© Aluna Publishing Wiadomości Lekarskie, VOLUME LXXIII, ISSUE 7, JULY 2020

ORIGINAL ARTICLE

APPLICATION OF NEURAL NETWORK TECHNOLOGIES

IN THE DENTAL CARIES FORECAST
DOI: 10.36740/WLek202007135

Oleksandr A. Udod1, Hanna S. Voronina1, Olena Yu. Ivchenkova2

1
DONETSK NATIONAL MEDICAL UNIVERSITY, LYMAN, UKRAINE
2
DONBASS STATE ENGINEERING ACADEMY, KRAMATORSK, UKRAINE

ABSTRACT
The aim: of the work was to develop and apply in the clinical trial a software product for the dental caries prediction based on neural network programming.
Materials and methods: Dental examination of 73 persons aged 6-7, 12-15 and 35-44 years was carried out. The data obtained during the survey were used as input for the
construction and training of the neural network. The output index was determined by the increase in the intensity of caries, taking into account the number of cavities. To build
a neural network, a high-level Python programming language with the NumPay extension was used.
Results: The intensity of carious dental lesions was the highest in 35-44 years old patients – 6.69 ± 0.38, in 6-7 years old children and 12-15 years old children it was 3.85 ±
0.27 and 2.15 ± 0.24, respectively (p <0.05). After constructing and training the neural network, 61 true and 12 false predictions were obtained based on these indices, the
accuracy of predicting the occurrence of caries was 83.56%. Based on these results, a graphical user interface for the “CariesPro” software application was created.
Conclusions: The resulting neural network and the software product based on it permit to predict the development of dental caries in persons of all ages with a probability
of 83.56%.
KEY WORDS: caries, prognosis, neural network, software

Wiad Lek. 2020;73(7):1499-1504

INTRODUCTION technologies can be applied to solve difficult and monu-

Innovative and promising research areas include the use mental challenging tasks with numerous parameters and
of intelligent systems, in particular, those based on neural characteristics in various fields of healthcare.
networks [1,2,3]. In health care, such systems are used in State of the present-day health care is drawing attention
human genome research, drug development, diagnosis and to preventive care. Against this background, it is quite
prognosis of certain diseases, such as cancer, cardiovascular promising and, certainly, important to predict the devel-
diseases, etc. [2,3,4,5]. In the processing of medical data, opment and course of diseases, in particular dental ones,
information systems based on neural networks have sev- among which, undoubtedly, the leading place is occupied
eral advantages over traditional methods of mathematical by dental caries and its complications [7,8]. To date, the
statistics. Neural networks are capable of simultaneous prevalence and intensity of this disease in pediatric and
processing a large array of different parameters, which adult populations in different regions of Ukraine remain
is extremely important in the diagnosis and prognosis of at a fairly significant level [8,9,10]. Integrated caries pre-
diseases, particularly if the patient has several diseases. vention measures applied to different segments of the
In such systems, after their creation, there is a possibility population, including organized contingents, taking into
for “extra training” of the initial model, for example, with account regional peculiarities, do not always show high
the emergence of new or more accurate data of additional efficiency due to, first of all, the lack of an individualized
studies or “extra training” of the model on the initial data approach. Caries prevention, based on the principles of
in changing external conditions, for example, in the case selectivity, i.e.sampling, and timeliness, based on deter-
of new drugs prescribing during treatment [1,4,5]. mining the individual disposition toward caries, creates the
The work of neural networks is based on algorithms that preconditions for improving its efficacy, for more rational
are, in their turn, based on the principles and patterns of use of the dentist’s working time as well as therapeutic and
signal transmission through neurons and synapses of the preventative means. In this sense, the predictive field of
human nervous system. Neural networks are capable of cariesology, i.e. prediction of this disease course and the
taking decisions and returning results based on the hid- possible results of preventive measures for a particular
den patterns found in the large data array. The peculiarity patient, seems quite appropriate and necessary.
of neural networks is that they are not programmed and There are various approaches to the prediction of dental
do not use any rules to draw a conclusion, but they are caries in children and adults, well-known are, for exam-
trained to do it by examples [1,6]. Thus, neural network ple, mathematical models, as well as numerous predictive

1499
Oleksandr A. Udod et al.

methods based on the determining oral fluid properties, in total. The output, i.e. final, index (Y) was determined
the state of the dental hard tissues, hygienic parameters of by the increase in the intensity of dental caries, taking
the oral cavity, etc. [11]. Clinical and laboratory studies of into account the number of cavities. The obtained data-
the tooth enamel acid resistance and its restitution after base was divided into two samples – the training and the
acid exposure are considered to be the most informative control ones, which included indices of 65 and 8 patients,
and prognostically valuable. These include determination respectively.
of structural and functional acid resistance of tooth enamel To build a neural network, we used a high-level Python
by the enamel resistance test, clinical assessment of enamel programming language with the NumPay extension. The
remineralization rate, CRT-test (color, reaction, time), but following software products were used during the work:
they have recently been used infrequently enough [11,12]. Pandas, TensorFlow, Keras, Scikit-learn, Matplotlib, Jupyter
Unfortunately, most of the proposed prediction meth- Notebook [15].
ods are technically sophisticated, sometimes multi-step At the first of the successive stages series, the data were
and time-consuming, and quite frequently their imple- processed and prepared. Initially, using the Pandas elec-
mentation is not possible in daily clinical dental practice. tronic library for data analysis and processing, the gener-
Therefore, it is important to develop fundamentally new ated database was read from Excel files, after which the
approaches to the prediction of dental caries based on downloaded data was divided into input (X) and output
software products, especially since computer technologies (Y). Further, in order to further normalize the data for
are being actively implemented into the medical-diagnostic each of the factors used, maximum values were determined
process, particularly in private-owned dental institutions. using the standard NumPay function, according to which
the normalization was performed. This was necessary to
improve the next neural network training process and to
THE AIM reduce the input to a single computing system.
The aim of the work was to develop and apply a software The final step of this stage was to convert the output
product to predict dental caries on the basis of neural coordinate to the unitary code (one-hot encoding) using
network programming. the standard Scikit-learn function. The use of unitary code
in the construction of neural networks is required in cases
where the number of outputs corresponds to the number of
MATERIALS AND METHODS possible output coordinate classes. When a neural network
Dental examination of 73 persons was carried out, which receives a test sample of factors, the maximum value will
was divided into three age groups. The first group included be only set to one of the outputs.
26 children aged 6-7 years, the second group included 34 At the second stage of the neural network construction
12-15 years old children, the third group comprised 13 its future structure and training was drafted. This was
persons aged 35-44 years. During the examination, the done using the sequential method of describing the neu-
caries intensity in temporary (if present) and permanent ral network using Sequential software. During the work,
teeth was determined according to the index of cfc (caries, we created a neural network containing five neurons in
filling, cavity) + cfcex (extracted teeth) or CFcEX, taking the input layer (based on the number of available dental
into account the number of tooth decay cavities, oral indices), one hidden fully connected layer and an output
hygiene condition according to the simplified hygienic layer with four neurons. The number of synaptic weights
Green-Vermillion index (OHI-S), structurally-functional and neurons in the hidden layer was determined according
acid resistance of tooth enamel (SFARE) according to the to the Kolmogorov – Arnold – Hecht – Nielsen theorem
enamel resistance test (TER) and the functional component [15]. A ReLU was selected as the activation function for
of the structural and functional acid resistance of enamel the input layer, and the softmax logistic function – for ac-
(ΔTER) [13,14]. After 1 year, a repeated clinical examina- tivation of the output layer. In order to avoid the problem
tion was performed, each patient was re-assesed by cfc + of the neural network “retraining”, which may cause the
CFcEX or CFcEX index and the increase in the intensity dependence of the output index on the input data to occur
of carious lesions was calculated according to the Δcfc + sharply nonlinear, .i.e. the result will be significantly and
CFcEX or Δ CFcEX index. The results of the study were unpredictably changed with slight fluctuations of the input
recorded in the medical card of the dental patient (form data, the method of “thinning” was applied, better known
No. 043 / o). We used a standard package of statistical as Dropout. Its essence is that instead of training a single
software for Microsoft Excel to prepare the patient source neural network, there is a simultaneous training of several
data tables for and to carry out calculations. ones, with further generalization of the results obtained.
The following indices were used as input (X) data, on The RMSprop algorithm was used as the optimization
the basis of which the caries probability was predicted: method, for the quality metric of the constructed neural
age of the patient (X1), dental caries intensity index taking network – Accuracy.
into account the number of cavities (X2), hygienic index The duration of the neural network training process was
OHI-S (X3), structural- functional acid resistance of tooth 25,000 epochs, with 10% of the available sample data not
enamel (X4) and functional component of structural and participating in the process, but used for test purposes
functional acid resistance of enamel (X5), i.e., five indices only. After training, the neural network and the software

1500
APPLICATION OF NEURAL NETWORK TECHNOLOGIES IN THE DENTAL CARIES FORECAST

Fig. 1. Diagram of loss-function based

on training data (blue - train) and test
data (orange - test).

Fig. 2. Diagrams for visualization of

the dental caries intensity growth rate
prediction based on the developed
neural network.

Fig. 3. Visualization of changes in

accuracy (acc metrics) in the training
process of the resulting neural network.

Fig. 4. Graphical interface to the

main user form of the “CariesPro”
application.

1501
Oleksandr A. Udod et al.

product based on it, called CariesPro, was used to achieve the indices in the re-survey were statistically significantly
the set purpose, i.e., predicting dental caries. Matplotlib (p <0.05) different from the primary ones. The increase
e-library was used to visualize the training outcomes and in the intensity of dental caries in these individuals by the
the results obtained. indices Δcfc + CFcEX or Δ CFcEX was 2.42 ± 0.15; 2.09 ±
All the studies were carried out in accordance with 0.15 and 2.23 ± 0.25, respectively, that is, indices did not
the principles of the Declaration of Helsinki, adopted by differ statistically (p <0.05).
the General Assembly of the World Medical Association The results obtained during the two surveys were used
(1997-2000) and approved by the University Bioethics to train the neural network. The training outcomes were
Commission. visualized by constructing a loss function in Matplotlib.
This function was used to calculate the error between the
existing and the obtained result, i.e. its ultimate purpose
RESULTS is to reduce this error. Thus, the loss function of both in
In order to create an input database, a primary examination the training data that participated in the training process
of persons was carried out, during which it was found that and in the test data that did not participate in the training
the intensity of tooth decay by cfc + CFcEX index, taking process was minimized (fig. 1). This testified to the ability
into account the number of cavities in children aged 6-7 of the neural network to learn. Significant fluctuations in
years, was 3.85 ± 0.27, which, according to the WHO stan- loss functions during the transition from epoch to epoch
dard criteria is considered medium. Statistically significant occurred due to the use of Dropout.
(p <0.05), this index was lower in 12-15 years old children, After the process of training the neural network, a whole
in whom it was equal to 2.15 ± 0.24, which corresponds to selection of the reported factors was provided. In response,
the low level according to the WHO recommendations. The it has generated a category of numbers for the individual
intensity of carious lesions in teeth according to the CFcEX increase in caries intensity (Y), among which index 4 was
index was very high by the WHO criteria in patients aged never met. This is due to the imbalance of the training sam-
35-44 years, it was 6.69 ± 0.38 and statistically significantly ple and to the very poor representation of this result in it.
(p <0.05) higher than all the others. The graphical overlay of the predicted results and the
Assessment of the oral cavity hygienic condition by the real results diagrams, which were constructed in the same
simplified hygienic index OHI-S revealed its unsatisfactory coordinate system, indicated a coincidence of the results
level in individuals of all three age groups, and their mean (fig. 2).
values were not statistically significantly different (р0.05), Visualization of the accuracy of the constructed Accuracy
being 1.98 ± 0.05 points , 1.87 ± 0.03 points and 1.86 ± (acc) neural network, which was calculated using Keras, has
0.05 points, respectively. confirmed the correct course of the neural network training
The structural and functional acid resistance of the teeth process, and oscillations during the transition from epoch
enamel according to the enamel resistance test was the to epoch were due to the use of Dropout (fig. 3).
worst in children aged 6-7 and 12-15 years, in particular, As a result of calculating work accuracy index of the
their mean values were 4.96 ± 0.33 points and 5.00 ± 0.34 Accuracy (acc) neural network, 61 true and 12 false pre-
points (p 0.05), respectively. Somewhat better, but with dictions were obtained from the entire sample, i.e. the
no statistically significant difference (p <0.05), structural prediction accuracy of the constructed and trained neural
and functional acid resistance of enamel (SFARE) was in network was 83.56%.
patients of the older age group, making 4.15 ± 0.66 points. Later, using the high-level Delphi programming lan-
After carrying out hemostimulation of the dental pulp, guage, a graphical user interface was created (fig. 4).
which was performed by rubbing hypertonic sodium chlo- In order to work with the presented software, it is neces-
ride solution into the palatal surface of the tested tooth for sary to fill in the patient’s passport data and enter the five
10 minutes, the most significant dynamics of structural dental indices obtained during the clinical dental exam-
and functional acid resistance was recorded in 6-7 and ination. The following buttons are present in this window:
12-15 years old children, who had a functional SFARE “Network calculation” – when this button is pressed, the
component, 1.73 ± 0.15 points and 1.71 ± 0.12 points (p value of the probable number of carious cavities index of
<0.05), respectively [13]. the patient is calculated according to the set values in the
Statistically significantly (p <0.05) less, in particular, analysis data;
almost by 11.5 times, this index was found in persons “Save” – pressing this button saves patient data and
of 35-44 years – 0.15 ± 0.10 points. Such differences are estimated values;
undoubtedly related to the age patterns regarding the dy- “Monthly visit report” – pressing this button generates
namics of the enamel acid resistance and the dominance a report on the status of teeth and oral hygiene in patients
of its structural component in the elderly patients [14]. who visited a dentist in the last month;
After 1 year, the intensity of dental caries in children aged “Conclusion” – pressing this button generates a conclusion
6-7 years increased by 1.6 times to 6.27 ± 0.35, in 12-15 that contains data on the condition of the teeth and oral
years old children this index increased almost by 2 times hygiene during the current and past examinations of the
to 4.23 ± 0.24 , in persons aged 35-44 years – up to 8,92 ± patient, as well as the results of predicting the probable num-
0,52, i.e. only by 1,3 times. In all the age groups surveyed, ber of caries cavities, based on the obtained dental indices.

1502
 APPLICATION OF NEURAL NETWORK TECHNOLOGIES IN THE DENTAL CARIES FORECAST

DISCUSSION REFERENCES
As of today, there is a large number of computer software to 1. Rethinavel Subramanian A. Neural Philosophy in Medical Applications.
predict dental caries, most of which are based on standard- Int. Journal of Engineering Research and Applications. 2014;4(1):237-
ized multifactorial questionnaires [16]. The most famous of 241.
them are NusCra (National University of Singapore Caries 2. Park W.J., Park J.B. History and application of artificial neural networks
Risk Assessment), CARIOGRAM, PreViser, ADA (caries in dentistry. Eur J Dent. 2018;12(4): 594–601. doi:10.4103/ejd.
risk assessment by the American Dental Association), ejd_325_18.
CAMBRA (Caries Management By Risk Assessment), CAT 3. Hwang J.J., Jung Y.H., Cho B.H. et al. An overview of deep learning in
(America Academy of Pediatric Dentistry’s Caries Assess- the field of dentistry. Imaging Sci Dent. 2019;49(1):1-7. doi: 10.5624/
ment Tool). Each software takes into account nine to twelve isd.2019.49.1.1.
factors, among which, besides clinical, are socio-demograph- 4. Nindrea R.D., Aryandono T., Lazuardi L. et al. Diagnostic Accuracy
ic (age, ethnicity), behavioral factors (diet) and others. of Different Machine Learning Algorithms for Breast Cancer Risk
One of the most widely used software in the world, called Calculation: a Meta-Analysis. Asian Pac J Cancer Prev. 2018;19(7):1747-
“CARIOGRAM”, is based on the multifactorial analysis of 1752.
factors that, from the developers’ point of view, are leading 5. Itchhaporia D., Snow P.B., Almassy R.J. et al. Artificial neural
in the caries development [17]. The software takes into networks: current status in cardiovascular medicine. J Am Coll Cardiol.
account ten factors, including the salivation rate, the buffer 1996;28(2):515-21.
capacity of the oral fluid, the content of lactobacilli and 6. Han S.H., Kim K.W., Kim S. et al. Artificial Neural Network: Understanding
streptococci in it, etc. Identifying some of them requires the Basic Concepts without Mathematics. Dement Neurocogn Disord.
laboratory testing, additional equipment, and therefore 2018;17(3):83-89. doi: 10.12779/dnd.2018.17.3.83.
unnecessary financial and time costs. The significance of 7. Frencken J.E., Sharma P., Stenhouse L. et al. Global epidemiology of
each factor is conditionally estimated from 0 to 3 points, dental caries and severe periodontitis – a comprehensive review. J Clin
without relying on any quantitative criteria, which should Periodontol. 2017;44(18):94-105. doi: 10.1111/jcpe.12677.
certainly be considered subjective enough. 8. Kassebaum N.J., Smith A.G.C., Bernabé E. et al. Global, Regional, and
The result is a graphical representation of a sectoral National Prevalence, Incidence, and Disability-Adjusted Life Years for
diagram including five parts of different colors and only Oral Conditions for 195 Countries, 1990-2015: A Systematic Analysis
demonstrating the influence of etiological factors. The for the Global Burden of Diseases, Injuries, and Risk Factors. J Dent Res.
risk of developing caries, according to the image, is rated 2017;96(4):380-387. doi: 10.1177/0022034517693566.
as high, medium or low, but the percentages given do not 9. Bezvushko E.V., Lahoda L.S. Stan tverdykh tkanyn zubiv u ditei mista
answer the question of the number of possible dental caries Lutska [Condition of hard teeth tissues in children of Lutsk]. Aktualni
lesions in a particular patient over time. problemy suchasnoi medytsyny: Visnyk ukrainskoi medychnoi
The developed “CariesPro” computer software does not stomatolohichnoi akademii. 2017;17(58):232-235. (in Ukrainian).
have the above disadvantages. In order to perform full- 10. Z adorozhna I.V., Povorozniuk V.V. Poshyrenist ta intensyvnist
time work with this software, one needs to identify only kariiesu zubiv u ditei Ukrainy: rezultaty kliniko-epidemiolohichnoho
five, which is important, exclusively clinical indices that obstezhennia [Prevalence and intensity of dental caries in children
can be established during a single dental examination, so of Ukraine: results of clinical and epidemiological examination]. Bil.
that additional examinations and visits are not required. Suhloby. Khrebet. 2014;1-2(13-14):26-29. (in Ukrainian).
In addition, the software takes into account the exact fig- 11. U dod O.A., Voronina H.S. Suchasni tendentsii ta pidkhody do
ures regarding the condition of the teeth and oral hygiene, prohnozuvannia kariiesu zubiv [Modern trends and approaches to
which are specific to the particular individual. predicting dental caries]. Novyny stomatolohii. 2019;1(98):S. 75-79.
The final index in working with the “CariesPro” software (in Ukrainian).
is the prediction of a certain number of carious cavities 12. Brudevold F., McCann H.G. Enamel solubility tests and their significance
probability after a certain time, which permits to plan in regard to dental caries. Ann N Y Acad Sci. 2006;153(1):20-51. doi:
and, if necessary, timely carry out the most individualized 10.1111/j.1749-6632.1968.tb11725.x.
measures of the selective caries prevention. “CariesPro’s” 13. Pat. 118523 Ukrainy na korysnu model. Sposib vyznachennia
high level of forecasting performance is in line with existing funktsionalnoi komponenty strukturno-funktsionalnoi kyslotostiikosti
software products, and the said software has the ability to emali zubiv [Method of determining functional component of structural
constantly learn and improve. and functional acid resistance of tooth enamel]. Udod OA, Voronina
HS-№ u201702237; zaiavl. 10.03.2017; opubl. 10.08.2017, Biul. №15.
(in Ukrainian).
CONCLUSION 14. Okushko V.R. Osnovy fiziologii zuba [Fundamentals of tooth physiology].
The resulting neural network obtained and the software Tiraspol’: Izd-vo Pridnestr. un-ta. 2005. 240 s. (in Russian).
product based on it permit to predict the development 15. Rudenko O., Bezsonov O., Romanyk O. Neural network time series
of dental caries in persons of all ages with a probability prediction based on multilayer perceptron. Development Management.
of 83.56%. Thanks to the developed form, the presented 2019;5(1):23-34. doi:10.21511/dm.5(1).2019.03.
software can be installed in a personal computer, which, 16. S enneby A., Mejare I., Sahlin N.E. et al. Diagnostic accuracy of
in turn, permits its efficient use for predicting dental caries different caries risk assessment methods. A systematic review. J Dent.
in daily clinical practice. 2015;43:1385–1393. doi: 10.1016/j.jdent.2015.10.011.

1503
Oleksandr A. Udod et al.

17. Bratthall D., Hansel-Petersson G. Cariogram – multifactorial risk Conflict of interest:

assessment model for multifactorial disease. Community Dent. Oral Authors declare no conflict of interest.
Epidemiol. 2005;33:256-264.

The present work is a fragment of the Stomatology Depart-
ment No. 1 of Donetsk national medical university “Clinical CORRESPONDING AUTHOR
and laboratory substantiation of improving technologies for Hanna S. Voronina
diagnostics, treatment, prognosis and prevention of dental Donetsk National Medical University
diseases” (state registration No. 0119 U 001447). 39 Mashinobudivnikiv blvd, 84313 Lyman, Ukraine
tel: 0509438625
ORCID and contributionship: e-mail: annavoronina2812@gmail.com
Oleksandr A. Udod: 0000-0001-6790-1936 А,E,F
Hanna S. Voronina: 0000-0002-3301-7808 В,D,E Received: 27.08.2019
Olena Yu. Ivchenkova: 0000-0003-4739-0192 C Accepted: 16.04.2020

A – Work concept and design, B – Data collection and analysis, C – Responsibility for statistical analysis,
D – Writing the article, E – Critical review, F – Final approval of the article

1504