- Email: [email protected]
Risk assessment of dental caries by using Classification and Regression Trees
journal of dentistry 39 (2011) 457–463
available at www.sciencedirect.com
journal homepage: www.intl.elsevierhealth.com/journals/jden
Risk assessment of dental caries by using Classification and Regression Trees Ataru Ito a, Mikako Hayashi a,*, Toshimitsu Hamasaki b, Shigeyuki Ebisu a a
Department of Restorative Dentistry and Endodontology, Osaka University Graduate School of Dentistry, 1-8 Yamadaoka, Suita, Osaka 565-0871, Japan b Department of Biomedical Statistics, Osaka University Graduate School of Medicine, 2-2 Yamadaoka, Suita, Osaka 565-0871, Japan
article info
abstract
Article history:
Objectives: Being able to predict an individual’s risks of dental caries would offer a poten-
Received 7 February 2011
tially huge natural step forward toward better oral heath. As things stand, preventive
Received in revised form
treatment against caries is mostly carried out without risk assessment because there is
4 April 2011
no proven way to analyse an individual’s risk factors. The purpose of this study was to try to
Accepted 5 April 2011
identify those patients with high and low risk of caries by using Classification and Regression Trees (CART). Methods: In this historical cohort study, data from 442 patients in a general practice who met
Keywords:
the inclusion criteria were analysed. CART was applied to the data to seek a model for
Caries
predicting caries by using the following parameters according to each patient: age, number
Risk assessment
of carious teeth, numbers of cariogenic bacteria, the secretion rate and buffer capacity of
CART
saliva, and compliance with a prevention programme. The risks of caries were presented by
Cariogenic bacteria
odds ratios. Multiple logistic regression analysis was performed to confirm the results
DMFT
obtained by CART. Results: CART identified high and low risk patients for primary caries with relative odds ratios of 0.41 (95%CI: 0.22–0.77, p = 0.0055) and 2.88 (95%CI: 1.49–5.59, p = 0.0018) according the numbers of cariogenic bacteria. High and low risk patients for secondary caries were also identified with the odds ratios of 0.07 (95%CI: 0.01–0.55, p = 0.00109) and 7.00 (95%CI: 3.50– 13.98, p < 0.0001) according the numbers of bacteria and existing caries. Conclusions: Cariogenic bacteria play a leading role in the incidence of caries. CART proved effective in identifying an individual patient’s risk of caries. # 2011 Elsevier Ltd. All rights reserved.
1.
Introduction
Being able to predict an individual’s risks of dental caries would offer a vital natural step forward in oral heath. However, the key task of preventive dental treatment is generally carried out without a proper caries risk assessment, even though various modern caries diagnosis and detection systems have been developed.1–5 There is now a considerable body of research showing that some people are more
vulnerable to caries and often develop new lesions in spite of regular dental check-ups.6,7 If people with higher risks can be identified and given improved intensive preventive care, this could offer both an efficient way of promoting individual and community oral health and a more economic use of health resources. The problem is how to identify patients at high or low risk. Several researchers have attempted to establish a system for predicting an individual’s future risk of caries.8–24 Although most of them studied children, recently some tackled the
* Corresponding author. Tel.: +81 6 6879 2928; fax: +81 6 6879 2928. E-mail address: [email protected] (M. Hayashi). 0300-5712/$ – see front matter # 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.jdent.2011.04.002
458
journal of dentistry 39 (2011) 457–463
more difficult problem of adults.11,18–24 A major difficulty with adults concerns the collection of appropriate clinical data. Many adult patients have already received numbers of restorations, and sometimes their restorative treatments are still going on; assessing their risk of caries may be difficult since the quality of treatment may itself be a risk factor. Data mining is a powerful statistical method of taking data from different perspectives for summarizing and analysing into useful and practical information that will identify important relationships.25 Technically, data mining is the process of finding correlations or patterns amongst dozens of fields in large relational databases. Classification and Regression Trees (CART), which is one method of data mining, uses a decision tree technique to classify data.26 It provides a set of rules that can be applied to an unclassified dataset to predict which factors are important. Binary tree-shaped structures for tackling classification and regression problems represent sets of decisions. Fig. 1 shows an example how CART can be applied to data to enable finding a set of rules to classify them. Back in 1936, Fisher took a multivariable data set of the iris flower and discriminant analysis to distinguish the three species of iris flowers.27 The dataset consists of 50 samples from each of three species of the iris flowers (Fig. 1 upper). Four features were measured from each sample, they are the length and the
[()TD$FIG]
width of sepal and petal. Based on the combination of the four features, Fisher developed a linear discriminant model to distinguish the species from each other (Fig. 1 middle). When CART was applied to the dataset, the three iris species could be clearly classified by the certain thresholds of petal length and width (Fig. 1 lower).28 When the petal length is shorter than 2.45 cm, the flower can be categorized as Setosa. On the other hand, when the petal length is 2.45 cm or longer and the petal width is 1.75 cm or wider, the flower can be categorized as Virginica. Decision trees are formed by a collection of rules based on values of certain variables in the modelling data set. Rules are selected based on how well splits can differentiate observations based on the dependent variable. Once a rule is selected and splits a node into two, the same logic is applied to each ‘‘child’’ node. We hypothesized that CART might be effective in identifying adult patients with higher and lower risks of caries. We were fortunate to obtain good records of treatment of patients going back to 15 years. These data seemed suitable for CART analysis to identify an individual’s risk of caries. The hypothesis tested was whether patients with higher and lower caries risk could be identified through CART.
2.
Materials and methods
2.1.
Patients included
The entire database of 4012 patients in a single general practice in Osaka and registered from May 1993 to February 2008 was screened. Patients aged between 20 and 64 who had been tested for levels of cariogenic bacteria, and for flow rate and buffer capacity of saliva when completing their initial restorative and periodontal treatments were considered eligible. Patients who had one of the following conditions were excluded: who failed to complete the initially planned treatments; who could not control plaque because of physical problems; and who received restorative treatments in other clinics. Protocols were approved by the ethics committee of Osaka University.
2.2.
Caries examination
The oral conditions of all patients were examined by dental hygienists, who had been trained and calibrated with representative cases, and then double checked by a dentist. All teeth were examined by visual inspection and radiographs. Any lesion that penetrated into one-third of the dentin was considered caries serious enough to need restorative treatment.
2.3.
Fig. 1 – CART applied to the Fisher’s Iris data.
Caries risk testing
Before the caries risk assessment, each patient completed initial restorative and periodontal treatments in order to eliminate possible negative effects of inferior oral conditions. Stimulated saliva flow rate was measured after the patients chewed for 5 min on paraffin pellets. Saliva buffering capacity and levels of mutans streptococci and Lactobacilli were
459
journal of dentistry 39 (2011) 457–463
assessed with Dentobuff1 Strip, Dentocult1 SM Strip, and Dentocult1 LB kits, respectively (Orion Diagnostica, Espoo, Finland). The trained dental hygienists conducted the biological sample collection. Dental hygienists continued to be checked annually throughout the testing period to ensure that there was 85% agreement in evaluating cariogenic bacteria.
2.4.
Preventive programme
All patients who completed initial restorative and periodontal treatments were advised to visit regularly at three- to sixmonth intervals to receive preventive treatments against caries and periodontitis. Preventive treatments included education on plaque control, scaling and polishing, and fluoride application. All patients daily used a toothpaste containing fluoride with a concentration to 900 ppm; drinking water in their residential areas was not fluoridated.
2.5.
Statistical analyses
For caries risk assessment, the following parameters for each patient were considered: age, numbers of decayed, missing
[()TD$FIG]
and filled teeth (DMFT); levels of cariogenic bacteria such as mutans streptococci (SM) and Lactobacilli (LB); the flow rate and buffer capacity of saliva. In addition, each patient’s compliance to the preventive programme was categorized as follows: regular visits, sometimes delayed visits, irregular visits, and visits only when needed. Then, CART was applied to develop a caries prediction model using a set of potentially significant factors referring to results of Akaike’s Information Criterion,29 which is often used in multiple logistic regression to identify the most explicable set of free variables. Higher and lower risks of the incidence of caries were presented by odds ratios. Multiple logistic regression analyses were conducted to clarify the most important factors in the onset of caries. Before doing so, we chose potentially influential factors with significant thresholds that were determined by assessing the results of a single logistic regression. Then the results of the multiple logistic regression were compared to those of CART. Throughout this study, data for primary and secondary caries were independently analysed, since the single logistic regression analyses showed that different factors influenced primary and secondary caries.
(n=4012) (n=235)
(n=3777) (n=3158)
(n=619)
(n=177) (n=442)
Fig. 2 – The selection process for patients included in this study.
460
journal of dentistry 39 (2011) 457–463
[()TD$FIG]
Fig. 3 – Distributions of patients according to their levels of mutans streptococci (SM) and Lactobacilli (LB).
3.
Results
The way we selected the patients for this study is summarized in Fig. 2. Altogether 619 patients received bacteria and saliva testing when completing their initial treatments and were advised to revisit for preventive treatments. However, 177 patients dropped out of the preventive programme. Finally, 442 patients (130 males and 312 females) with an average age of 41.2 (SD: 12.5) were eligible. The average observation period of those patients was 1418 days (median: 1077 days). The average number of DMFT at their first visit was 15.4 (SD: 6.5, median: 16). The average secretion rate of saliva was 7.55 ml (SD: 3.63, median: 7 ml). As to the buffer capacity of saliva, 66.5% of the patients tested had high levels, 20.8% were moderate, and 12.7% were low. A check on compliance to the prevention programme showed that more than 75% of the patients attended regularly or with
[()TD$FIG]
Fig. 4 – CART for incidence of primary caries within 3 years ([TD$INLE] needed restorations [TD$INLE] did not need restorations). Three nodes were identified according to the levels of SM and LB. When the level of SM is higher than 1 T 106 CFU/ ml, the patient can be categorized as high-risk group with odds ratios of 2.88. Whilst, when the levels of SM and LB are lower than 1 T 106 CFU/ml and LB 1 T 104 CFU/ml, respectively, the patient can be categorized as low-risk group with odds ratios of 0.41.
occasional delays (Fig. 2). The distributions of levels of cariogenic bacteria in patients are shown in Fig. 3. Approximately 70% of patients showed moderate levels of SM (score 1 and 2) and the rest were categorized as low (score 0) or high (score 3) levels. More than 90% of the patients had very low (score 0) to moderate (score 2) levels of LB levels from, and less than 10% had high levels (score 3). Carious lesions that needed restorations were found in 86 patients (19.5%) within three years. Patients who suffered primary caries totalled 39 (8.8%), whilst another 47 (10.6%) developed secondary caries. Fig. 4 shows CART for the incidence of primary caries within three years. CART showed that patients with higher and lower risks of primary caries could be identified by their SM and LB levels. When the SM is 1 106 colony forming unit (CFU)/ml or higher, the patient can be categorized as high-risk with an odds ratio of 2.88 (95%CI: 1.49–5.59, p = 0.0018). On the other hand, when the levels of SM and LB are lower than 1 106 CFU/ml and LB 1 104 CFU/ml, respectively, the patient can be categorized as low-risk with an odds ratio of 0.41 (95%CI: 0.22–0.77, p = 0.0055). The sensitivity and specificity in identifying the higher risk patients were 34.0% and 84.8% (Table 1). Fig. 5 shows CART for incidence of secondary caries within three years. Patients with higher or lower risk for secondary caries were identified by DMFT and their SM and LB levels. When the DMFT is higher than 17 and the LB level is 1 104 CFU/ml or higher, the patients can be categorized as high-risk with odds ratios of 7.00 (95%CI: 3.50–13.98, p < 0.0000). Patients with DMFT of 17 or fewer and the levels of SM lower than 1 105 CFU/ml categorized as low-risk with an odds ratio of 0.07 (95%CI: 0.01–0.55, p = 0.0109). The
Table 1 – Sensitivity, specificity and related values for primary and secondary caries prediction by CART.
Sensitivity (%) Specificity (%) Positive predictive value (%) Negative predictive value (%) Proportion correctly classified (%)
Primary caries
Secondary caries
34.0 84.8 21.0 91.5 79.6
71.8 73.0 20.4 96.4 72.9
journal of dentistry 39 (2011) 457–463
[()TD$FIG]
Fig. 5 – CART for incidence of secondary caries within 3 years ([TD$INLE] needed restorations [TD$INLE] did not need restorations). Four nodes were identified according to DMFT and the levels of SM and LB. When the DMFT is higher than 17 and the LB level is 1 T 104 CFU/ml or higher, the patients can be categorized as high-risk group with odds ratios of 7.00 (95%CI: 3.50–13.89, p < 0.0000). Patients with DMFT fewer than 17 and the levels of SM lower than 1 T 105 CFU/ml categorized as low-risk with an odds ratio of 0.07 (95%CI: 0.01–0.55, p = 0.0109).
Table 2 – Results of multiple logistic regression analyses for primary and secondary caries. Variables Primary caries DMFT SM (1 106 CFU/ml) LB (1 103 CFU/ml) Compliance to prevention programme Secondary caries DMFT Saliva 5 ml/min SM (5 105 CFU/ml) LB (1 104 CFU/ml) Compliance to prevention programme
Odds ratio
95%CI
p-value
1.00 2.41 1.75 1.24
0.96–1.05 1.20–4.86 0.91–3.39 0.63–2.45
0.9203 0.0136 0.0946 0.5346
1.11 1.00 2.22 3.04 0.93
1.04–1.18 0.90–1.11 1.06–4.62 1.35–6.85 0.42–2.05
0.0010 0.9713 0.0339 0.0072 0.8557
sensitivity and specificity in identifying the higher risk patients in this model were 71.8% and 73.0% (Table 1). Results of the multiple logistic regression analyses are shown in Table 2. Patients with an SM level higher than 106 CFU/ml were vulnerable to primary caries with a odds ratio of 2.41 (95%CI: 1.20–4.86, p = 0.0136). Patients with an SM level higher than 5 105 CFU/ml and with an LB level higher than 1 104 CFU/ml were vulnerable to secondary caries with odds ratios of 2.22 (95%CI: 1.06–4.62, p = 0.0339) and 3.04 (95%CI: 1.35–6.85, p = 0.0072), respectively.
4.
Discussion
Results of the multiple regression analysis showed that SM alone was a contributing factor in the onset of primary caries: DMFT and levels of SM and LB were significant for the onset of secondary caries. These factors were also selected by CART to identify the patients with higher and lower caries risk. This
461
consistency indicates the appropriateness of using CART in helping to identify an individual’s caries risk. In primary caries, high levels of SM, of 1 106 CFU/ml or higher, are critical factors in identifying a higher risk patient. A patient found to have a lower risk was characterized by a combination of moderate levels of both SM and LB. Clinically this is significant: it means that low risk of primary caries does not necessarily mean that a patient must have extremely low levels of cariogenic bacteria. This suggests that primary caries can be controlled by a preventive programme. In secondary caries, a combination of DMFT along with LB was found to be the strongest indicator of a higher risk patient. It should be noted that the threshold of LB was low. The clinical significance is that LB probably accumulates in the marginal ditch of a restoration.30 LB is acknowledged to have low ability of adhering to enamel surfaces.31,32 This is the likely reason why many restorations accelerate the risks for secondary caries. Even low levels of LB can be a risk factor for secondary caries in a patient with multiple restorations. This finding suggests the importance of high quality restorations with good marginal adaptation to prevent secondary caries. Mjo¨r et al. claimed that secondary caries should be considered as primary caries that originated at a marginal area of a restoration.33,34 In the present study, the significant variables that affected the onset of primary and secondary caries were not exactly same in CART and the logistic regression analysis. In addition, although both SM and LB were identified as significant factors in helping to produce primary and secondary caries, the levels and roles of bacteria in the onset of these lesions were different. Patients targeted for preventive programmes need different emphasis in dealing with primary and secondary caries. A programme to prevent primary caries should be strictly applied to patients with high SM levels of more than 1 106 CFU/ml, which 17.2% of the patients of this study exhibited. On the other hand, for secondary caries, special attention should be paid to patients with more than 18 restorations in conjunction with LB levels more than 1 104 CFU/ml. Sensitivity and specificity are statistical measures of the performance of a binary classification test. Sensitivity measures the proportion of actual positives which are correctly identified as such, whilst specificity measures the proportion of negatives. A suitable caries prediction model should have both sensitivity and specificity each over 80%,35,36 or together should be greater than 160%.37 However, only a few could successfully achieve these recommendations,9,12,13,24 and no model has yet been developed to predict caries in adults. In this study, only the specificity for primary caries reached the critical 80% level. The low sensitivity for primary caries may be because its onset was masked by the previous restorative and periodontal treatments. On the other hand, the influential factors for secondary caries still existed after the initial and preventive treatments, so the both sensitivity and specificity were more than 70%. More important, the odds ratio for high-risk patients of developing secondary caries was high, illustrating the clinical value of CART. In addition, the negative predictive values for primary and secondary caries were 91.5% and 96.4%, respectively (Table 1). This means that CART could predict low risk patients effectively and precisely.
462
journal of dentistry 39 (2011) 457–463
From a clinical standpoint, identifying low-risk patients is still useful when planning how frequently they should be checked. Saliva has been considered as influential in the onset of caries.38,39 However, in the present study, neither the flow rate nor buffer capacity of saliva was identified as an influential factor. Saliva might be an effective modifier to accelerate the onset of caries in cooperation with bacteria, and might be a significant indispensable in multiple caries. Further analysis, including investigating the backgrounds of patients with multiple caries should be conducted to clarify the detailed roles of saliva. In some studies on children, their eating habits and plaque index are regarded as important factors for the onset of caries.40–44 Those factors have to be considered carefully, because they may also reflect lifestyles of patients. Neither was analysed in this study. Instead, the samples of cariogenic bacteria were collected after the initial restorative and periodontal treatments, since the quantity and quality of cariogenic bacteria can be largely affected by untreated cavities and inferior restorations. This made the analyses of inherent characteristics of cariogenic bacteria of an individual possible and meaningful. Obviously this is a study based on a single dental practice. The patients included were carefully selected from large numbers of registered ones by using rigorous criteria in order to eliminate any sampling bias. However, the conclusions from this study can only apply to motivated patients who have continuously received the preventive treatments. The validity of the results obtained in this study should be confirmed by conducting further analyses using data of patients from other clinics. The results by the CART analyses suggest that clinicians should consider the level of cariogenic bacteria and DMFT as predictive indicators of caries. For higher caries risk patients identified by the CART analyses, further intensive care in addition to an existing programme should be planned to prevent the onset of new lesions. This study showed that cariogenic bacteria play a leading role in the incidence of caries, but saliva is less indicative. CART is an effective tool in identifying the caries risk of an individual patient.
Acknowledgment This study was supported by Grants-in-Aid for Scientific Research (22390358) from the Japan Society for the Promotion of Science.
references
1. Jablonski-Momeni A, Ricketts DNJ, Stachniss V, HeinzelGutenbrunner RMM, Pieper K. Occlusal caries: evaluation of direct microscopy versus digital imaging used for two histological classification systems. Journal of Dentistry 2009;38:204–11. 2. Zhang W, McGrath C, Lo ECM. A comparison of root caries diagnosis based on visual-tactile criteria and DIAGNO dent in vivo. Journal of Dentistry 2009;38:509–13.
3. Zakian CM, Taylor AM, Ellwood PR, Pretty IA. Occlusal caries detection by using thermal imaging. Journal of Dentistry 2010;38:788–95. 4. Mitropoulos P, Rahiotis C, Stamatakis H, Kakaboura A. Diagnostic performance of the visual caries classification system ICDAS II versus radiography and micro-computed tomography for proximal caries detection: an in vitro study. Journal of Dentistry 2010;38:859–67. 5. Rindal DB, Gordan VV, Litaker MS, Bader JD, Fellows JL, Qvist V, et al. Methods dentists use to diagnose primary caries lesions prior to restorative treatment: findings from the dental PBRN. Journal of Dentistry 2010;38:1027–32. 6. Burt BA. Prevention policies in the light of the changed distribution of dental caries. Acta Odontologica Scandinavica 1998;56:179–86. 7. Nishi M, Stjernswa¨rd J, Carlsson P, Bratthall D. Caries experience of some countries and areas expressed by the Significant Caries Index. Community Dentistry and Oral Epidemiology 2002;30:296–301. 8. Pitts NB, Chestnutt IG, Evans D, White D, Chadwick B, Steele JG. The dental caries experience of children in the United Kingdom, 2003. British Dental Journal 2006;200:313–20. 9. Stamm J, Disney J, Graves R, Bohannan H, Abernathy J. The University of North Carolina caries risk assessment study I: rationale and content. Journal of Public Health Dentistry 1988;48:225–32. 10. Stewart PW, Stamm JW. Classification tree prediction models for dental caries from clinical, microbiological, and interview data. Journal of Dental Research 1991;70: 1239–51. 11. Disney JA, Graves RC, Stamm JW, Bohannan HM, Abernathy JR, Zack DD. The University of North Carolina caries risk assessment study: further developments in caries risk prediction. Community Dentistry and Oral Epidemiology 1992;20:64–75. 12. Steiner M, Helfenstein U, Marthaler TM. Dental predictors of high caries increment in children. Journal of Dental Research 1992;71:1926–33. 13. Leverett DH, Proskin HM, Featherstone JD, Adair SM, Eisenberg AD, Mundorff-Shrestha SA, et al. Caries risk assessment in a longitudinal discrimination study. Journal of Dental Research 1993;72:538–43. 14. Grindefjord M, Dahllof G, Nilsson B, Modeer T. Prediction of dental caries development in 1-year-old children. Caries Research 1995;29:343–8. 15. Powell LV. Caries risk assessment: relevance to the practitioner. Journal of the American Dental Association 1998;129:349–53. 16. Reich E, Lussi A, Newbrun E. Caries-risk assessment. International Dental Journal 1999;49:15–26. 17. Messer LB. Assessing caries risk in children. Australian Dental Journal 2000;45:10–6. 18. Hansel-Petersson G, Fure S, Bratthall D. Evaluation of computer-based caries risk assessment program in an elderly group of individuals. Acta Odontologica Scandinavica 2003;61:164–71. 19. Bader JD, Perrin NA, Maupome G, Rindel B, Rush WA. Validation of a simple approach to caries risk assessment. Journal of Public Health Dentistry 2005;65:76–81. 20. Rindal DB, Rush WA, Perrin NA, Maupome G, Bader JD. Outcomes associated with dentist’ risk assessment. Community Dentistry and Oral Epidemiology 2006;34: 381–6. 21. Ruiz MA, Montiel CJM, Almerich SJM. Evaluation of caries risk in a young adult population. Medicina Oral Patologı´a Oral Cirugı´a Bucal 2007;12:E412–8. 22. Roberts-Thomson K, Stewart JF. Risk indicators of caries experience among young adults. Australian Dental Journal 2008;53:122–7.
journal of dentistry 39 (2011) 457–463
23. Sonbul H, Al-Otaibi M, Bikhed D. Risk profile of adults with several dental restorations using the Cariogram model. Acta Odontologica Scandinavica 2008;66:351–7. 24. Gao XL, Hsu CYS, Xu Y, Hwarng HB, Loh T, Koh D. Building caries risk assessment models for children. Journal of Dental Research 2010;89:637–43. 25. Kantardzic M. Data mining: concepts, models, methods, and algorithms. New York: John Wiley & Sons; 2003. p. 43–56. 26. Freitas AA. Data mining and knowledge discovery with evolutionary algorithms. Humburg: Springer; 2002. p. 13–43. 27. Fisher RA. The use of multiple measurements in taxonomic problems. Annals of Eugenics 1936;7:179–88. 28. http://www.ism.ac.jp/lectures/20a.html. 29. Akaike H. A Bayesian analysis of the minimum AIC procedure. Annals of the Institute of Statistical Mathematics 1978;30A:9–14. 30. McGrady JA, Butcher WG, Beighton D, Switalski LM. Specific and charge interactions mediate collagen recognition by oral lactobacilli. Journal of Dental Research 1995;74:649–57. 31. Milnes A, Bowden G. The microflora associated with the developing lesions of nursing caries. Caries Research 1985;19:289–97. 32. MacPherson LMD, MacFarlane TW, Stephen KW. An intraoral appliance study of the plaque microflora associated with early enamel demineralization. Journal of Dental Research 1990;69:1712–6. 33. Mjo¨r IA. The location of clinically diagnosed secondary caries. Quintessence International 1998;29:313–7. 34. Mjo¨r IA, Toffenetti F. Secondary caries: a literature review with case reports. Quintessence International 2000;31:165–79. 35. Wilson RF, Ashley FR. Identification of caries risk in schoolchildren: salivary buffering capacity and bacterial
36. 37.
38.
39.
40.
41.
42.
43.
44.
463
counts, sugar intake and caries experience as predictors of 2-year and 3-year caries increment. British Dental Journal 1989;167:99–102. Hausen H. Caries prediction – state of the art. Community Dentistry and Oral Epidemiology 1997;25:87–96. Fleiss JL, Kingman A. Statistical management of data in clinical research. Critical Reviews in Oral Biology and Medicine 1990;1:55–66. Tenovuo J. Salivary parameters of relevance for assessing caries activity in individuals and populations. Community Dentistry and Oral Epidemiology 1997;25:82–6. Eliasson L, Carle´n A, Almsta˚hl A, Wilkstro¨m M, Lingstro¨m P. Dental plaque pH and micro-organisms during hyposalivation. Journal of Dental Research 2006;85: 334–8. Rugg-Gunn AJ, Hackett AF, Appleton DR, Jenkins GN, Eastoe JE. Relationship between dietary habits and caries increment assessed over two years in 405 English school children. Archives of Oral Biology 1984;29:983–92. Firestone AR, Mu¨hlemann HR. In vivo pH of plaque-covered and plaque-free interdental surfaces in humans following a sucrose rinse. Clinical Preventive Dentistry 1985;7:24–6. Burt BA, Eklund SA, Morgan KJ, Larkin FE, Guire KE, Brown LO, et al. The effects of sugars intake and frequency of ingestion on dental caries increment in a three-year longitudinal study. Journal of Dental Research 1988;67:1422–9. Bjertness E. The importance of oral hygiene on variation in dental caries in adults. Acta Odontologica Scandinavica 1991;49:97–102. Krasse B. The Vipeholm dental caries study: recollections and reflections 50 Years Late. Journal of Dental Research 2001;80:1785–8.
available at www.sciencedirect.com
journal homepage: www.intl.elsevierhealth.com/journals/jden
Risk assessment of dental caries by using Classification and Regression Trees Ataru Ito a, Mikako Hayashi a,*, Toshimitsu Hamasaki b, Shigeyuki Ebisu a a
Department of Restorative Dentistry and Endodontology, Osaka University Graduate School of Dentistry, 1-8 Yamadaoka, Suita, Osaka 565-0871, Japan b Department of Biomedical Statistics, Osaka University Graduate School of Medicine, 2-2 Yamadaoka, Suita, Osaka 565-0871, Japan
article info
abstract
Article history:
Objectives: Being able to predict an individual’s risks of dental caries would offer a poten-
Received 7 February 2011
tially huge natural step forward toward better oral heath. As things stand, preventive
Received in revised form
treatment against caries is mostly carried out without risk assessment because there is
4 April 2011
no proven way to analyse an individual’s risk factors. The purpose of this study was to try to
Accepted 5 April 2011
identify those patients with high and low risk of caries by using Classification and Regression Trees (CART). Methods: In this historical cohort study, data from 442 patients in a general practice who met
Keywords:
the inclusion criteria were analysed. CART was applied to the data to seek a model for
Caries
predicting caries by using the following parameters according to each patient: age, number
Risk assessment
of carious teeth, numbers of cariogenic bacteria, the secretion rate and buffer capacity of
CART
saliva, and compliance with a prevention programme. The risks of caries were presented by
Cariogenic bacteria
odds ratios. Multiple logistic regression analysis was performed to confirm the results
DMFT
obtained by CART. Results: CART identified high and low risk patients for primary caries with relative odds ratios of 0.41 (95%CI: 0.22–0.77, p = 0.0055) and 2.88 (95%CI: 1.49–5.59, p = 0.0018) according the numbers of cariogenic bacteria. High and low risk patients for secondary caries were also identified with the odds ratios of 0.07 (95%CI: 0.01–0.55, p = 0.00109) and 7.00 (95%CI: 3.50– 13.98, p < 0.0001) according the numbers of bacteria and existing caries. Conclusions: Cariogenic bacteria play a leading role in the incidence of caries. CART proved effective in identifying an individual patient’s risk of caries. # 2011 Elsevier Ltd. All rights reserved.
1.
Introduction
Being able to predict an individual’s risks of dental caries would offer a vital natural step forward in oral heath. However, the key task of preventive dental treatment is generally carried out without a proper caries risk assessment, even though various modern caries diagnosis and detection systems have been developed.1–5 There is now a considerable body of research showing that some people are more
vulnerable to caries and often develop new lesions in spite of regular dental check-ups.6,7 If people with higher risks can be identified and given improved intensive preventive care, this could offer both an efficient way of promoting individual and community oral health and a more economic use of health resources. The problem is how to identify patients at high or low risk. Several researchers have attempted to establish a system for predicting an individual’s future risk of caries.8–24 Although most of them studied children, recently some tackled the
* Corresponding author. Tel.: +81 6 6879 2928; fax: +81 6 6879 2928. E-mail address: [email protected] (M. Hayashi). 0300-5712/$ – see front matter # 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.jdent.2011.04.002
458
journal of dentistry 39 (2011) 457–463
more difficult problem of adults.11,18–24 A major difficulty with adults concerns the collection of appropriate clinical data. Many adult patients have already received numbers of restorations, and sometimes their restorative treatments are still going on; assessing their risk of caries may be difficult since the quality of treatment may itself be a risk factor. Data mining is a powerful statistical method of taking data from different perspectives for summarizing and analysing into useful and practical information that will identify important relationships.25 Technically, data mining is the process of finding correlations or patterns amongst dozens of fields in large relational databases. Classification and Regression Trees (CART), which is one method of data mining, uses a decision tree technique to classify data.26 It provides a set of rules that can be applied to an unclassified dataset to predict which factors are important. Binary tree-shaped structures for tackling classification and regression problems represent sets of decisions. Fig. 1 shows an example how CART can be applied to data to enable finding a set of rules to classify them. Back in 1936, Fisher took a multivariable data set of the iris flower and discriminant analysis to distinguish the three species of iris flowers.27 The dataset consists of 50 samples from each of three species of the iris flowers (Fig. 1 upper). Four features were measured from each sample, they are the length and the
[()TD$FIG]
width of sepal and petal. Based on the combination of the four features, Fisher developed a linear discriminant model to distinguish the species from each other (Fig. 1 middle). When CART was applied to the dataset, the three iris species could be clearly classified by the certain thresholds of petal length and width (Fig. 1 lower).28 When the petal length is shorter than 2.45 cm, the flower can be categorized as Setosa. On the other hand, when the petal length is 2.45 cm or longer and the petal width is 1.75 cm or wider, the flower can be categorized as Virginica. Decision trees are formed by a collection of rules based on values of certain variables in the modelling data set. Rules are selected based on how well splits can differentiate observations based on the dependent variable. Once a rule is selected and splits a node into two, the same logic is applied to each ‘‘child’’ node. We hypothesized that CART might be effective in identifying adult patients with higher and lower risks of caries. We were fortunate to obtain good records of treatment of patients going back to 15 years. These data seemed suitable for CART analysis to identify an individual’s risk of caries. The hypothesis tested was whether patients with higher and lower caries risk could be identified through CART.
2.
Materials and methods
2.1.
Patients included
The entire database of 4012 patients in a single general practice in Osaka and registered from May 1993 to February 2008 was screened. Patients aged between 20 and 64 who had been tested for levels of cariogenic bacteria, and for flow rate and buffer capacity of saliva when completing their initial restorative and periodontal treatments were considered eligible. Patients who had one of the following conditions were excluded: who failed to complete the initially planned treatments; who could not control plaque because of physical problems; and who received restorative treatments in other clinics. Protocols were approved by the ethics committee of Osaka University.
2.2.
Caries examination
The oral conditions of all patients were examined by dental hygienists, who had been trained and calibrated with representative cases, and then double checked by a dentist. All teeth were examined by visual inspection and radiographs. Any lesion that penetrated into one-third of the dentin was considered caries serious enough to need restorative treatment.
2.3.
Fig. 1 – CART applied to the Fisher’s Iris data.
Caries risk testing
Before the caries risk assessment, each patient completed initial restorative and periodontal treatments in order to eliminate possible negative effects of inferior oral conditions. Stimulated saliva flow rate was measured after the patients chewed for 5 min on paraffin pellets. Saliva buffering capacity and levels of mutans streptococci and Lactobacilli were
459
journal of dentistry 39 (2011) 457–463
assessed with Dentobuff1 Strip, Dentocult1 SM Strip, and Dentocult1 LB kits, respectively (Orion Diagnostica, Espoo, Finland). The trained dental hygienists conducted the biological sample collection. Dental hygienists continued to be checked annually throughout the testing period to ensure that there was 85% agreement in evaluating cariogenic bacteria.
2.4.
Preventive programme
All patients who completed initial restorative and periodontal treatments were advised to visit regularly at three- to sixmonth intervals to receive preventive treatments against caries and periodontitis. Preventive treatments included education on plaque control, scaling and polishing, and fluoride application. All patients daily used a toothpaste containing fluoride with a concentration to 900 ppm; drinking water in their residential areas was not fluoridated.
2.5.
Statistical analyses
For caries risk assessment, the following parameters for each patient were considered: age, numbers of decayed, missing
[()TD$FIG]
and filled teeth (DMFT); levels of cariogenic bacteria such as mutans streptococci (SM) and Lactobacilli (LB); the flow rate and buffer capacity of saliva. In addition, each patient’s compliance to the preventive programme was categorized as follows: regular visits, sometimes delayed visits, irregular visits, and visits only when needed. Then, CART was applied to develop a caries prediction model using a set of potentially significant factors referring to results of Akaike’s Information Criterion,29 which is often used in multiple logistic regression to identify the most explicable set of free variables. Higher and lower risks of the incidence of caries were presented by odds ratios. Multiple logistic regression analyses were conducted to clarify the most important factors in the onset of caries. Before doing so, we chose potentially influential factors with significant thresholds that were determined by assessing the results of a single logistic regression. Then the results of the multiple logistic regression were compared to those of CART. Throughout this study, data for primary and secondary caries were independently analysed, since the single logistic regression analyses showed that different factors influenced primary and secondary caries.
(n=4012) (n=235)
(n=3777) (n=3158)
(n=619)
(n=177) (n=442)
Fig. 2 – The selection process for patients included in this study.
460
journal of dentistry 39 (2011) 457–463
[()TD$FIG]
Fig. 3 – Distributions of patients according to their levels of mutans streptococci (SM) and Lactobacilli (LB).
3.
Results
The way we selected the patients for this study is summarized in Fig. 2. Altogether 619 patients received bacteria and saliva testing when completing their initial treatments and were advised to revisit for preventive treatments. However, 177 patients dropped out of the preventive programme. Finally, 442 patients (130 males and 312 females) with an average age of 41.2 (SD: 12.5) were eligible. The average observation period of those patients was 1418 days (median: 1077 days). The average number of DMFT at their first visit was 15.4 (SD: 6.5, median: 16). The average secretion rate of saliva was 7.55 ml (SD: 3.63, median: 7 ml). As to the buffer capacity of saliva, 66.5% of the patients tested had high levels, 20.8% were moderate, and 12.7% were low. A check on compliance to the prevention programme showed that more than 75% of the patients attended regularly or with
[()TD$FIG]
Fig. 4 – CART for incidence of primary caries within 3 years ([TD$INLE] needed restorations [TD$INLE] did not need restorations). Three nodes were identified according to the levels of SM and LB. When the level of SM is higher than 1 T 106 CFU/ ml, the patient can be categorized as high-risk group with odds ratios of 2.88. Whilst, when the levels of SM and LB are lower than 1 T 106 CFU/ml and LB 1 T 104 CFU/ml, respectively, the patient can be categorized as low-risk group with odds ratios of 0.41.
occasional delays (Fig. 2). The distributions of levels of cariogenic bacteria in patients are shown in Fig. 3. Approximately 70% of patients showed moderate levels of SM (score 1 and 2) and the rest were categorized as low (score 0) or high (score 3) levels. More than 90% of the patients had very low (score 0) to moderate (score 2) levels of LB levels from, and less than 10% had high levels (score 3). Carious lesions that needed restorations were found in 86 patients (19.5%) within three years. Patients who suffered primary caries totalled 39 (8.8%), whilst another 47 (10.6%) developed secondary caries. Fig. 4 shows CART for the incidence of primary caries within three years. CART showed that patients with higher and lower risks of primary caries could be identified by their SM and LB levels. When the SM is 1 106 colony forming unit (CFU)/ml or higher, the patient can be categorized as high-risk with an odds ratio of 2.88 (95%CI: 1.49–5.59, p = 0.0018). On the other hand, when the levels of SM and LB are lower than 1 106 CFU/ml and LB 1 104 CFU/ml, respectively, the patient can be categorized as low-risk with an odds ratio of 0.41 (95%CI: 0.22–0.77, p = 0.0055). The sensitivity and specificity in identifying the higher risk patients were 34.0% and 84.8% (Table 1). Fig. 5 shows CART for incidence of secondary caries within three years. Patients with higher or lower risk for secondary caries were identified by DMFT and their SM and LB levels. When the DMFT is higher than 17 and the LB level is 1 104 CFU/ml or higher, the patients can be categorized as high-risk with odds ratios of 7.00 (95%CI: 3.50–13.98, p < 0.0000). Patients with DMFT of 17 or fewer and the levels of SM lower than 1 105 CFU/ml categorized as low-risk with an odds ratio of 0.07 (95%CI: 0.01–0.55, p = 0.0109). The
Table 1 – Sensitivity, specificity and related values for primary and secondary caries prediction by CART.
Sensitivity (%) Specificity (%) Positive predictive value (%) Negative predictive value (%) Proportion correctly classified (%)
Primary caries
Secondary caries
34.0 84.8 21.0 91.5 79.6
71.8 73.0 20.4 96.4 72.9
journal of dentistry 39 (2011) 457–463
[()TD$FIG]
Fig. 5 – CART for incidence of secondary caries within 3 years ([TD$INLE] needed restorations [TD$INLE] did not need restorations). Four nodes were identified according to DMFT and the levels of SM and LB. When the DMFT is higher than 17 and the LB level is 1 T 104 CFU/ml or higher, the patients can be categorized as high-risk group with odds ratios of 7.00 (95%CI: 3.50–13.89, p < 0.0000). Patients with DMFT fewer than 17 and the levels of SM lower than 1 T 105 CFU/ml categorized as low-risk with an odds ratio of 0.07 (95%CI: 0.01–0.55, p = 0.0109).
Table 2 – Results of multiple logistic regression analyses for primary and secondary caries. Variables Primary caries DMFT SM (1 106 CFU/ml) LB (1 103 CFU/ml) Compliance to prevention programme Secondary caries DMFT Saliva 5 ml/min SM (5 105 CFU/ml) LB (1 104 CFU/ml) Compliance to prevention programme
Odds ratio
95%CI
p-value
1.00 2.41 1.75 1.24
0.96–1.05 1.20–4.86 0.91–3.39 0.63–2.45
0.9203 0.0136 0.0946 0.5346
1.11 1.00 2.22 3.04 0.93
1.04–1.18 0.90–1.11 1.06–4.62 1.35–6.85 0.42–2.05
0.0010 0.9713 0.0339 0.0072 0.8557
sensitivity and specificity in identifying the higher risk patients in this model were 71.8% and 73.0% (Table 1). Results of the multiple logistic regression analyses are shown in Table 2. Patients with an SM level higher than 106 CFU/ml were vulnerable to primary caries with a odds ratio of 2.41 (95%CI: 1.20–4.86, p = 0.0136). Patients with an SM level higher than 5 105 CFU/ml and with an LB level higher than 1 104 CFU/ml were vulnerable to secondary caries with odds ratios of 2.22 (95%CI: 1.06–4.62, p = 0.0339) and 3.04 (95%CI: 1.35–6.85, p = 0.0072), respectively.
4.
Discussion
Results of the multiple regression analysis showed that SM alone was a contributing factor in the onset of primary caries: DMFT and levels of SM and LB were significant for the onset of secondary caries. These factors were also selected by CART to identify the patients with higher and lower caries risk. This
461
consistency indicates the appropriateness of using CART in helping to identify an individual’s caries risk. In primary caries, high levels of SM, of 1 106 CFU/ml or higher, are critical factors in identifying a higher risk patient. A patient found to have a lower risk was characterized by a combination of moderate levels of both SM and LB. Clinically this is significant: it means that low risk of primary caries does not necessarily mean that a patient must have extremely low levels of cariogenic bacteria. This suggests that primary caries can be controlled by a preventive programme. In secondary caries, a combination of DMFT along with LB was found to be the strongest indicator of a higher risk patient. It should be noted that the threshold of LB was low. The clinical significance is that LB probably accumulates in the marginal ditch of a restoration.30 LB is acknowledged to have low ability of adhering to enamel surfaces.31,32 This is the likely reason why many restorations accelerate the risks for secondary caries. Even low levels of LB can be a risk factor for secondary caries in a patient with multiple restorations. This finding suggests the importance of high quality restorations with good marginal adaptation to prevent secondary caries. Mjo¨r et al. claimed that secondary caries should be considered as primary caries that originated at a marginal area of a restoration.33,34 In the present study, the significant variables that affected the onset of primary and secondary caries were not exactly same in CART and the logistic regression analysis. In addition, although both SM and LB were identified as significant factors in helping to produce primary and secondary caries, the levels and roles of bacteria in the onset of these lesions were different. Patients targeted for preventive programmes need different emphasis in dealing with primary and secondary caries. A programme to prevent primary caries should be strictly applied to patients with high SM levels of more than 1 106 CFU/ml, which 17.2% of the patients of this study exhibited. On the other hand, for secondary caries, special attention should be paid to patients with more than 18 restorations in conjunction with LB levels more than 1 104 CFU/ml. Sensitivity and specificity are statistical measures of the performance of a binary classification test. Sensitivity measures the proportion of actual positives which are correctly identified as such, whilst specificity measures the proportion of negatives. A suitable caries prediction model should have both sensitivity and specificity each over 80%,35,36 or together should be greater than 160%.37 However, only a few could successfully achieve these recommendations,9,12,13,24 and no model has yet been developed to predict caries in adults. In this study, only the specificity for primary caries reached the critical 80% level. The low sensitivity for primary caries may be because its onset was masked by the previous restorative and periodontal treatments. On the other hand, the influential factors for secondary caries still existed after the initial and preventive treatments, so the both sensitivity and specificity were more than 70%. More important, the odds ratio for high-risk patients of developing secondary caries was high, illustrating the clinical value of CART. In addition, the negative predictive values for primary and secondary caries were 91.5% and 96.4%, respectively (Table 1). This means that CART could predict low risk patients effectively and precisely.
462
journal of dentistry 39 (2011) 457–463
From a clinical standpoint, identifying low-risk patients is still useful when planning how frequently they should be checked. Saliva has been considered as influential in the onset of caries.38,39 However, in the present study, neither the flow rate nor buffer capacity of saliva was identified as an influential factor. Saliva might be an effective modifier to accelerate the onset of caries in cooperation with bacteria, and might be a significant indispensable in multiple caries. Further analysis, including investigating the backgrounds of patients with multiple caries should be conducted to clarify the detailed roles of saliva. In some studies on children, their eating habits and plaque index are regarded as important factors for the onset of caries.40–44 Those factors have to be considered carefully, because they may also reflect lifestyles of patients. Neither was analysed in this study. Instead, the samples of cariogenic bacteria were collected after the initial restorative and periodontal treatments, since the quantity and quality of cariogenic bacteria can be largely affected by untreated cavities and inferior restorations. This made the analyses of inherent characteristics of cariogenic bacteria of an individual possible and meaningful. Obviously this is a study based on a single dental practice. The patients included were carefully selected from large numbers of registered ones by using rigorous criteria in order to eliminate any sampling bias. However, the conclusions from this study can only apply to motivated patients who have continuously received the preventive treatments. The validity of the results obtained in this study should be confirmed by conducting further analyses using data of patients from other clinics. The results by the CART analyses suggest that clinicians should consider the level of cariogenic bacteria and DMFT as predictive indicators of caries. For higher caries risk patients identified by the CART analyses, further intensive care in addition to an existing programme should be planned to prevent the onset of new lesions. This study showed that cariogenic bacteria play a leading role in the incidence of caries, but saliva is less indicative. CART is an effective tool in identifying the caries risk of an individual patient.
Acknowledgment This study was supported by Grants-in-Aid for Scientific Research (22390358) from the Japan Society for the Promotion of Science.
references
1. Jablonski-Momeni A, Ricketts DNJ, Stachniss V, HeinzelGutenbrunner RMM, Pieper K. Occlusal caries: evaluation of direct microscopy versus digital imaging used for two histological classification systems. Journal of Dentistry 2009;38:204–11. 2. Zhang W, McGrath C, Lo ECM. A comparison of root caries diagnosis based on visual-tactile criteria and DIAGNO dent in vivo. Journal of Dentistry 2009;38:509–13.
3. Zakian CM, Taylor AM, Ellwood PR, Pretty IA. Occlusal caries detection by using thermal imaging. Journal of Dentistry 2010;38:788–95. 4. Mitropoulos P, Rahiotis C, Stamatakis H, Kakaboura A. Diagnostic performance of the visual caries classification system ICDAS II versus radiography and micro-computed tomography for proximal caries detection: an in vitro study. Journal of Dentistry 2010;38:859–67. 5. Rindal DB, Gordan VV, Litaker MS, Bader JD, Fellows JL, Qvist V, et al. Methods dentists use to diagnose primary caries lesions prior to restorative treatment: findings from the dental PBRN. Journal of Dentistry 2010;38:1027–32. 6. Burt BA. Prevention policies in the light of the changed distribution of dental caries. Acta Odontologica Scandinavica 1998;56:179–86. 7. Nishi M, Stjernswa¨rd J, Carlsson P, Bratthall D. Caries experience of some countries and areas expressed by the Significant Caries Index. Community Dentistry and Oral Epidemiology 2002;30:296–301. 8. Pitts NB, Chestnutt IG, Evans D, White D, Chadwick B, Steele JG. The dental caries experience of children in the United Kingdom, 2003. British Dental Journal 2006;200:313–20. 9. Stamm J, Disney J, Graves R, Bohannan H, Abernathy J. The University of North Carolina caries risk assessment study I: rationale and content. Journal of Public Health Dentistry 1988;48:225–32. 10. Stewart PW, Stamm JW. Classification tree prediction models for dental caries from clinical, microbiological, and interview data. Journal of Dental Research 1991;70: 1239–51. 11. Disney JA, Graves RC, Stamm JW, Bohannan HM, Abernathy JR, Zack DD. The University of North Carolina caries risk assessment study: further developments in caries risk prediction. Community Dentistry and Oral Epidemiology 1992;20:64–75. 12. Steiner M, Helfenstein U, Marthaler TM. Dental predictors of high caries increment in children. Journal of Dental Research 1992;71:1926–33. 13. Leverett DH, Proskin HM, Featherstone JD, Adair SM, Eisenberg AD, Mundorff-Shrestha SA, et al. Caries risk assessment in a longitudinal discrimination study. Journal of Dental Research 1993;72:538–43. 14. Grindefjord M, Dahllof G, Nilsson B, Modeer T. Prediction of dental caries development in 1-year-old children. Caries Research 1995;29:343–8. 15. Powell LV. Caries risk assessment: relevance to the practitioner. Journal of the American Dental Association 1998;129:349–53. 16. Reich E, Lussi A, Newbrun E. Caries-risk assessment. International Dental Journal 1999;49:15–26. 17. Messer LB. Assessing caries risk in children. Australian Dental Journal 2000;45:10–6. 18. Hansel-Petersson G, Fure S, Bratthall D. Evaluation of computer-based caries risk assessment program in an elderly group of individuals. Acta Odontologica Scandinavica 2003;61:164–71. 19. Bader JD, Perrin NA, Maupome G, Rindel B, Rush WA. Validation of a simple approach to caries risk assessment. Journal of Public Health Dentistry 2005;65:76–81. 20. Rindal DB, Rush WA, Perrin NA, Maupome G, Bader JD. Outcomes associated with dentist’ risk assessment. Community Dentistry and Oral Epidemiology 2006;34: 381–6. 21. Ruiz MA, Montiel CJM, Almerich SJM. Evaluation of caries risk in a young adult population. Medicina Oral Patologı´a Oral Cirugı´a Bucal 2007;12:E412–8. 22. Roberts-Thomson K, Stewart JF. Risk indicators of caries experience among young adults. Australian Dental Journal 2008;53:122–7.
journal of dentistry 39 (2011) 457–463
23. Sonbul H, Al-Otaibi M, Bikhed D. Risk profile of adults with several dental restorations using the Cariogram model. Acta Odontologica Scandinavica 2008;66:351–7. 24. Gao XL, Hsu CYS, Xu Y, Hwarng HB, Loh T, Koh D. Building caries risk assessment models for children. Journal of Dental Research 2010;89:637–43. 25. Kantardzic M. Data mining: concepts, models, methods, and algorithms. New York: John Wiley & Sons; 2003. p. 43–56. 26. Freitas AA. Data mining and knowledge discovery with evolutionary algorithms. Humburg: Springer; 2002. p. 13–43. 27. Fisher RA. The use of multiple measurements in taxonomic problems. Annals of Eugenics 1936;7:179–88. 28. http://www.ism.ac.jp/lectures/20a.html. 29. Akaike H. A Bayesian analysis of the minimum AIC procedure. Annals of the Institute of Statistical Mathematics 1978;30A:9–14. 30. McGrady JA, Butcher WG, Beighton D, Switalski LM. Specific and charge interactions mediate collagen recognition by oral lactobacilli. Journal of Dental Research 1995;74:649–57. 31. Milnes A, Bowden G. The microflora associated with the developing lesions of nursing caries. Caries Research 1985;19:289–97. 32. MacPherson LMD, MacFarlane TW, Stephen KW. An intraoral appliance study of the plaque microflora associated with early enamel demineralization. Journal of Dental Research 1990;69:1712–6. 33. Mjo¨r IA. The location of clinically diagnosed secondary caries. Quintessence International 1998;29:313–7. 34. Mjo¨r IA, Toffenetti F. Secondary caries: a literature review with case reports. Quintessence International 2000;31:165–79. 35. Wilson RF, Ashley FR. Identification of caries risk in schoolchildren: salivary buffering capacity and bacterial
36. 37.
38.
39.
40.
41.
42.
43.
44.
463
counts, sugar intake and caries experience as predictors of 2-year and 3-year caries increment. British Dental Journal 1989;167:99–102. Hausen H. Caries prediction – state of the art. Community Dentistry and Oral Epidemiology 1997;25:87–96. Fleiss JL, Kingman A. Statistical management of data in clinical research. Critical Reviews in Oral Biology and Medicine 1990;1:55–66. Tenovuo J. Salivary parameters of relevance for assessing caries activity in individuals and populations. Community Dentistry and Oral Epidemiology 1997;25:82–6. Eliasson L, Carle´n A, Almsta˚hl A, Wilkstro¨m M, Lingstro¨m P. Dental plaque pH and micro-organisms during hyposalivation. Journal of Dental Research 2006;85: 334–8. Rugg-Gunn AJ, Hackett AF, Appleton DR, Jenkins GN, Eastoe JE. Relationship between dietary habits and caries increment assessed over two years in 405 English school children. Archives of Oral Biology 1984;29:983–92. Firestone AR, Mu¨hlemann HR. In vivo pH of plaque-covered and plaque-free interdental surfaces in humans following a sucrose rinse. Clinical Preventive Dentistry 1985;7:24–6. Burt BA, Eklund SA, Morgan KJ, Larkin FE, Guire KE, Brown LO, et al. The effects of sugars intake and frequency of ingestion on dental caries increment in a three-year longitudinal study. Journal of Dental Research 1988;67:1422–9. Bjertness E. The importance of oral hygiene on variation in dental caries in adults. Acta Odontologica Scandinavica 1991;49:97–102. Krasse B. The Vipeholm dental caries study: recollections and reflections 50 Years Late. Journal of Dental Research 2001;80:1785–8.