Prediction of Caries Risk
Principles of risk prediction
Some basic principles have to be followed for successful and cost-effective caries
prediction, caries prevention, and caries control:
1. The higher the risk of developing caries for most of the population, the more
significant the effects of one single preventive measure and the stronger the
correlations between one single etiologic or modifying risk factor and the risk for
caries development.
2. In populations in which only a minority of the people will develop new carious
lesions, it is necessary to use accurate risk predictive measures to select at-risk
individuals and introduce needs-related combinations of caries-preventive measures,
in other words, a “high-risk strategy.”
In most child populations today, caries incidence is skewed. The majority of the
children in most age groups develop no or very few new carious surfaces, while a
small minority, 5% to 10%, develop several new carious surfaces each year.
Therefore, accurate caries risk predictors are useful. However, the dilemma is that no
method will guarantee that 100% of the selected high-risk individuals are “true” highrisk
individuals. The percentage of “true” high-risk individuals among the selected
group of high-risk individuals is termed sensitivity of the risk predictive method.
Similarly, methods are used to select nonrisk or low-risk individuals. The percentage
of “true” nonrisk individuals is termed specificity of the predictive method used.
These principles may be exemplified in Fig 41 (a to c) showing the outline of a
typical study with the aim of evaluating the predictive power of a risk marker of
dental caries. In the beginning, the baseline caries status and the level of the selected
risk marker are assessed. Caries recordings at the end of the follow-up period make it
possible to assess the true caries incidence during the period.
Prediction studies deal with two dichotomies: (1) individuals for whom it is believed
that the risk is high or low, and (2) the ones for whom true high or low caries
incidence is observed. Thus group a in Fig 41 (a) consists of correctly classified
individuals, true-positives, for whom it was believed that the risk was high and whose
actual caries incidence was high. Correspondingly, group d represents correctly
classified true-negatives. For individuals falling into groups b and c, misclassification
has occurred. For the false positives, in group “b,” a high risk was assumed, but the
true caries incidence was low. Correspondingly the false-negatives in group c were
believed to have a low risk, but their actual caries incidence was high.
This design is only usable for one predictor at a time. In practice, several predictors
are often regarded in prediction studies. In the case of multiple predictors, each of
them can be considered separately, which leads to predictor-specific numbers of trueor
false-positives and true- or false-negatives. Alternatively, the information of many
predictors can be condensed into a single variable on the level of which the prediction
of high or low risk is based. The techniques for such condensing range from simple
summaries to sophisticated regression-based multivariable models.
Even in the case of one predictor, the risk markers are seldom natural dichotomies. To
generate the four groups of interest¾true- and false-positives, and true- and falsenegatives
¾it is necessary to artificially dichotomize both multi- ple-level predictors
and the outcome, true caries incidence, which, in the data collection, is usually
regarded as the number of new carious tooth surfaces, not as a dichotomy.
The dichotomization can be done in different ways. Each threshold level for believed
high risk and for observed high true caries incidence leads to a different distribution
of the study subjects into the four groups of interest (true- and false-positives, and
true- and false-negatives). Thus, when the results of a prediction study are evaluated,
it is of utmost importance to consider the threshold levels that have been used. To
estimate the accuracy of the classification of the four groups in Fig 41 (a), the
quantities a, b, c, and d are organized in the form of a 2 x 2 contingency table (Fig 41
(b)).
Six different measures of accuracy and their estimators are given in Fig 41 (c). As
mentioned earlier, sensitivity is the proportion of those who were believed to have a
high risk among the individuals whose actual caries incidence during the follow-up
was high. Specificity is the proportion of those who were believed to have a low risk
among the patients whose actual caries incidence during the follow-up was low.
False-positive rate and false-negative rate carry exactly the same information as
sensitivity and specificity but, in contrast, reveal proportions of misclassified subjects.
False-positive rate is the proportion of those who were believed to have a high risk
among the subjects whose actual caries incidence during the follow-up was low.
False-negative rate is the proportion of subjects who were believed to have a low risk
among those individuals whose actual caries incidence was high.
Positive predictive value is the proportion of those whose actual caries incidence was
high among the subjects who were believed to have a high risk. Negative predictive
value is the proportion of subjects whose actual caries incidence was low among the
patients for whom a low risk was predicted.
All six of these measures should always be examined pairwise. For instance,
sensitivity has no meaning if the specificity is not known. This is important from a
cost-effectiveness point of view during screening of populations of children with
relatively low caries prevalence and incidence before needs-related caries preventive
programs are designed. However, for the individual patient the consequences of being
a false nonrisk or low-risk individual are very different from those of being selected
as a false-positive high-risk individual.
