- Email: [email protected]
Using association rule mining to identify risk factors for early childhood caries
COMM-3952; No. of Pages 7
ARTICLE IN PRESS c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
journal homepage: www.intl.elsevierhealth.com/journals/cmpb
Using association rule mining to identify risk factors for early childhood caries ˇ b , Jasmina Tusek ˇ c , Marko Knezevi´ ˇ c a, Vladimir Ivanˇcevi´c a,∗ , Ivan Tusek Salaheddin Elheshk a , Ivan Lukovi´c a a b c
University of Novi Sad, Faculty of Technical Sciences, Trg Dositeja Obradovi´ca 6, 21 000 Novi Sad, Serbia University of Novi Sad, Medical Faculty, Hajduk Veljkova 3, 21 000 Novi Sad, Serbia ˇ Private Dental Office Palmadent, Mazurani´ ceva 4, 21 131 Petrovaradin, Serbia
a r t i c l e
i n f o
a b s t r a c t
Article history:
Background and objective: Early childhood caries (ECC) is a potentially severe disease affecting
Received 9 February 2015
children all over the world. The available findings are mostly based on a logistic regression
Received in revised form
model, but data mining, in particular association rule mining, could be used to extract more
19 June 2015
information from the same data set.
Accepted 20 July 2015
Methods: ECC data was collected in a cross-sectional analytical study of the 10% sample ˇ of preschool children in the South Backa area (Vojvodina, Serbia). Association rules were
Keywords:
extracted from the data by association rule mining. Risk factors were extracted from the
Early childhood caries
highly ranked association rules.
Risk factor
Results: Discovered dominant risk factors include male gender, frequent breastfeeding (with
Association rule mining
other risk factors), high birth order, language, and low body weight at birth. Low health
Objective measure of
awareness of parents was significantly associated to ECC only in male children.
interestingness
Conclusions: The discovered risk factors are mostly confirmed by the literature, which cor-
Data mining
roborates the value of the methods. © 2015 Elsevier Ireland Ltd. All rights reserved.
1.
Introduction
Early childhood caries (ECC) is a special form of caries of the primary dentition that affects the teeth after eruption, has a rapid progression, and later leads to a number of symptoms and complications. The American Academy of Pediatric Dentists (AAPD) defined ECC as “the presence of one or more decayed (noncavitated or cavitated lesions), missing (due to caries), or filled tooth surfaces in any primary tooth in a child under the age of six” [1]. Available data, often grouped into different age ranges, indicate dramatic variation in prevalence across different parts of the world, e.g., from 1% to 12% in Australia, England, Sweden, and Finland [2,3], 19.3% in USA
∗
[4], 28.4% in northeastern Brazil [5], 29.1% in central Trinidad [6], 44% in southwestern India [7], 56.2% in Poland [8], from 59% to 94% in northern Philippines [9], and 98.9% in Canada’s northern Manitoba [10]. Although ECC is widely spread in Serbia, it has not been a frequent research topic. According to the data used in the present study, ECC prevalence in preschool children in the ˇ area (SBA) of the Province of Vojvodina, Serbia, South Backa was 30.5%. This lies in the moderate category in relation to the low prevalence reported in UK, Sweden, and Finland [5] and the high prevalence in the Middle-East [11], some African countries [12], Hispano-Americans [13], and in some regions of Australia [14]. In 1998, Vulovic´ and Carevic´ [15] reported that the ECC prevalence in three-year-old children in SBA was
Corresponding author. Tel.: +381 21 485 2441; fax: +381 21 6350727. ˇ ´ E-mail address: [email protected] (V. Ivancevi c).
http://dx.doi.org/10.1016/j.cmpb.2015.07.008 0169-2607/© 2015 Elsevier Ireland Ltd. All rights reserved.
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008
COMM-3952; No. of Pages 7
2
ARTICLE IN PRESS c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
22.07%. It appears that ECC in Serbia may be on increase, which could be linked to the rapid decrease in living standards, therapeutic approach to ECC treatment, as well as to specific demographic, psychosocial and behavioral characteristics of the environment. For these reasons, researchers need ECC models that correspond to the current epidemiological situation in the affected area. In a literature review on caries models [16], logistic regression (LR) analysis appears to be the most prominent, while the only study that featured data mining (DM) relied on classification and regression tree (CART) analysis [17]. In a comparison between LR and two DM techniques, namely CART and artificial neural networks (ANNs), Gansky [18] concluded that Knowledge Discovery and Data Mining (KDD) methods may provide advantages over traditional statistical methods in dental data. In more recent caries studies that employed DM, predictive models in Japan have included C 5.0 [19], ANNs [19], and CART [20], of which the last one has been applied in the USA as well [21]. According to the cited studies and to the best of our knowledge, DM-based approaches to ECC examination have been published for the USA and Japan (primarily concerning decision trees and ANNs), but rarely, if ever, for Serbia or Europe. For ECC in SBA, various individual risk factors were identified in a study that relied on LR [22]. The collection of the relevant data sample about ECC was a challenge because such data are scarce in Serbia. In an attempt to extract additional knowledge that could not be provided by the previously employed statistical methods, we are now performing DM, namely association rule mining (ARM), on the same data. ARM, which is a group of DM techniques used to detect associations in data, may be utilized in an ECC study to detect relationships between ECC and potential risk factors. When employing LR, it may be difficult to understand impact of an individual risk factor or interplay between multiple risk factors. Researchers typically need to formulate a hypothesis for each risk factor combination before doing formal evaluation, which may become practically infeasible even for a moderately sized set of variables. On the other hand, in DM, many patterns may be extracted in a single run, but many resulting formats are of low readability. ARM may be used to avoid these problems because it provides: • numerous readable patterns (rules) that describe interaction between variables; • more straightforward interpretation than for the LR coefficients; and • numerous interpretable measures of rule interestingness, which facilitate identification of relevant rules and rule comprehension. Nonetheless, ARM may produce an enormous number of rules, but this may be alleviated by rule pruning and ranking. Therefore, in this exploratory study, we first performed ARM to generate ECC-related association rules (ARs) and then, with the goal of uncovering ECC risk factors and their interaction, we pruned, ranked, and inspected the highly ranked association rules. More information about the used data set and ARM methods is provided in Section 2. The uncovered risk factors, as well as the comparison of our findings to
those from other studies and regions, are presented in Section 3.
2.
Material and methods
2.1.
Data Set
2.1.1.
Data collection
The aim of the data collection was to investigate potential risk factors and the prevalence of ECC, as well as the degree of severity (white spot lesion, type 1-5) among different social and ethnic groups of preschool children in SBA, Republic of Serbia. The survey was a cross-sectional analytical study of the 10% sample of preschool children in SBA, aged 13–71 months of both sexes, and different ethnicities, social status, and human environmental (urban, rural) groups. The presence or absence of ECC was recorded depending on the presence of nocavity caries (white spot lesions), or cavity caries. All primary teeth were examined and caries was recorded using WHO recognized indices DMFT and DMFS, which are calculated as the total number of decayed (D), missing due to caries (M), and filled (F) teeth (T) and surfaces (S), respectively [23]. The diagnosis and the clinical form of ECC were defined by dental check-up according to the Wyne’s modified criteria [22]. Epidemiological data of the different social and ethnic groups, as well social status, habits, attitudes, behavior and health knowledge, were obtained by the interview of the parents of the examined children through a series of closed questions (see Supplementary materials 1). Unerupted or congenitally missing teeth were excluded from the DMFT and DMFS scores. Ethical considerations: All identifiable personal information was adequately disguised in the data in order to preserve the anonymity of the individuals involved. This study was approved by the Committee of Human Research of the Medical Faculty of Novi Sad, University of Novi Sad.
2.1.2.
Data overview
In the collected data set, there are 341 individual records collected at different locations in SBA (see Supplementary materials 2). Besides the binary variable denoting ECC presence/absence, there are 35 categorical variables carefully selected by the domain experts as potential risk factors (Table 1).
2.2.
Association rule overview
An association rule (AR) in this study is of the form A ⇒ B, where A (the left-hand-side or the antecedent) and B (the right-hand-side or the consequent) are disjoint sets of attribute-value pairs (AVPs), e.g., attribute A = value A1, attribute B = value B1 ⇒ attribute C = value C1. The rule is valid for a case from the data set if all the AVPs from the rule hold true for the particular case. Support (SUPP) of a rule denotes the percentage of cases from the whole data
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008
ARTICLE IN PRESS
COMM-3952; No. of Pages 7
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
Table 1 – Explanatory variables in the ECC data set. Child-related variables Ethnicity Age Gender Serbian language Birth order Birth weight Breastfeeding Breastfeeding frequency Breastfeeding during night Bottle feeding Infant formulas Additional food sweetening Fluoride supplements Fluoride toothpaste Oral hygiene Tooth brushing Diarrhea during infancy Medical syrups First dentist visit Mother-related variables Age Marital status Ethnicity Serbian language Number of children Education level Employment status Sweets during pregnancy Fluoride supplements during pregnancy Oral health during pregnancy Health awareness Father-related variables Health awareness Family-related variables City Quality of housing Housing conditions Household monthly income
set for which the rule is valid. Confidence (CONF) of a rule indicates the same but only within the subset of cases satisfying the antecedent of the rule. Lift of a rule indicates the ratio of the observed support of the rule and the expected support if the antecedent and the consequent were independent. ARM is performed on the ECC data set featuring 341 cases and 36 categorical variables. A restriction that is imposed on the format of the generated rules is that the consequent of each rule contains a single AVP denoting either ECC presence or absence. The generated rules are ranked to facilitate the identification of relevant relationships between ECC and potential risk factors.
2.3.
Rule generation and pruning
ARM is performed using the Apriori algorithm implementation from the arules package for R [24]. We mine for two groups of rules whose consequent contains only the ECC variable: • R1 rules concerning the presence of ECC, where SUPP ≥ 0.05 and CONF ≥ 0.6; and • R2 rules concerning the absence of ECC, where SUPP ≥ 0.15 and CONF ≥ 0.6.
3
In both cases, the starting support threshold was set to 0.3 (30% of all cases), but had to be gradually decreased in order to increase the number of rules for later inspection. The support threshold for R1 rules had to be further decreased because the data set is imbalanced in favor of ECC absence, i.e., ECC is present in only 104 cases (30.5%). The confidence threshold was set to 0.6 (60%) in order to find AVPs that are more often associated with ECC (either presence or absence) than they are not, i.e., associated in at least 60% of the cases containing the evaluated AVPs and not associated in at most 40% of the same cases. The lists PR1 and PR2 were the result of pruning the lists R1 and R2, respectively. This was performed in the following manner: • Rules that are present less than it may be expected were removed, i.e., we eliminated all rules that have lift or odds ratio below one. • More general rules are favored, i.e., we removed each rule for which there is a rule that has the same or higher lift, the same consequent, but a more general antecedent (its antecedent is a proper subset of the antecedent of the removed rule).
2.4.
Rule Ranking
As ARM usually results in a large number of association rules, the notion of rule interestingness [25,26] is important when evaluating the generated ARs. In order to evaluate the PR1 and PR2 rules, we combined several objective measures of rule interestingness (OMORI) [27] using average ranking for rules [28], and ranked PR1 and PR2 rules separately. In this manner, we obtained two ranked rule lists: RPR1 (ranked PR1) and RPR2 (ranked PR2). Out of many OMORIs, we considered five: support, conviction, phi, cosine, and odds ratio. The usefulness of individual OMORIs, as indicated by their agreement with expert evaluation, may depend on the domain of application and the actual topic [25]. As there is not a set of recommended measures, we had initially calculated values for 16 common OMORIs, but, owing to strong correlation between various measures [29], had to exclude the highly correlated measures through correlogram analysis. The resulting five measures correspond to relatively well-separated groups of the measure clusters identified in [26]. When there were several candidates for exclusion during correlogram analysis, we tended to keep measures with “good” properties. For instance, cosine satisfies the null addition property, i.e., the addition of new unrelated cases should not hide the previously discovered patterns. Furthermore, odds ratio is a scaling invariant measure, i.e., if data are scaled, the measure value should not change.
3.
Results and discussion
Rule generation resulted in 18 R1 rules and 88142 R2 rules. After pruning, there remained 15 PR1 rules and 6306 PR2 rules. In Table 2, we present all AVPs featured in the antecedents of the PR1 rules. One AVP may appear in multiple rules and one
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008
ARTICLE IN PRESS
COMM-3952; No. of Pages 7
4
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
Table 2 – AVPs in PR1 antecedents. AVP
Frequency in PR1 rules (N = 15)
General AVPs (father only) Father health awareness = low General AVPs (mother only) Mother Serbian language = reading and writing Mother employment status = unemployed Marital status = married General AVPs (child only) Child gender = male Child age = 3 years Birth order ≥ 3rd Feeding-related AVPs (child only) Medical syrups > 5 times a year Breastfeeding frequency > 8 times a day Breastfeeding during night = yes Fluoride-related AVPs (child and mother) Child fluoride supplements = no Fluoride supplements during pregnancy = no Child fluoride toothpaste = no
11 4 3 2 1 1 1 9 2 1 6 2 1
3.2.
Table 3 – The top three RPR1 rules. No.
Rule
1
Child gender = male, father health awareness = low ⇒ ECC = yes Breastfeeding frequency> 8 times a day, child fluoride supplements = no, medical syrups> 5 times a year ⇒ ECC = yes Child age = 3 years, birth order ≥ 3rd ⇒ ECC = yes
2
3
Support
Confidence
Lift
Average rank
0.07
0.64
2.09
3.3
0.05
0.72
2.36
3.6
0.06
0.66
2.15
4.6
rule may contain multiple AVPs in its antecedent. After rule ranking, we obtain the RPR1 and RPR2 lists. We discuss the top three RPR1 rules (Table 3) in the following subsections. The top three RPR2 rules share three child-related AVPs in their antecedents: understands and speaks Serbian, born as the first one, and birth weight greater than 2500 g. Other AVPs include comfortable housing, mother employed, and mother occasionally consuming sweets during pregnancy. As opposed to the RPR1 rules, the RPR2 rules have higher support (from 0.27 to 0.30) and confidence (greater than 80%), but lower lift (about 1.2), i.e., they cover more cases with more confidence but they are not far from the expected.
3.1.
to the father’s role model in the family, as it was shown that father’s dental hygiene habits and caries history, as well as the mother’s, are related to the greater than zero DMFT score of the children [30]. If we evaluate the relationship between father’s health awareness and child’s ECC using the 2 test, we obtain 2 (2, N = 191) = 16.68 (p < .001) for male children, and 2 (2, N = 150) = 2.83 (p = .24) for female children. On the other hand, when the same test is used to evaluate the relationship between mother’s health awareness and child’s ECC, the results are 2 (2, N = 191) = 10.70 (p = .005) for male children, and 2 (2, N = 150) = 5.64 (p = .06) for female children. At the significance level p < 0.01, the only significant relationship in this context is between the child’s male gender and the health awareness of the parent, irrespective of the parent gender. The observed higher frequency of ECC in male children coincides with the results of Croatian [31] and Brazilian [32] researchers.
Gender
Rule 1 (Table 3) indicates that the low health awareness of the father is related to ECC in male children. This could be due
Feeding
In Rule 2 (Table 3), frequent breastfeeding is combined with the lack of fluoride supplements and frequent consumption of medical syrups. The significance of fluoride in this context is well supported by previous research [33]. Moreover, we observed direct correlation between ECC and the frequent use of medical syrups (more than 5 times per year). Some medical syrups could lead to ECC [34]. The use of the inappropriate acidic food may undoubtedly cause the erosion on the tooth enamel. However, in our study, an important factor is the high concentration of saccharose in medical syrups still produced in Serbia. This implies that mutans streptococci and lactobacilli as the principle bacteria which cause ECC take saccharose as a substrate to produce acid resulting in enamel demineralization and dentin protein matrix disintegration. Syrups contain a high percentage of sugar, which is a caries risk factor in children with chronic diseases [35,36]. There are studies that indicate correlation between nursing and ECC development [37], but the general relation between frequent breastfeeding and ECC remains controversial [2]. Our study indicates that the middle form of disease, as well as the severe form with or without complications, was considerably more frequent for prolonged nursing. In this context, Hallet and O’Rourke [14] pointed out that lactation after 12 months could expose children to higher frequency of disease and to severe forms of ECC. Literature data [38] showed that mutans streptococci could increase the lactose fermentation with frequent human milk contact, which could lead to the possibility of caries development on the deciduous teeth as the result of prolonged nursing. In Turkey [36], a considerably higher ECC prevalence was reported for children who experienced prolonged nursing (after 37 months). On the other hand, human milk may influence the pH dental plaque values and reduced dissolubility of the enamel by decreasing the process of enamel remineralization [38]. It is also important to emphasize that IgA in human milk inhibits the growth of the great amount of bacteria including mutans streptococci [39]. In addition to this, human milk contains a complex of the defensive and protecting system that can reduce the risk of child illness and, consequently, it has an indirect influence
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008
ARTICLE IN PRESS
COMM-3952; No. of Pages 7
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
on ECC prevention by decreasing the therapeutic use of sugary medical syrups, which are still produced in Serbia. The higher frequency of ECC in the children breastfed for more than 12 months can be linked with qualitative and quantitative changing of the isoenzyme complex of the lactodehydrogenase in children saliva [40] and with the quick drop of IgA in human milk [41]. With respect to this topic, further biochemical and immunological studies are required. In general, human milk has a clinically negligible decay potential, except when there is a habit of frequent lactation, especially at night, when the salivary protective factors are reduced, or during children illnesses, or when it is necessary to use medicaments that can cause decrease in salivary flow (xerostomia, Mikulicz’s syndrome, etc.) [38].
3.3.
Child birth order and age
In Rule 3 (Table 3), it appears that the third and every next born child in a family has a higher risk for ECC than the previous children, which is similar to the findings presented in [14]. Our study also showed a direct correlation between child’s age and prevalence of ECC, which was detected in other studies as well [23,36]. As the number of teeth was increasing due to age, it was logical to expect an increase of caries, as well as the severity of disease in environmentally unchangeable conditions such as eating habits, oral hygiene, and fluoride use.
3.4.
Language
In the top three RPR2 rules (ECC absence), there is an AVP about child’s lack of fluency in Serbian, the official language of the country. In the last official census from 2011, 25 categories for nationality were defined and the multiethnic population of SBA (Province of Vojvodina) was pointed out [42]. The highest prevalence was noted in Roma children, followed by Ruthenian, Slovakian, and other nationalities who, according to our data, did not speak Serbian fluently. The spoken language is the basis of interpersonal communication, education, socialization, and equal participation in all spheres of social life. Considering these facts, there is a need for understanding the problem of multilingualism of the minority population groups living in Vojvodina. In our study, we noted that children who could not understand Serbian, apart from ethnicity, were more significantly exposed to ECC as opposed to children who could understand and speak Serbian. This translates to a more difficult approach to health information and mass media (television, radio, newspapers) and consequently to a lower level of health education, and increased possibility for ECC occurrence. As a result of the language barriers, the low level of understanding between parent–child–dentist may lead to distrust and feeling of discrimination among parents themselves and their children as well [43].
3.5.
Child birth weight
The prevalence of the moderate, middle and severe form of ECC with complications was higher in children who were born with body mass less than 2500 g as opposed to children of
5
normal weight at birth. The relationship between ECC and low birth weight was also detected in [44]. Prematurely born children mostly had lower birth weight and accordingly a significantly higher prevalence of linear enamel defects (LEDs) and other developmental anomalies [45]. They were also more predisposed to frequent diseases and rapid progression of ECC. In [46], it was stated that teeth with LEDs were more affected with ECC when compared to healthy teeth. Davies [47] explained that the teeth with LEDs are more predisposed to caries because they had rough surface with the presence of the expressed recess on the enamel which could increase plaque accumulation and mutans streptococci colonization. The studies of Seow et al. [48] and Douglass et al. [49] also indicated the significance of the LEDs as a predisposed factor for ECC and the high prevalence of this defect of up to 62% in prematurely born children with very low birth weight.
4.
Conclusion
The described methodology could be useful in ECC research because the resulting patterns include associations of potential risk factors. As ECC is a multifactorial disease, researchers need to combine factors to explain the risk of ECC appearance. Researchers have attempted to expand basic microbiological models for ECC development, and to include various social, demographic and behavioral factors [50] such as ethnicity, family income, maternal education level, family status and parental knowledge, as well as inappropriate nursing habit, since the reason for disease appearance is still unknown. The main strengths of our study when compared to other DM-based studies on ECC [19–21] are: (i) examination of a different set of potential risk factors; (ii) usage of different techniques, namely ARM; and (iii) better readability of the results. Moreover, as opposed to using DM for ECC prediction, which is the main goal of other studies, we focus on ECC description. The used approach may lead to a list of potential risk factors for the analyzed environment. We identified the following ECC-related patterns: (i) male children; (ii) male children whose parents are not well informed about dental health; (iii) children who were frequently breastfed; (iv) third or subsequent child in the family; (v) lack of fluency in Serbian (the official language); and (vi) low child’s body weight at birth. Contrary to the expectations, the relationship between parent health awareness and ECC [30] was significant only in male children. The sometimes-questioned relationship between breastfeeding and ECC [37] seems to hold true when breastfeeding is frequent and coupled with other factors. As discussed in Section 3, the remaining risk factors are confirmed or recognized in other studies [14,32,43,44]. In our future work, we plan to formally model the domain knowledge about ECC, and use it during rule selection and interpretation. We also intend to experiment with summarization of ARs into a decision tree and check if that would lead to understandable predictive models.
Conflict of interest None.
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008
COMM-3952; No. of Pages 7
6
ARTICLE IN PRESS c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
Acknowledgement The research presented in this paper was supported by the Ministry of Education, Science, and Technological Development of the Republic of Serbia under Grant III-44010.
Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.cmpb. 2015.07.008.
references
[1] American Academy of Pediatric Dentists, Policy on early childhood caries (ECC): classifications, consequences, and preventive strategies, Ref. Man. 36 (2014) 50–52. [2] M.G. Gussy, E.G. Waters, O. Walsh, N.M. Kilpatrick, Early childhood caries: current evidence for aetiology and prevention, J. Paediatr. Child Health 42 (2006) 37–43. [3] A.R. Milnes, Description and epidemiology of nursing caries, J. Public Health Dent. 56 (1996) 38–50. [4] National Center for Health Statistics, Health, United States, 2013: With Special Feature on Prescription Drugs, National Center for Health Statistics, Hyattsville, MD, 2014. [5] A. Rosenblatt, P. Zarzar, Breast-feeding and early childhood caries: an assessment among Brazilian infants, Int. J. Paediatr. Dent. 14 (2004) 439–445. [6] R. Naidu, J. Nunn, A. Kelly, Socio-behavioural factors and early childhood caries: a cross-sectional study of preschool children in central Trinidad, BMC Oral Health 13 (2013) 30. [7] B. Jose, N.M. King, Early childhood caries lesions in preschool children in Kerala, India, Pediatr. Dent. 25 (2003) 594–600. [8] F. Szatko, M. Wierzbicka, E. Dybizbanska, I. Struzycka, E. Iwanicka-Frankowska, Oral health of Polish three-year-olds and mothers’ oral health-related knowledge, Community Dent. Health 21 (2004) 175–180. ˜ K. Shinada, Y. Kawaguchi, Early childhood [9] K.M.G. Carino, caries in northern Philippines, Community Dent. Oral Epidemiol. 31 (2003) 81–89. [10] R.J. Schroth, P.J. Smith, J.C. Whalen, Ch. Lekic, M.E.K. Moffatt, Prevalence of caries among preschool-aged children in a northern Manitoba community, J. Can. Dent. Assoc. 71 (2005), 27-27f. [11] L.D. Rajab, M. Hamdan, Early childhood caries and risk factors in Jordan, Community Dent. Health 19 (2002) 224–229. [12] S.N. Kiwanuka, A.N. Åstrøm, T.A. Trovik, Dental caries experience and its relationship to social and behavioural factors among 3–5-year-old children in Uganda, Int. J. Paediatr. Dent. 14 (2004) 336–346. [13] L.M. Kaste, R.H. Selwitz, R.J. Oldakowski, J. Brunelle, D.M. Winn, L.J. Brown, Coronal caries in the primary and permanent dentition of children and adolescents 1–17 years of age: United States, 1988–1991, J. Dent. Res. 75 (1996) 631–641. [14] K. Hallett, P. O’Rourke, Social and behavioural determinants of early childhood caries, Aust. Dent. J. 48 (2003) 27–33. ´ M. Carevic, ´ An infective nature of dental caries, [15] M. Vulovic, Stomatoloˇski glasnik Srbije 45 (1998) 5–9 (Serbian). [16] L. Powell, Caries prediction: a review of the literature, Community Dent. Oral Epidemiol. 26 (1998) 361–371. [17] L. Breiman, J.H. Friedman, R.A. Olshen, C.J. Stone, Classification and Regression Trees, Wadsworth, Belmont CA, 1984.
[18] S.A. Gansky, Dental data mining: potential pitfalls and practical issues, Adv. Dent. Res. 17 (2003) 109–114. [19] Y. Tamaki, Y. Nomura, S. Katsumura, A. Okada, H. Yamada, S. Tsuge, Y. Kadoma, N. Hanada, Construction of a dental caries prediction model by data mining, J. Oral Sci. 51 (2009) 61–68. [20] A. Ito, M. Hayashi, T. Hamasaki, S. Ebisu, Risk assessment of dental caries by using classification and regression trees, J. Dent. 39 (2011) 457–463. [21] H.-F. Li, Data Mining and Pattern Discovery using Exploratory and Visualization Methods for Large Multidimensional Datasets (PhD Thesis), University of Kentucky, 2013. [22] I. Tuˇsek, The Influence of Social Environment and Ethnicity on Caries Prevalence in the Early Childhood (PhD Thesis), University of Belgrade, 2009 (Serbian). [23] C.G.A. Nobile, L. Fortunato, A. Bianco, C. Pileggi, M. Pavia, Pattern and severity of early childhood caries in Southern Italy: a preschool-based cross-sectional study, BMC Public Health 14 (2014) 206. [24] M. Hahsler, B. Grün, K. Hornik, arules — a computational environment for mining association rules and frequent item sets, J. Stat. Softw. 14 (2005) 1–25. [25] M. Ohsaki, H. Abe, S. Tsumoto, H. Yokoi, T. Yamaguchi, Evaluation of rule interestingness measures in medical knowledge discovery in databases, Artif. Intell. Med. 41 (2007) 177–196. [26] C. Tew, C. Giraud-Carrier, K. Tanner, S. Burton, Behavior-based clustering and analysis of interestingness measures for association rule mining, Data Min. Knowl. Discov. 28 (2014) 1004–1045. [27] L. Geng, H.J. Hamilton, Interestingness measures for data mining: a survey, ACM Comput. Surv. 38 (2006), Article 9. [28] D.R. Carvalho, A.A. Freitas, N. Ebecken, Evaluating the correlation between objective rule interestingness measures and real human interest, in: 2005 9th European Conference on Principles and Practice of Knowledge Discovery in Databases, 2005, pp. 453–461. [29] P.-N. Tan, V. Kumar, J. Srivastava, Selecting the right objective measure for association analysis, Inf. Syst. 29 (2004) 293–313. [30] M.L. Mattila, P. Rautava, M. Sillanpää, P. Paunio, Caries in five-year-old children and associations with family-related factors, J. Dent. Res. 79 (2000) 875–881. ´ ´ H. Juric, ´ W. Dukic, ´ D. Glavina, Factors [31] O. Lulic-Duki c, predisposing to early childhood caries (ECC) in children of pre-school age in the city of Zagreb, Croatia, Coll. Antropol. 25 (2001) 297–302. [32] S.M. Maciel, W. Marcenes, R.G. Watt, A. Sheiham, The relationship between sweetness preference and dental caries in mother/child pairs from Maringá-Pr, Brazil, Int. Dent. J. 51 (2001) 83–88. [33] S.M. Adair, Evidence-based use of fluoride in contemporary pediatric dental practice, Pediatr. Dent. 28 (2006) 133–142. [34] A.S. Siu, F.C. Chu, H.K. Yip, Cough syrup addiction and rampant caries: a report of two cases, Prim. Dent. Care 9 (2002) 27–30. [35] E.J. McDerra, M.A. Pollard, M.E. Curzon, The dental status of asthmatic British school children, Pediatr. Dent. 20 (1998) 281–287. [36] S. Ölmez, M. Uzamis¸, G. Erdem, Association between early childhood caries and clinical, microbiological, oral hygiene and dietary variables in rural Turkish children, Turk. J. Pediatr. 45 (2003) 231–236. [37] T. Kato, T. Yorifuji, M. Yamakawa, S. Inoue, K. Saito, H. Doi, I. Kawachi, Association of breast feeding with early childhood dental caries: Japanese population-based study, BMJ Open 5 (2015) e006982.
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008
COMM-3952; No. of Pages 7
ARTICLE IN PRESS c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
[38] D. Birkhed, T. Imfeld, S. Edwardsson, pH changes in human dental plaque from lactose and milk before and after adaptation, Caries Res. 27 (1993) 43–50. [39] L.M. Calvano, Immunological power of mother’s milk, in: M.R. de Carvalho, R.N. Tamez (Eds.), Amamentac¸ão: Bases Científicas para a Prática Profissional, Guanabara Koogan, Rio de Janeiro, 2002, pp. 88–95 (Portuguese). [40] J. Bai, Q. Zhou, Z. Bao, X. Li, M. Qin, Comparison of salivary proteins between children with early childhood caries and children without caries, Zhonghua Kou Qiang Yi Xue Za Zhi 42 (2007) 21–23 (Chinese). [41] A.-K. Holm, Diet and caries in high-risk groups in developed and developing countries, Caries Res. 24 (1990) 44–52. [42] The Statistical Office of the Republic of Serbia, The Republic of Serbia 2011 Census of population, households and dwellings in the Republic of Serbia, 2012. Retrieved from www.stat.gov.rs (accessed 15.06.15). [43] M.S. Amin, M. Elyasi, R.J. Schroth, A. Azarpazhooh, S. Compton, L. Keenan, R. Wolfe, Improving the oral health of young children of newcomer families: a forum for community members, researchers, and policy-makers, J. Can. Dent. Assoc. 80 (2014) e64.
7
[44] V.E. dos Santos Junior, R.M. de Sousa, M.C. Oliveira, A.F. de Caldas Junior, A. Rosenblatt, Early childhood caries and its relationship with perinatal, socioeconomic and nutritional risks: a cross sectional study, BMC Oral Health 14 (2014) 47. [45] P. Prakash, P. Subramaniam, B.H. Durgesh, S. Konde, Prevalence of early childhood caries and associated risk factors in preschool children of urban Bangalore, India: a cross-sectional study, Eur. J. Dent. 6 (2012) 141–152. [46] P.W. Caufield, Y. Li, T.G. Bromage, Hypoplasia-associated severe early childhood caries-a proposed definition, J. Dent. Res. 91 (2012) 544–550. [47] G.N. Davies, Early childhood caries — a synopsis, Community Dent. Oral Epidemiol. 26 (1998) 106–116. [48] W.K. Seow, C. Humphrys, D.I. Tudehope, Increased prevalence of developmental dental defects in low birth-weight, prematurely born children: a controlled study, Pediatr. Dent. 9 (1987) 221–225. [49] J.M. Douglass, A.B. Douglass, H.J. Silk, A practical guide to infant oral health, Am. Fam. Physician 70 (2004) 2113–2120. [50] S.T. Reisine, W. Psoter, Socioeconomic status and selected behavioral determinants as risk factors for dental caries, J. Dent. Educ. 65 (2001) 1009–1016.
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008
ARTICLE IN PRESS c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
journal homepage: www.intl.elsevierhealth.com/journals/cmpb
Using association rule mining to identify risk factors for early childhood caries ˇ b , Jasmina Tusek ˇ c , Marko Knezevi´ ˇ c a, Vladimir Ivanˇcevi´c a,∗ , Ivan Tusek Salaheddin Elheshk a , Ivan Lukovi´c a a b c
University of Novi Sad, Faculty of Technical Sciences, Trg Dositeja Obradovi´ca 6, 21 000 Novi Sad, Serbia University of Novi Sad, Medical Faculty, Hajduk Veljkova 3, 21 000 Novi Sad, Serbia ˇ Private Dental Office Palmadent, Mazurani´ ceva 4, 21 131 Petrovaradin, Serbia
a r t i c l e
i n f o
a b s t r a c t
Article history:
Background and objective: Early childhood caries (ECC) is a potentially severe disease affecting
Received 9 February 2015
children all over the world. The available findings are mostly based on a logistic regression
Received in revised form
model, but data mining, in particular association rule mining, could be used to extract more
19 June 2015
information from the same data set.
Accepted 20 July 2015
Methods: ECC data was collected in a cross-sectional analytical study of the 10% sample ˇ of preschool children in the South Backa area (Vojvodina, Serbia). Association rules were
Keywords:
extracted from the data by association rule mining. Risk factors were extracted from the
Early childhood caries
highly ranked association rules.
Risk factor
Results: Discovered dominant risk factors include male gender, frequent breastfeeding (with
Association rule mining
other risk factors), high birth order, language, and low body weight at birth. Low health
Objective measure of
awareness of parents was significantly associated to ECC only in male children.
interestingness
Conclusions: The discovered risk factors are mostly confirmed by the literature, which cor-
Data mining
roborates the value of the methods. © 2015 Elsevier Ireland Ltd. All rights reserved.
1.
Introduction
Early childhood caries (ECC) is a special form of caries of the primary dentition that affects the teeth after eruption, has a rapid progression, and later leads to a number of symptoms and complications. The American Academy of Pediatric Dentists (AAPD) defined ECC as “the presence of one or more decayed (noncavitated or cavitated lesions), missing (due to caries), or filled tooth surfaces in any primary tooth in a child under the age of six” [1]. Available data, often grouped into different age ranges, indicate dramatic variation in prevalence across different parts of the world, e.g., from 1% to 12% in Australia, England, Sweden, and Finland [2,3], 19.3% in USA
∗
[4], 28.4% in northeastern Brazil [5], 29.1% in central Trinidad [6], 44% in southwestern India [7], 56.2% in Poland [8], from 59% to 94% in northern Philippines [9], and 98.9% in Canada’s northern Manitoba [10]. Although ECC is widely spread in Serbia, it has not been a frequent research topic. According to the data used in the present study, ECC prevalence in preschool children in the ˇ area (SBA) of the Province of Vojvodina, Serbia, South Backa was 30.5%. This lies in the moderate category in relation to the low prevalence reported in UK, Sweden, and Finland [5] and the high prevalence in the Middle-East [11], some African countries [12], Hispano-Americans [13], and in some regions of Australia [14]. In 1998, Vulovic´ and Carevic´ [15] reported that the ECC prevalence in three-year-old children in SBA was
Corresponding author. Tel.: +381 21 485 2441; fax: +381 21 6350727. ˇ ´ E-mail address: [email protected] (V. Ivancevi c).
http://dx.doi.org/10.1016/j.cmpb.2015.07.008 0169-2607/© 2015 Elsevier Ireland Ltd. All rights reserved.
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008
COMM-3952; No. of Pages 7
2
ARTICLE IN PRESS c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
22.07%. It appears that ECC in Serbia may be on increase, which could be linked to the rapid decrease in living standards, therapeutic approach to ECC treatment, as well as to specific demographic, psychosocial and behavioral characteristics of the environment. For these reasons, researchers need ECC models that correspond to the current epidemiological situation in the affected area. In a literature review on caries models [16], logistic regression (LR) analysis appears to be the most prominent, while the only study that featured data mining (DM) relied on classification and regression tree (CART) analysis [17]. In a comparison between LR and two DM techniques, namely CART and artificial neural networks (ANNs), Gansky [18] concluded that Knowledge Discovery and Data Mining (KDD) methods may provide advantages over traditional statistical methods in dental data. In more recent caries studies that employed DM, predictive models in Japan have included C 5.0 [19], ANNs [19], and CART [20], of which the last one has been applied in the USA as well [21]. According to the cited studies and to the best of our knowledge, DM-based approaches to ECC examination have been published for the USA and Japan (primarily concerning decision trees and ANNs), but rarely, if ever, for Serbia or Europe. For ECC in SBA, various individual risk factors were identified in a study that relied on LR [22]. The collection of the relevant data sample about ECC was a challenge because such data are scarce in Serbia. In an attempt to extract additional knowledge that could not be provided by the previously employed statistical methods, we are now performing DM, namely association rule mining (ARM), on the same data. ARM, which is a group of DM techniques used to detect associations in data, may be utilized in an ECC study to detect relationships between ECC and potential risk factors. When employing LR, it may be difficult to understand impact of an individual risk factor or interplay between multiple risk factors. Researchers typically need to formulate a hypothesis for each risk factor combination before doing formal evaluation, which may become practically infeasible even for a moderately sized set of variables. On the other hand, in DM, many patterns may be extracted in a single run, but many resulting formats are of low readability. ARM may be used to avoid these problems because it provides: • numerous readable patterns (rules) that describe interaction between variables; • more straightforward interpretation than for the LR coefficients; and • numerous interpretable measures of rule interestingness, which facilitate identification of relevant rules and rule comprehension. Nonetheless, ARM may produce an enormous number of rules, but this may be alleviated by rule pruning and ranking. Therefore, in this exploratory study, we first performed ARM to generate ECC-related association rules (ARs) and then, with the goal of uncovering ECC risk factors and their interaction, we pruned, ranked, and inspected the highly ranked association rules. More information about the used data set and ARM methods is provided in Section 2. The uncovered risk factors, as well as the comparison of our findings to
those from other studies and regions, are presented in Section 3.
2.
Material and methods
2.1.
Data Set
2.1.1.
Data collection
The aim of the data collection was to investigate potential risk factors and the prevalence of ECC, as well as the degree of severity (white spot lesion, type 1-5) among different social and ethnic groups of preschool children in SBA, Republic of Serbia. The survey was a cross-sectional analytical study of the 10% sample of preschool children in SBA, aged 13–71 months of both sexes, and different ethnicities, social status, and human environmental (urban, rural) groups. The presence or absence of ECC was recorded depending on the presence of nocavity caries (white spot lesions), or cavity caries. All primary teeth were examined and caries was recorded using WHO recognized indices DMFT and DMFS, which are calculated as the total number of decayed (D), missing due to caries (M), and filled (F) teeth (T) and surfaces (S), respectively [23]. The diagnosis and the clinical form of ECC were defined by dental check-up according to the Wyne’s modified criteria [22]. Epidemiological data of the different social and ethnic groups, as well social status, habits, attitudes, behavior and health knowledge, were obtained by the interview of the parents of the examined children through a series of closed questions (see Supplementary materials 1). Unerupted or congenitally missing teeth were excluded from the DMFT and DMFS scores. Ethical considerations: All identifiable personal information was adequately disguised in the data in order to preserve the anonymity of the individuals involved. This study was approved by the Committee of Human Research of the Medical Faculty of Novi Sad, University of Novi Sad.
2.1.2.
Data overview
In the collected data set, there are 341 individual records collected at different locations in SBA (see Supplementary materials 2). Besides the binary variable denoting ECC presence/absence, there are 35 categorical variables carefully selected by the domain experts as potential risk factors (Table 1).
2.2.
Association rule overview
An association rule (AR) in this study is of the form A ⇒ B, where A (the left-hand-side or the antecedent) and B (the right-hand-side or the consequent) are disjoint sets of attribute-value pairs (AVPs), e.g., attribute A = value A1, attribute B = value B1 ⇒ attribute C = value C1. The rule is valid for a case from the data set if all the AVPs from the rule hold true for the particular case. Support (SUPP) of a rule denotes the percentage of cases from the whole data
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008
ARTICLE IN PRESS
COMM-3952; No. of Pages 7
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
Table 1 – Explanatory variables in the ECC data set. Child-related variables Ethnicity Age Gender Serbian language Birth order Birth weight Breastfeeding Breastfeeding frequency Breastfeeding during night Bottle feeding Infant formulas Additional food sweetening Fluoride supplements Fluoride toothpaste Oral hygiene Tooth brushing Diarrhea during infancy Medical syrups First dentist visit Mother-related variables Age Marital status Ethnicity Serbian language Number of children Education level Employment status Sweets during pregnancy Fluoride supplements during pregnancy Oral health during pregnancy Health awareness Father-related variables Health awareness Family-related variables City Quality of housing Housing conditions Household monthly income
set for which the rule is valid. Confidence (CONF) of a rule indicates the same but only within the subset of cases satisfying the antecedent of the rule. Lift of a rule indicates the ratio of the observed support of the rule and the expected support if the antecedent and the consequent were independent. ARM is performed on the ECC data set featuring 341 cases and 36 categorical variables. A restriction that is imposed on the format of the generated rules is that the consequent of each rule contains a single AVP denoting either ECC presence or absence. The generated rules are ranked to facilitate the identification of relevant relationships between ECC and potential risk factors.
2.3.
Rule generation and pruning
ARM is performed using the Apriori algorithm implementation from the arules package for R [24]. We mine for two groups of rules whose consequent contains only the ECC variable: • R1 rules concerning the presence of ECC, where SUPP ≥ 0.05 and CONF ≥ 0.6; and • R2 rules concerning the absence of ECC, where SUPP ≥ 0.15 and CONF ≥ 0.6.
3
In both cases, the starting support threshold was set to 0.3 (30% of all cases), but had to be gradually decreased in order to increase the number of rules for later inspection. The support threshold for R1 rules had to be further decreased because the data set is imbalanced in favor of ECC absence, i.e., ECC is present in only 104 cases (30.5%). The confidence threshold was set to 0.6 (60%) in order to find AVPs that are more often associated with ECC (either presence or absence) than they are not, i.e., associated in at least 60% of the cases containing the evaluated AVPs and not associated in at most 40% of the same cases. The lists PR1 and PR2 were the result of pruning the lists R1 and R2, respectively. This was performed in the following manner: • Rules that are present less than it may be expected were removed, i.e., we eliminated all rules that have lift or odds ratio below one. • More general rules are favored, i.e., we removed each rule for which there is a rule that has the same or higher lift, the same consequent, but a more general antecedent (its antecedent is a proper subset of the antecedent of the removed rule).
2.4.
Rule Ranking
As ARM usually results in a large number of association rules, the notion of rule interestingness [25,26] is important when evaluating the generated ARs. In order to evaluate the PR1 and PR2 rules, we combined several objective measures of rule interestingness (OMORI) [27] using average ranking for rules [28], and ranked PR1 and PR2 rules separately. In this manner, we obtained two ranked rule lists: RPR1 (ranked PR1) and RPR2 (ranked PR2). Out of many OMORIs, we considered five: support, conviction, phi, cosine, and odds ratio. The usefulness of individual OMORIs, as indicated by their agreement with expert evaluation, may depend on the domain of application and the actual topic [25]. As there is not a set of recommended measures, we had initially calculated values for 16 common OMORIs, but, owing to strong correlation between various measures [29], had to exclude the highly correlated measures through correlogram analysis. The resulting five measures correspond to relatively well-separated groups of the measure clusters identified in [26]. When there were several candidates for exclusion during correlogram analysis, we tended to keep measures with “good” properties. For instance, cosine satisfies the null addition property, i.e., the addition of new unrelated cases should not hide the previously discovered patterns. Furthermore, odds ratio is a scaling invariant measure, i.e., if data are scaled, the measure value should not change.
3.
Results and discussion
Rule generation resulted in 18 R1 rules and 88142 R2 rules. After pruning, there remained 15 PR1 rules and 6306 PR2 rules. In Table 2, we present all AVPs featured in the antecedents of the PR1 rules. One AVP may appear in multiple rules and one
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008
ARTICLE IN PRESS
COMM-3952; No. of Pages 7
4
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
Table 2 – AVPs in PR1 antecedents. AVP
Frequency in PR1 rules (N = 15)
General AVPs (father only) Father health awareness = low General AVPs (mother only) Mother Serbian language = reading and writing Mother employment status = unemployed Marital status = married General AVPs (child only) Child gender = male Child age = 3 years Birth order ≥ 3rd Feeding-related AVPs (child only) Medical syrups > 5 times a year Breastfeeding frequency > 8 times a day Breastfeeding during night = yes Fluoride-related AVPs (child and mother) Child fluoride supplements = no Fluoride supplements during pregnancy = no Child fluoride toothpaste = no
11 4 3 2 1 1 1 9 2 1 6 2 1
3.2.
Table 3 – The top three RPR1 rules. No.
Rule
1
Child gender = male, father health awareness = low ⇒ ECC = yes Breastfeeding frequency> 8 times a day, child fluoride supplements = no, medical syrups> 5 times a year ⇒ ECC = yes Child age = 3 years, birth order ≥ 3rd ⇒ ECC = yes
2
3
Support
Confidence
Lift
Average rank
0.07
0.64
2.09
3.3
0.05
0.72
2.36
3.6
0.06
0.66
2.15
4.6
rule may contain multiple AVPs in its antecedent. After rule ranking, we obtain the RPR1 and RPR2 lists. We discuss the top three RPR1 rules (Table 3) in the following subsections. The top three RPR2 rules share three child-related AVPs in their antecedents: understands and speaks Serbian, born as the first one, and birth weight greater than 2500 g. Other AVPs include comfortable housing, mother employed, and mother occasionally consuming sweets during pregnancy. As opposed to the RPR1 rules, the RPR2 rules have higher support (from 0.27 to 0.30) and confidence (greater than 80%), but lower lift (about 1.2), i.e., they cover more cases with more confidence but they are not far from the expected.
3.1.
to the father’s role model in the family, as it was shown that father’s dental hygiene habits and caries history, as well as the mother’s, are related to the greater than zero DMFT score of the children [30]. If we evaluate the relationship between father’s health awareness and child’s ECC using the 2 test, we obtain 2 (2, N = 191) = 16.68 (p < .001) for male children, and 2 (2, N = 150) = 2.83 (p = .24) for female children. On the other hand, when the same test is used to evaluate the relationship between mother’s health awareness and child’s ECC, the results are 2 (2, N = 191) = 10.70 (p = .005) for male children, and 2 (2, N = 150) = 5.64 (p = .06) for female children. At the significance level p < 0.01, the only significant relationship in this context is between the child’s male gender and the health awareness of the parent, irrespective of the parent gender. The observed higher frequency of ECC in male children coincides with the results of Croatian [31] and Brazilian [32] researchers.
Gender
Rule 1 (Table 3) indicates that the low health awareness of the father is related to ECC in male children. This could be due
Feeding
In Rule 2 (Table 3), frequent breastfeeding is combined with the lack of fluoride supplements and frequent consumption of medical syrups. The significance of fluoride in this context is well supported by previous research [33]. Moreover, we observed direct correlation between ECC and the frequent use of medical syrups (more than 5 times per year). Some medical syrups could lead to ECC [34]. The use of the inappropriate acidic food may undoubtedly cause the erosion on the tooth enamel. However, in our study, an important factor is the high concentration of saccharose in medical syrups still produced in Serbia. This implies that mutans streptococci and lactobacilli as the principle bacteria which cause ECC take saccharose as a substrate to produce acid resulting in enamel demineralization and dentin protein matrix disintegration. Syrups contain a high percentage of sugar, which is a caries risk factor in children with chronic diseases [35,36]. There are studies that indicate correlation between nursing and ECC development [37], but the general relation between frequent breastfeeding and ECC remains controversial [2]. Our study indicates that the middle form of disease, as well as the severe form with or without complications, was considerably more frequent for prolonged nursing. In this context, Hallet and O’Rourke [14] pointed out that lactation after 12 months could expose children to higher frequency of disease and to severe forms of ECC. Literature data [38] showed that mutans streptococci could increase the lactose fermentation with frequent human milk contact, which could lead to the possibility of caries development on the deciduous teeth as the result of prolonged nursing. In Turkey [36], a considerably higher ECC prevalence was reported for children who experienced prolonged nursing (after 37 months). On the other hand, human milk may influence the pH dental plaque values and reduced dissolubility of the enamel by decreasing the process of enamel remineralization [38]. It is also important to emphasize that IgA in human milk inhibits the growth of the great amount of bacteria including mutans streptococci [39]. In addition to this, human milk contains a complex of the defensive and protecting system that can reduce the risk of child illness and, consequently, it has an indirect influence
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008
ARTICLE IN PRESS
COMM-3952; No. of Pages 7
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
on ECC prevention by decreasing the therapeutic use of sugary medical syrups, which are still produced in Serbia. The higher frequency of ECC in the children breastfed for more than 12 months can be linked with qualitative and quantitative changing of the isoenzyme complex of the lactodehydrogenase in children saliva [40] and with the quick drop of IgA in human milk [41]. With respect to this topic, further biochemical and immunological studies are required. In general, human milk has a clinically negligible decay potential, except when there is a habit of frequent lactation, especially at night, when the salivary protective factors are reduced, or during children illnesses, or when it is necessary to use medicaments that can cause decrease in salivary flow (xerostomia, Mikulicz’s syndrome, etc.) [38].
3.3.
Child birth order and age
In Rule 3 (Table 3), it appears that the third and every next born child in a family has a higher risk for ECC than the previous children, which is similar to the findings presented in [14]. Our study also showed a direct correlation between child’s age and prevalence of ECC, which was detected in other studies as well [23,36]. As the number of teeth was increasing due to age, it was logical to expect an increase of caries, as well as the severity of disease in environmentally unchangeable conditions such as eating habits, oral hygiene, and fluoride use.
3.4.
Language
In the top three RPR2 rules (ECC absence), there is an AVP about child’s lack of fluency in Serbian, the official language of the country. In the last official census from 2011, 25 categories for nationality were defined and the multiethnic population of SBA (Province of Vojvodina) was pointed out [42]. The highest prevalence was noted in Roma children, followed by Ruthenian, Slovakian, and other nationalities who, according to our data, did not speak Serbian fluently. The spoken language is the basis of interpersonal communication, education, socialization, and equal participation in all spheres of social life. Considering these facts, there is a need for understanding the problem of multilingualism of the minority population groups living in Vojvodina. In our study, we noted that children who could not understand Serbian, apart from ethnicity, were more significantly exposed to ECC as opposed to children who could understand and speak Serbian. This translates to a more difficult approach to health information and mass media (television, radio, newspapers) and consequently to a lower level of health education, and increased possibility for ECC occurrence. As a result of the language barriers, the low level of understanding between parent–child–dentist may lead to distrust and feeling of discrimination among parents themselves and their children as well [43].
3.5.
Child birth weight
The prevalence of the moderate, middle and severe form of ECC with complications was higher in children who were born with body mass less than 2500 g as opposed to children of
5
normal weight at birth. The relationship between ECC and low birth weight was also detected in [44]. Prematurely born children mostly had lower birth weight and accordingly a significantly higher prevalence of linear enamel defects (LEDs) and other developmental anomalies [45]. They were also more predisposed to frequent diseases and rapid progression of ECC. In [46], it was stated that teeth with LEDs were more affected with ECC when compared to healthy teeth. Davies [47] explained that the teeth with LEDs are more predisposed to caries because they had rough surface with the presence of the expressed recess on the enamel which could increase plaque accumulation and mutans streptococci colonization. The studies of Seow et al. [48] and Douglass et al. [49] also indicated the significance of the LEDs as a predisposed factor for ECC and the high prevalence of this defect of up to 62% in prematurely born children with very low birth weight.
4.
Conclusion
The described methodology could be useful in ECC research because the resulting patterns include associations of potential risk factors. As ECC is a multifactorial disease, researchers need to combine factors to explain the risk of ECC appearance. Researchers have attempted to expand basic microbiological models for ECC development, and to include various social, demographic and behavioral factors [50] such as ethnicity, family income, maternal education level, family status and parental knowledge, as well as inappropriate nursing habit, since the reason for disease appearance is still unknown. The main strengths of our study when compared to other DM-based studies on ECC [19–21] are: (i) examination of a different set of potential risk factors; (ii) usage of different techniques, namely ARM; and (iii) better readability of the results. Moreover, as opposed to using DM for ECC prediction, which is the main goal of other studies, we focus on ECC description. The used approach may lead to a list of potential risk factors for the analyzed environment. We identified the following ECC-related patterns: (i) male children; (ii) male children whose parents are not well informed about dental health; (iii) children who were frequently breastfed; (iv) third or subsequent child in the family; (v) lack of fluency in Serbian (the official language); and (vi) low child’s body weight at birth. Contrary to the expectations, the relationship between parent health awareness and ECC [30] was significant only in male children. The sometimes-questioned relationship between breastfeeding and ECC [37] seems to hold true when breastfeeding is frequent and coupled with other factors. As discussed in Section 3, the remaining risk factors are confirmed or recognized in other studies [14,32,43,44]. In our future work, we plan to formally model the domain knowledge about ECC, and use it during rule selection and interpretation. We also intend to experiment with summarization of ARs into a decision tree and check if that would lead to understandable predictive models.
Conflict of interest None.
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008
COMM-3952; No. of Pages 7
6
ARTICLE IN PRESS c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
Acknowledgement The research presented in this paper was supported by the Ministry of Education, Science, and Technological Development of the Republic of Serbia under Grant III-44010.
Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.cmpb. 2015.07.008.
references
[1] American Academy of Pediatric Dentists, Policy on early childhood caries (ECC): classifications, consequences, and preventive strategies, Ref. Man. 36 (2014) 50–52. [2] M.G. Gussy, E.G. Waters, O. Walsh, N.M. Kilpatrick, Early childhood caries: current evidence for aetiology and prevention, J. Paediatr. Child Health 42 (2006) 37–43. [3] A.R. Milnes, Description and epidemiology of nursing caries, J. Public Health Dent. 56 (1996) 38–50. [4] National Center for Health Statistics, Health, United States, 2013: With Special Feature on Prescription Drugs, National Center for Health Statistics, Hyattsville, MD, 2014. [5] A. Rosenblatt, P. Zarzar, Breast-feeding and early childhood caries: an assessment among Brazilian infants, Int. J. Paediatr. Dent. 14 (2004) 439–445. [6] R. Naidu, J. Nunn, A. Kelly, Socio-behavioural factors and early childhood caries: a cross-sectional study of preschool children in central Trinidad, BMC Oral Health 13 (2013) 30. [7] B. Jose, N.M. King, Early childhood caries lesions in preschool children in Kerala, India, Pediatr. Dent. 25 (2003) 594–600. [8] F. Szatko, M. Wierzbicka, E. Dybizbanska, I. Struzycka, E. Iwanicka-Frankowska, Oral health of Polish three-year-olds and mothers’ oral health-related knowledge, Community Dent. Health 21 (2004) 175–180. ˜ K. Shinada, Y. Kawaguchi, Early childhood [9] K.M.G. Carino, caries in northern Philippines, Community Dent. Oral Epidemiol. 31 (2003) 81–89. [10] R.J. Schroth, P.J. Smith, J.C. Whalen, Ch. Lekic, M.E.K. Moffatt, Prevalence of caries among preschool-aged children in a northern Manitoba community, J. Can. Dent. Assoc. 71 (2005), 27-27f. [11] L.D. Rajab, M. Hamdan, Early childhood caries and risk factors in Jordan, Community Dent. Health 19 (2002) 224–229. [12] S.N. Kiwanuka, A.N. Åstrøm, T.A. Trovik, Dental caries experience and its relationship to social and behavioural factors among 3–5-year-old children in Uganda, Int. J. Paediatr. Dent. 14 (2004) 336–346. [13] L.M. Kaste, R.H. Selwitz, R.J. Oldakowski, J. Brunelle, D.M. Winn, L.J. Brown, Coronal caries in the primary and permanent dentition of children and adolescents 1–17 years of age: United States, 1988–1991, J. Dent. Res. 75 (1996) 631–641. [14] K. Hallett, P. O’Rourke, Social and behavioural determinants of early childhood caries, Aust. Dent. J. 48 (2003) 27–33. ´ M. Carevic, ´ An infective nature of dental caries, [15] M. Vulovic, Stomatoloˇski glasnik Srbije 45 (1998) 5–9 (Serbian). [16] L. Powell, Caries prediction: a review of the literature, Community Dent. Oral Epidemiol. 26 (1998) 361–371. [17] L. Breiman, J.H. Friedman, R.A. Olshen, C.J. Stone, Classification and Regression Trees, Wadsworth, Belmont CA, 1984.
[18] S.A. Gansky, Dental data mining: potential pitfalls and practical issues, Adv. Dent. Res. 17 (2003) 109–114. [19] Y. Tamaki, Y. Nomura, S. Katsumura, A. Okada, H. Yamada, S. Tsuge, Y. Kadoma, N. Hanada, Construction of a dental caries prediction model by data mining, J. Oral Sci. 51 (2009) 61–68. [20] A. Ito, M. Hayashi, T. Hamasaki, S. Ebisu, Risk assessment of dental caries by using classification and regression trees, J. Dent. 39 (2011) 457–463. [21] H.-F. Li, Data Mining and Pattern Discovery using Exploratory and Visualization Methods for Large Multidimensional Datasets (PhD Thesis), University of Kentucky, 2013. [22] I. Tuˇsek, The Influence of Social Environment and Ethnicity on Caries Prevalence in the Early Childhood (PhD Thesis), University of Belgrade, 2009 (Serbian). [23] C.G.A. Nobile, L. Fortunato, A. Bianco, C. Pileggi, M. Pavia, Pattern and severity of early childhood caries in Southern Italy: a preschool-based cross-sectional study, BMC Public Health 14 (2014) 206. [24] M. Hahsler, B. Grün, K. Hornik, arules — a computational environment for mining association rules and frequent item sets, J. Stat. Softw. 14 (2005) 1–25. [25] M. Ohsaki, H. Abe, S. Tsumoto, H. Yokoi, T. Yamaguchi, Evaluation of rule interestingness measures in medical knowledge discovery in databases, Artif. Intell. Med. 41 (2007) 177–196. [26] C. Tew, C. Giraud-Carrier, K. Tanner, S. Burton, Behavior-based clustering and analysis of interestingness measures for association rule mining, Data Min. Knowl. Discov. 28 (2014) 1004–1045. [27] L. Geng, H.J. Hamilton, Interestingness measures for data mining: a survey, ACM Comput. Surv. 38 (2006), Article 9. [28] D.R. Carvalho, A.A. Freitas, N. Ebecken, Evaluating the correlation between objective rule interestingness measures and real human interest, in: 2005 9th European Conference on Principles and Practice of Knowledge Discovery in Databases, 2005, pp. 453–461. [29] P.-N. Tan, V. Kumar, J. Srivastava, Selecting the right objective measure for association analysis, Inf. Syst. 29 (2004) 293–313. [30] M.L. Mattila, P. Rautava, M. Sillanpää, P. Paunio, Caries in five-year-old children and associations with family-related factors, J. Dent. Res. 79 (2000) 875–881. ´ ´ H. Juric, ´ W. Dukic, ´ D. Glavina, Factors [31] O. Lulic-Duki c, predisposing to early childhood caries (ECC) in children of pre-school age in the city of Zagreb, Croatia, Coll. Antropol. 25 (2001) 297–302. [32] S.M. Maciel, W. Marcenes, R.G. Watt, A. Sheiham, The relationship between sweetness preference and dental caries in mother/child pairs from Maringá-Pr, Brazil, Int. Dent. J. 51 (2001) 83–88. [33] S.M. Adair, Evidence-based use of fluoride in contemporary pediatric dental practice, Pediatr. Dent. 28 (2006) 133–142. [34] A.S. Siu, F.C. Chu, H.K. Yip, Cough syrup addiction and rampant caries: a report of two cases, Prim. Dent. Care 9 (2002) 27–30. [35] E.J. McDerra, M.A. Pollard, M.E. Curzon, The dental status of asthmatic British school children, Pediatr. Dent. 20 (1998) 281–287. [36] S. Ölmez, M. Uzamis¸, G. Erdem, Association between early childhood caries and clinical, microbiological, oral hygiene and dietary variables in rural Turkish children, Turk. J. Pediatr. 45 (2003) 231–236. [37] T. Kato, T. Yorifuji, M. Yamakawa, S. Inoue, K. Saito, H. Doi, I. Kawachi, Association of breast feeding with early childhood dental caries: Japanese population-based study, BMJ Open 5 (2015) e006982.
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008
COMM-3952; No. of Pages 7
ARTICLE IN PRESS c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 5 ) xxx–xxx
[38] D. Birkhed, T. Imfeld, S. Edwardsson, pH changes in human dental plaque from lactose and milk before and after adaptation, Caries Res. 27 (1993) 43–50. [39] L.M. Calvano, Immunological power of mother’s milk, in: M.R. de Carvalho, R.N. Tamez (Eds.), Amamentac¸ão: Bases Científicas para a Prática Profissional, Guanabara Koogan, Rio de Janeiro, 2002, pp. 88–95 (Portuguese). [40] J. Bai, Q. Zhou, Z. Bao, X. Li, M. Qin, Comparison of salivary proteins between children with early childhood caries and children without caries, Zhonghua Kou Qiang Yi Xue Za Zhi 42 (2007) 21–23 (Chinese). [41] A.-K. Holm, Diet and caries in high-risk groups in developed and developing countries, Caries Res. 24 (1990) 44–52. [42] The Statistical Office of the Republic of Serbia, The Republic of Serbia 2011 Census of population, households and dwellings in the Republic of Serbia, 2012. Retrieved from www.stat.gov.rs (accessed 15.06.15). [43] M.S. Amin, M. Elyasi, R.J. Schroth, A. Azarpazhooh, S. Compton, L. Keenan, R. Wolfe, Improving the oral health of young children of newcomer families: a forum for community members, researchers, and policy-makers, J. Can. Dent. Assoc. 80 (2014) e64.
7
[44] V.E. dos Santos Junior, R.M. de Sousa, M.C. Oliveira, A.F. de Caldas Junior, A. Rosenblatt, Early childhood caries and its relationship with perinatal, socioeconomic and nutritional risks: a cross sectional study, BMC Oral Health 14 (2014) 47. [45] P. Prakash, P. Subramaniam, B.H. Durgesh, S. Konde, Prevalence of early childhood caries and associated risk factors in preschool children of urban Bangalore, India: a cross-sectional study, Eur. J. Dent. 6 (2012) 141–152. [46] P.W. Caufield, Y. Li, T.G. Bromage, Hypoplasia-associated severe early childhood caries-a proposed definition, J. Dent. Res. 91 (2012) 544–550. [47] G.N. Davies, Early childhood caries — a synopsis, Community Dent. Oral Epidemiol. 26 (1998) 106–116. [48] W.K. Seow, C. Humphrys, D.I. Tudehope, Increased prevalence of developmental dental defects in low birth-weight, prematurely born children: a controlled study, Pediatr. Dent. 9 (1987) 221–225. [49] J.M. Douglass, A.B. Douglass, H.J. Silk, A practical guide to infant oral health, Am. Fam. Physician 70 (2004) 2113–2120. [50] S.T. Reisine, W. Psoter, Socioeconomic status and selected behavioral determinants as risk factors for dental caries, J. Dent. Educ. 65 (2001) 1009–1016.
ˇ ´ et al., Using association rule mining to identify risk factors for early childhood caries, Comput. Please cite this article in press as: V. Ivancevi c, Methods Programs Biomed. (2015), http://dx.doi.org/10.1016/j.cmpb.2015.07.008