopic asthma significantly attenuated allergen-induced bronchospasm and
airway inflammation [13].
One of the key challenges in the control of bronchial asthma is its heterogeneity in severity
and susceptibility to prescribed medications. Modern and effective biomarkers, which include
TSLP, can optimize the diagnosis and treatment of patients with asthma and help predict
patient functional and clinical outcomes. To determine the laboratory parameters of TSLP in
blood, reagents are relatively expensive for patients in countries such as Ukraine and are often
unavailable in the laboratory [14]. These factors limit the possibility of widespread use of
TSLP for the diagnosis of the disease and require the development of alternative effective and
low-cost methods for diagnosing bronchial asthma.
Application of prediction models based on neural network to analyze the data of
immunological examination of a patient opens new perspectives for creation of accessible and
inexpensive methods of diagnostics of bronchial asthma course severity, that determines the
significance and relevance of the present study.
2. Materials and methods
Human TSLP ELISA kit was used for the determination of TSLP level in serum. This method is
a 90-minute sandwich immunoassay in a 96-well microtiter plate. The amount of TSLP is
measured in picograms per 1 mL of serum (pg/mL). Our study involved 90 children aged 6 to
18 years, divided into two groups. The main group consisted of 70 children diagnosed with
bronchial asthma. The control group included 20 healthy children. The average age of
children with bronchial asthma is 11 years. The main group included 20 patients with
intermittent, 20 patients with mild, 20 patients with moderate persistent and 10 patients with
severe persistent asthma. The study was conducted with respect for human rights and in
accordance with international ethical requirements; it doesn't violate any scientific ethical
standards and standards of biomedical research [15]. Anamnestic data of patients, clinical
symptoms of the disease and indicators of laboratory tests are studied. For every patient,
information on 142 factors that could be the cause of bronchial asthma was gathered,
processed and analyzed. As a result of preliminary research, factors that can act as basic
factors for the creation of alternative methods of asthma diagnosis are identified.
Mathematical statistics methods based on neural network is used to build prediction models.
These methods are a tool for building dependencies between the degree of asthma severity
and: a) the value of TSLP and the degree of bronchial asthma severity; b) the values of factors
based on clinical blood tests. Comparative analysis of prediction results using the two types
of dependencies will make it possible to assess the accuracy of prediction. As a result of a
preliminary study, factors are identified that can act as basic factors for the creation of
alternative methods for diagnosing asthma.
In the comparative analysis, the following actions are performed for each patient: 1)
general clinical blood test parameters, biochemical studies of TSLP, Ig E and immunogram
data are determined; 2) classification of the disease severity degree is performed based on the
unified clinical protocol of primary and secondary (specialized) medical care “Bronchial
asthma in children” [16].
To analyze the results of the study, two models are built using a neural network: a) the first
model for predicting the severity of bronchial asthma is based on the patient immunogram
values; b) the second model for predicting the severity of the disease is based on the values of
factors for clinical blood analysis. The outcome of the comparative analysis is the prediction
accuracy of each of the two models for the tested group of patients. Before the comparative
analysis, the patients are divided into two groups, one of which is used for training the neural
network and the other for testing.
3. Problem statement
To comparatively analyze the prediction accuracy, two models are built using neural network.
Model Preparation 1. To determine the significance of TSLP in the formation of disease
severity, the level of cytokine depending on the severity of AD is analyzed. Evaluation of
TSLP levels in children with AD depending on disease severity revealed significant
fluctuations of 0.14...149.01 pg/mL and no directly proportional dependence of asthma severity
on serum TSLP concentration. On the one hand, in intermittent, mild persistent, and moderate
persistent forms of asthma, the median and interquartile intervals of TSLP levels are lower
than in controls and did not correlate with disease severity. A probable increase of TSLP level
in children with severe persistent course of AD compared with the control group (almost 5-
fold increase) and with milder asthma is found, which confirmed the presence of this cytokine
in the pathogenesis of some severe forms of AD and substantiated the search for the influence
of TSLP on the formation of certain clinical characteristics of the disease. A subgroup of
patients with severe asthma showed an increase in TSLP expression despite therapy with high
doses of inhaled or oral corticosteroids [17]. To determine the values of TSLP in the
pathogenesis of bronchial asthma, the levels of TSLP according to the clinical features of the
disease are calculated and used in the present study (Table 1).
Table 1
Dependence of TSLP on clinical features of asthma
Clinical sign of Number Level of TSLP, pg/ml Р1 Р2
BA of Median (Q1;Q3)
patients
Manifestation early (under 3 20 17.93(6.13;40.63) 0.783 0.547
years)
of asthma late (after 3 years) 50 12.44(5.95;28.01) 0.706
Duration less than 3 years 16 7.84(5.95;19.93) 0.413 0.272
of disease more than 3 years 54 13.17(5.04;32.64) 0.836
less than 7 years 31 7.84(5.22;20.47) 0.133 0.275
more than 7 years 39 13.94(5.95;36.45) 0.936
Atopy increased IgE level 59 11.76(4.50;27.01) 0.502 0.676
normal IgE level 11 13.21(9.58;20.47) 0.555
Clinical eosinophilia 26 8.71(4.50;16.85) 0.099 0.277
blood count level of eosinophils 44 13.57(5.95;30.28) 0.971
<5%
Comorbid atopic dermatitis 6 36.08(5.04;101.63) 0.195 0.324
conditions atopic dermatitis 64 11.76(5/22;21.93) 0.472
allergic rhinitis 39 13.21(5.22;59.69) 0.131 0.731
no rhinitis 31 8.13(5.22;19.02) 0.159
Allergy negative 30 15.03(7.37;55.33) 0.027 0.593
heredity positive 40 7.99(4.50;21.2) 0.218
Asthma negative 17 17.57(5.95;76.03) 0.148 0.437
heredity positive 53 11.76(5.22;21.93) 0.340
Р1 – comparison between the clinical sign presence and absence groups’
Р2 – comparison with the control group
Output parameters of the model are severity degrees of bronchial asthma: SEVERE
PERSISTENT, MODERATE PERSISTENT, MILD PERSISTENT, INTERMITTENT. Numerical
characteristics of the model output factors are summarized in Table 2.
Preparation Model 2. The general clinical blood test refers to the most common type of
blood test and is administered at the first stage of any examination. A blood smear is
examined to quantify hematologic parameters: Hemoglobin content (HGB) count; Red blood
cell (RBC) count; White Blood Cell (WBC) count; Eosinophils %, Basophils %, Band neutrophils
%, Segmented neutrophils %, Lymphocytes %, Monocytes %; Blood platelets (platelets);
Erythrocytes Sedimentation Rate (ESR). Determination of total immunoglobulin E (Ig E) is
used as a screening test to detect susceptibility to allergic reactions. The selected factors
Leucocytes, Neutrophils %, CD3 %, CD4 %, CD8 %, CD16 %, CD22 %, CD25 % refer to
indicators of immune status - quantitative and qualitative indicators of cellular and humoral
immunity. Numerical statistical characteristics of the input set of factors of the model and
observed values are summarized in Table 2.
Table 2
Numerical characteristics of factors determining the severity of bronchial asthma
Cod
Regressor name
e
RBC 4.6764 0.2632
HGB 136.7297 20.115
WBC 6.6722 1.9
Eosinophils % 4.0 3.5311
Basophils % 0.1429 0.3927
Band neutrophils % 1.2813 0.78
Segmented neutrophils % 58.9531 11.8631
Lymphocytes % 31.1094 10.4957
Monocytes % 4.4688 2.537
ESR 3.8594 1.9435
Ig E 0.8308 0.375
Leucocytes 6.4912 1.3099
Neutrophils % 59.45 7.3449
CD3 % 69.6 2.9983
CD4 % 40.0 1.4318
CD8 % 29.6375 1.6374
CD16 % 12.975 1.565
CD22 % 18.7 1.0654
CD25 % 25.5875 7.7164
SEVERE PERSISTENT 0.0444 0.2082
MODERATE PERSISTENT 0.3111 0.4657
MILD PERSISTENT 0.3111 0.4562
INTERMITTENT 0.3333 0.4740
Calculated indicators of correlation coefficients between the model factors , and between
the factors of the model and the observed value are also presented in Table 3.
Table 3
Correlation coefficients values ,
0.32 -
0.06 0.09 -
-0.07 -0.14 -0.01 -
0.04 0.05 0.17 0.12 -
0.11 0.12 0.07 0.03 0.07 -
0.25 0.18 0.28 -0.4 0.02 -0.05 -
-0.29 -0.22 -0.35 0.09 -0.11 -0.04 -0.91 -
0.04 0.2 0.15 0.15 0.03 0.07 -0.29 0.04 -
0.1 -0.01 0.02 -0.00 0.06 0.58 0.04 -0.11 0.08 -
-0.1
0.11 -0.18 -0.09 0.25 0.07 -0.11 -0.02 -0.04 -0.05
4
-0.01 0.09 -0.12 -0.09 0.03 0.13 -0.04 0.06 0.05 0.16
-0.02 0.03 0.12 -0.09 -0.13 -0.03 0.23 -0.17 -0.25 0.07
-0,1
-0.05 0.02 0.08 -0.02 0.04 -0,02 -0,16 0,13 0,23
2
In this study, a multilayer perceptron will be used to create each of the two models described
above for predicting the severity of bronchial asthma in children. The neural network
architecture contains several layers, each of which is connected both to the previous layer,
from which it receives data, and to the subsequent layer, which is influenced. The problem of
classifying the severity of the course of bronchial asthma for each of the sets of input factors
of the first and second models is considered. The calculation of the observed values in each
model utilized the same network architecture for the hidden and output layers, with each
node having the same activation functions for the two models. Softmax activation function
generates the values of the output layer nodes in each model, thus it is guaranteed that the
output nodes take positive values and the sum of all output node values is equal to one. The
output values of the output factors will be interpreted as the probability of the course degree
of bronchial asthma.
4. Prediction model of bronchial asthma severity
Let us consider the construction of comparative prediction models of bronchial asthma
severity based on a neural network represented by a multilayer perceptron with the
architecture . M is the number of input factors in the prediction model. is
the number of nodes in the -th hidden layer (k=1..L). is the number of nodes in the output
layer. The minimum number of neurons in the hidden layer is determined from the conditions
[18, 19]:
, , (1)
for a training sample of size . Hidden layer nodes are characterized by a Sigmoid activation
function
(2)
For the output layer nodes, Softmax function is set as the activation function:
(3)
Softmax function makes it possible to normalize the values of the model output parameters,
that makes it possible to predict different degrees of severity of bronchial asthma disease in
the studied patients, treating them as probability of disease in accordance with the
classification of bronchial asthma by severity of course. Discrete categories in the studied
models are treated as a set of values from a common probability distribution. Softmax of the
activation function converts the values of the output layer of the neural network into actual
discrete probability distributions of the severity of bronchial asthma disease course: SEVERE
PERSISTENT, MODERATE PERSISTENT, MILD PERSISTENT, INTERMITTENT.
The neural network architecture for model 1 and model 2 is shown in Figure 1, Figure 2,
respectively. The architecture of the models differs in the number of neurons in the input
layer. The architecture of hidden layers and output layer in model 1 and model 2 are the same.
Figure 1: Neural network architecture in a model 1 for predicting the severity of the course of
bronchial asthma (M-M1-M2-M3-Z , model input parameters: TSLP).
Figure 2: Neural network architecture in a model 2 for predicting the severity of the course of
bronchial asthma (M-M1-M2-M3-Z , input parameters of the model: factors of clinical blood
analysis).
For each of the models, weights and bias are initialized by a normal distribution of values with
parameters. When training neural networks for each of the prediction models, the following
hyperparameters are used: a) learning rate ; b) number of training epochs:
.
5. Analyzing the results
A comparative analysis of two models for predicting the bronchial asthma severity in children
presents a study of the prediction accuracy. Both models use neural networks but differ in
their input parameters. The TSLP-model is based on thymus stromal lymphopoietin level as
the main predictor, while the CBT-Model includes a broader set of factors derived from
clinical blood tests.
Preliminary analysis of groups of factors [20], with which the severity of bronchial asthma
is associated, allowed us to choose the group of factors of clinical blood analysis as the initial
set of regressors. A separate study will be devoted to the detailed issue of the factor group
selection. The present paper focuses on the comparative analysis of the prediction accuracy
for the two models. The first step in the comparative analysis of the two methods for
diagnosing the bronchial asthma severity course is the choice of neural network architecture
for the prediction models. The TSLP-model uses only one input parameter, the TSLP score in
clinical analysis. The CBT-model uses clinical blood test values as input parameters. To ensure
training, the weights and bias parameters in both models were initialized using a normal
distribution with specified characteristics. The training is performed using hyperparameters: a
learning rate of 0.001 and 20000 epochs. A sigmoid activation function is used in the hidden
layers to promote nonlinear transformations of the input data, while a Softmax function in the
output layer allowed probabilistic predictions of the asthma severity.
For the comparative analysis of the prediction models, the architecture 1-35-50-35-5
(Figure 1) for the TSLP-model and the architecture 19-35-50-35-5 (Figure 2) for the CBT-
model are chosen. Two neural network architectures differ in the number of the nodes in the
input layer corresponding to the input factors of the models. The number of the nodes in the
hidden layers is chosen based on the condition that the training sample containing about 10 5
patients under study will be used in subsequent experiments [21]. Taking into account the
formula (1), the number of hidden nodes for the model with one hidden layer is determined by
the inequality
, (4)
which corresponds to 35-50 nodes for each of the three hidden layers for the selected
architectures (Figure 1, Figure 2). When training the neural network for each of the prediction
models (model 1: TSLP model, Figure 1; model 2: CBT model, Figure 2), the sample for training
the neural network is divided into two data sets. The first dataset directly serves to train the
neural network and is ~80%, the second dataset (test dataset), which is ~20%, is used to verify
the accuracy of the training process. After the training, both models are tested on the
validation dataset to evaluate their prediction accuracy. The training process of the neural
network for TSLP-model is shown in Figure 3. To demonstrate the neural network training
process, four samples with maximum accurancy value are selected containing the above two
datasets. The maximum accurancy value is 0.67 for the training process of TSLP-model.
Quantitative indicators characterizing the quality of the neural network training process are
presented in Table 4.
a) b)
c) d)
Figure 3: Training a TSLP-model neural network (basic neural network architecture)
Table 4
TSLP-model prediction result (baseline architecture)
initial predict
PERSISTENT
PERSISTENT
MODERATE
PERSISTEN
PERSISTEN
MITTENT
MODERAT
PERSISTEN
PERSISTEN
MITTENT
HEALTY
SEVERE
SEVERE
HEALTY
MILD
INTER
INTER
MILD
N#
T
T
T
E
1 0 0 1 0 0 0,0133 0,2876 0,3964 0,1115 0,1911
2 0 0 1 0 0 0,0444 0,1898 0,2184 0,4214 0,1260
3 0 1 0 0 0 0,2818 0,1769 0,0582 0,0863 0,3968
4 0 0 0 1 0 0,0348 0,2078 0,2463 0,4005 0,1106
5 0 0 0 1 0 0,0161 0,3009 0,3888 0,1774 0,1167
6 0 0 1 0 0 0,0674 0,1789 0,1887 0,3897 0,1753
7 0 0 0 1 0 0,2477 0,1733 0,0669 0,1007 0,4115
8 0 1 0 0 0 0,0161 0,3009 0,3888 0,1774 0,1167
9 1 0 0 0 0 0,1077 0,1773 0,1537 0,2910 0,2702
10 0 0 1 0 0 0,0844 0,1779 0,1737 0,3485 0,2155
… … … … … … … … … … …
An exact match is presented for ~30% of the studied patients. In ~80% of cases, the TSLP-
model detects either an exact match or a neighboring group in the qualification table, which is
a reasonably good prediction result for the TSLP-model.
For comparative analysis, the training process of the neural network for TSLP model is
shown in Figure 4. As for the TSLP-model, the CBT-model shows a maximum accurancy value
of 0.67 for the training process, equal to the value obtained for the TSLP-model. This is a good
enough result to suggest the use of the CBT-model as an alternative to the TSLP model for
prediagnosis of bronchial asthma severity.
a) b)
c) d)
Figure 4 Training a neural network of a CBT-model (basic architecture of a neural network)
One difference is that for a given number of epochs ( for each of the model
), loss, accurancy and test loss functions asymptotically tend to the value
determined by the model parameters, while the learning process for the CBT-model neural
network is not steady-state. It should also be noted that for the TSLP-model the loss and test
loss functions asymptotically approach each other. Such behavior is not observed for the CBT
model, which also indicates that the learning process for the CBT-model should contain a
larger number of epochs. The latter circumstance is to some extent explained by the fact that
the CBT-model contains significantly more input factors than the TSLP-model.
For a comparative analysis of the two prediction models, quantitative indicators
characterizing the quality of the training process of the neural network of the CBT-model are
presented in Table 5.
Table 5
CBT-model prediction results
initial predict
PERSISTENT
PERSISTENT
MODERATE
PERSISTENT
PERSISTENT
MODERATE
PERSISTENT
PERSISTENT
MITTENT
MITTENT
HEALTY
SEVERE
SEVERE
HEALTY
MILD
INTER
INTER
MILD
N#
1 0 0 0 1 0 0,0018 0,0649 0,0387 0,8931 0,0015
2 0 0 1 0 0 0,1549 0,3465 0,2732 0,2067 0,0187
3 0 0 0 1 0 0,0312 0,0658 0,0474 0,7738 0,0817
4 0 0 0 0 1 0,0051 0,0139 0,0053 0,0838 0,8919
5 0 0 0 1 0 0,0080 0,0053 0,0018 0,9835 0,0014
6 0 1 0 0 0 0,1524 0,3559 0,2564 0,2162 0,0191
7 0 0 1 0 0 0,0011 0,0567 0,9387 0,0034 0,0001
8 0 0 1 0 0 0,1524 0,3559 0,2564 0,2163 0,0191
9 1 0 0 0 0 0,8815 0,0032 0,0329 0,0582 0,0241
10 0 0 1 0 0 0,1519 0,3575 0,2560 0,2156 0,0190
… … … … … … … … … … …
An exact match in the present study for the training and test dataset is found for ~70% of the
patients studied. In each case, the CBT-model predicted either an exact match or a
neighboring group in the qualification table, which is a rather unexpected result obtained. The
second result achieved is also noteworthy. The values of probability of predicting the severity
of bronchial asthma disease course for CBT-model is much higher, reaching the value from 0.8
to 1.0 for a rather large group of analyzed patients. The obtained prediction results for the
CBT-model are explained by a large number of input factors obtained from clinical blood tests.
Figure 5: Neural network architecture in a TSLP-model for predicting the severity of the
course of bronchial asthma (neural network architecture 1-15-15-5)
In addition to this study, let us consider the effect of changing the neural network
architecture on the quality of the model for predicting the severity of bronchial asthma
disease. The process of training the neural network for TSLP model with 1-15-15-5 neural
network architecture (Figure 5) is shown in Figure 6.
As for the basic architecture, four samples with the maximum accuracy value were selected
to demonstrate the process of training the neural network. When training the neural network,
the maximum accuracy value decreased slightly. For the number of epochs defined in the base
case, the neural network training process did not reach steady-state. Quantitative indicators
characterizing the quality of the neural network training process are presented in Table 6.
An exact match is presented for ~20% of the study patients, which is lower than in the base
case. Also, as in the base case, for ~80% of cases the TSLP-model identifies either an exact
match or a neighboring group in the qualification table.
a) b)
c) d)
Figure 6: Training a TSLP-model neural network (neural network architecture
1-15-15-5)
Table 6
TSLP-model prediction result (neural network architecture 1-15-15-5)
initial predict
PERSISTENT
PERSISTENT
MODERATE
PERSISTENT
PERSISTENT
MODERATE
PERSISTENT
PERSISTENT
MITTENT
MITTENT
HEALTY
SEVERE
SEVERE
HEALTY
MILD
INTER
INTER
MILD
N#
1 0 0 1 0 0 0,0093 0,2475 0,3546 0,2121 0,1764
2 0 0 1 0 0 0,0480 0,1887 0,2110 0,4347 0,1176
3 0 1 0 0 0 0,2643 0,2137 0,0559 0,0754 0,3907
4 0 0 0 1 0 0,0346 0,2143 0,2390 0,4136 0,0985
5 0 0 0 1 0 0,0112 0,3065 0,4011 0,1868 0,0944
6 0 0 1 0 0 0,0778 0,1674 0,1885 0,3934 0,1728
7 0 0 0 1 0 0,2372 0,1907 0,0676 0,0881 0,4164
8 0 1 0 0 0 0,0112 0,3065 0,4011 0,1868 0,0944
9 1 0 0 0 0 0,1225 0,1574 0,1624 0,2820 0,2757
10 0 0 1 0 0 0,0978 0,1617 0,1782 0,3457 0,2165
… … … … … … … … … … …
The comparative analysis shows that the CBT model, which incorporates multiple clinical
factors, is not inferior in predicting the quality of the TLSP-model based solely on the TSLP
level. Although TSLP is an important biomarker associated with asthma severity, it does not
encompass the entire multifactorial nature of the disease. The sufficiently high accuracy of the
CBT model indicates that the integration of multiple clinical indicators provides a more
complete picture of disease severity.
In addition, the sensitivity and specificity of both models were evaluated to better
understand their diagnostic capabilities. The CBT-model showed higher sensitivity in
detecting severe and moderate cases of asthma, while maintaining an acceptable level of
specificity. The latter indicates that the CBT-model is an alternative for clinical applications
where accurate determination of asthma severity plays a key role in treatment planning.
In conclusion, comparative model analysis provides valuable results integrating different
clinical parameters, providing an accurate and reliable method for predicting the severity of
bronchial asthma in children. This approach is in line with the general trend in medicine
towards precision diagnosis and personalized treatment.
6. Conclusion
The present study successfully demonstrated the feasibility of using neural network-based
predictive models to determine the severity of bronchial asthma in children by analyzing
various clinical and immunological parameters. Two models developed, one based on thymic
stromal lymphopoietin (TSLP) levels and the other on factors of the total clinical blood count,
provided a clear indication of the correlation between model regressors and disease severity.
Although there was no direct proportional relationship between TSLP levels and asthma
severity in all cases, the TSLP-based model emphasized the potential of TSLP as a biomarker,
especially in severe asthma. On the other hand, the model based on total clinical blood count
showed significant predictive accuracy, offering an alternative method for assessing disease
severity.
The comparative analysis of these models demonstrates that, although each model has its
own merits, the integration of several biomarkers and clinical parameters can improve the
accuracy and reliability of asthma severity predictions. It is important to emphasize that the
model based on clinical blood count factors demonstrated a prediction accuracy
commensurate with that based on thymic stromal lymphopoietin (TSLP) levels, which
supports the hypothesis that this model can be used as one method of reasonably accurate
diagnosis of asthma severity.
In the absence of TSLP reagents in clinical institutions, a prediction model based on the
results of the general clinical blood test can be used to diagnose the severity of the course of
bronchial asthma quite effectively. If TSLP-based paraclinical tests are available, the developed
diagnostic models can act as verification of TSLP laboratory test results.
The fact that the TSLP model relies on a single biomarker limits its ability to predict
asthma severity across the disease spectrum. In addition, it should be considered that the
models were trained on a relatively small dataset for comparative analysis of prediction
performance. Future research could focus on expanding the dataset and incorporating
additional biomarkers and laboratory examination scores to further improve the prediction
accuracy of the models. In addition, investigating alternative neural network architectures,
such as convolutional or recurrent networks, may provide a deeper understanding of the
temporal dynamics of asthma severity. Integrating data on disease dynamics over time will
improve the ability of models to predict changes in asthma severity and facilitate the
development of personalized treatment strategies.
The results emphasize the importance of developing cost-effective and accessible
diagnostic tools tailored to the specific needs of resource-limited regions. Further research and
refinement of these models could lead to improved, personalized treatment plans, ultimately
improving patient outcomes and reducing the burden of bronchial asthma on health care
systems.
References
[1] Global Initiative for Asthma. Global Strategy for Asthma Management and Prevention,
2024, https://ginasthma.org/reports/.
[2] M. E. Kuruvilla, F. Eun-Hyung Lee, G. B. Lee, Understanding asthma phenotypes,
endotypes, and mechanisms of disease, Clin Rev Allergy Immunol, 2019, vol. 56(2),
pp. 219–233, doi: 10.1007/s12016-018-8712-1.
[3] G. Varricchi, A. Pecoraro, G. Marone, G. Criscuolo, G. Spadaro, A. Genovese et al, Thymic
stromal lymphopoietin isoforms, Inflammatory Disorders, and Cancer, Frontiers in
Immunology, 2018, 9, doi: 10.3389/fimmu.2018.01595.
[4] G. M. Gauvreau, R. Sehmi, C. S. Ambrose, J. M. Griffiths, Thymic stromal lymphopoietin:
its role and potential as a therapeutic target in asthma, Expert Opin Ther Targets, 2020,
vol. 24(8), pp. 777–792, doi: 10.1080/14728222.2020.1783242.
[5] A. Berraies, B. Hamdi, J. Ammar, K. Hamzaoui, A. Hamzaoui, Increased expression of
thymic stromal lymphopoietin in induced sputum from asthmatic children, Immunol.
Lett, 2016, vol. 178, pp. 85–91, https://doi.org/10.1016/j.imlet.2016.08.004.
[6] D. Al-Sajee, R. Sehmi, T.J. Hawke et al, Expression of IL-33 and TSLP and their receptors
in asthmatic airways after inhaled allergen challenge, Am J Respir Crit Care Med, 2018,
vol.198(6), pp. 805–807, doi: 10.1164/rccm.201712-2468LE.
[7] W. Wang, Y. Li, Z. Lv, Y. Chen, Y. Li, et al, Bronchial allergen challenge of patients with
atopic asthma triggers an alarmin (IL-33, TSLP, and IL-25) response in the airways
epithelium and submucosa, J Immunol, 2018, vol. 201(8), pp. 2221–2231,
https://doi.org/10.4049/jimmunol.1800709.
[8] H. Hong, S. Liao, F. Chen, Q. Yang, De-Yun Wang, Role of IL-25, IL-33, and TSLP in
triggering united airway diseases toward type 2 inflammation, Allergy, 2020, vol.75(11),
pp. 2794-2804, doi: 10.1111/all.14526.
[9] H. Guo, X. Ji, G.Yang, Y. Jin, Abnormal thymic stromal lymphopoietin expression in the
gastrointestinal mucosa of patients with eosinophilic gastroenteritis, Jornal de Pediatria,
2020, vol. 96(3), pp. 350-355, doi: 10.1016/j.jpedp.2019.03.002.
[10] S. Colicino, D. Munblit, C. Minelli, A. Custovic, P. Cullinan, Validation of childhood
asthma predictive tools: a systematic review, Clin Exp Allergy, 2019; 49(4), pp. 410- 418,
doi: 10.1111/cea.13336.
[11] O. Kozhyna, O. Pihnastyi, Covariance coefficients factors from a clinical study of the
severity of bronchial asthma in children of the Kharkov region, 2017, Mendeley Data, 1,
2019.
[12] J. R. Parnes, N. A. Molfino, G. Colice, U. Martin, J. Corren et al, Targeting TSLP in
asthma, J Asthma Allergy, 2022, vol.15, pp. 749-765, doi: 10.2147/JAA.S275039.
[13] G. Gauvreau, J. M. Hohlfeld, J. M. FitzGerald, L.-P. Boulet, D. W. Cockcroft et al, Inhaled
anti-TSLP antibody fragment, ecleralimab, blocks responses to allergen in mild asthma,
Eur Respir J, 2023, vol. 61(3):2201193, doi: 10.1183/13993003.01193-2022.
[14] https://www.elkbiotech.com/pro/ELK1083
[15] UNESCO.ORG, Universal Declaration on Bioethics and Human Rights, 2005,
http://portal.unesco.org/en/ev.php. 2005.
[16] https://zakon.rada.gov.ua/rada/show/v2856282-21
[17] A. Shikotra, D. F. Choy, C. M. Ohri, E. Doran, C. Butler et al, Increased expression of
immunoreactive thymic stromal lymphopoietin in patients with severe asthma, J Allergy
Clin Immunol, 2012, vol.129(1):104-11.e1-9, doi: 10.1016/j.jaci.2011.08.031.
[18] M. Paul, M. Chakraborty, Observation on training neural network for diagnosing
schizophrenia, International Journal of Advanced Research in Computer Science, 2018,
vol. 9(1), pp. 163–166, http://dx.doi.org/10.26483/ijarcs.v9i1.5279.
[19] D. Stathakis, How many hidden layers and nodes? International Journal of Remote
Sensing, 2009, vol. 30, pp. 2133-2147, doi:10.1080/01431160802549278.
[20] O. Kozhyna, O. Pihnastyi, Data Structure of Clinical Research, Human Health & Disease,
2019, vol. 3(9), p. 71-79, http://dx.doi.org/10.2139/ssrn.3336301.
[21] V. Klymenko, O. Kozhyna, K. Zemlianskyi, Prevalence of bronchial asthma symptomatic
manifestation among children of Kharkiv, The world of medicine and biology, 2019,
vol. 2(68), p.61-65, doi:10.26.724/2079-8334-2019-2-68-61-65.