International Journal of Legal Medicine

, Volume 119, Issue 1, pp 22–26

Validation of common classification systems for assessing the mineralization of third molars

Authors

  • Andreas Olze
    • Institute of Legal Medicine (Charité)Humboldt University of Berlin
  • Dominique Bilang
    • Institute of Legal Medicine (Charité)Humboldt University of Berlin
  • Sven Schmidt
    • Institute of Legal Medicine (Charité)Humboldt University of Berlin
  • Klaus-Dieter Wernecke
    • Institute of Medical Biometry (Charité)Humboldt University of Berlin DRK-Kliniken Westend
  • Gunther Geserick
    • Institute of Legal Medicine (Charité)Humboldt University of Berlin
    • Institute of Legal Medicine (Charité)Humboldt University of Berlin
Original Article

DOI: 10.1007/s00414-004-0489-5

Cite this article as:
Olze, A., Bilang, D., Schmidt, S. et al. Int J Legal Med (2005) 119: 22. doi:10.1007/s00414-004-0489-5

Abstract

One major criterion for dental age estimation is the evaluation of third molar mineralization. There are various methods for evaluating tooth mineralization based on classification by stages. The aim of the present work is to assess the validity of the common classification systems. To this end, we analyzed 420 conventional orthopantomograms of German females aged 12–25 years old. The mineralization status of tooth 38 was determined using the stages defined by Gleiser and Hunt, Demirjian et al., Gustafson and Koch, Harris and Nortje and Kullman et al., respectively. Of the methods tested, the most accurate results were obtained with Demirjian et al.’s classification system, which performed best not only for observer agreement but also for the correlation between estimated and true age. It is argued that this is due to the fact that Demirjian et al.’s classification is based on a sufficient number of stages which are defined independently of speculative estimations of length. This leads to the conclusion that the method devised by Demirjian et al. should be used for evaluating the mineralization of third molars for purposes of forensic age determination.

Keywords

Forensic age estimationDental ageMineralizationThird molars

Introduction

Today forensic age diagnostics is an established research sector of legal medicine in its own right (Ohtani et al. 2003; Olze et al. 2004; Ritz-Timme et al. 2003; Schmeling et al. 2004; Takasaki et al. 2003). In recent years it has become increasingly important to determine, in particular, the age of living persons (Schmeling et al. 2001b). From a legal perspective, such age estimates are carried out to determine whether a suspect without valid identification documents has reached the age of criminal responsibility and whether general criminal law in force for adults is to be applied. In many countries the age thresholds of relevance to criminal prosecution lie between 14 and 18 years (Dünkel et al. 1997).

In line with recommendations drawn up by the international interdisciplinary Study Group on Forensic Age Diagnostics (http://www.charite.de/rechtsmedizin/agfad/index.htm), a forensic age diagnosis for the purpose of criminal investigations should consist of a clinical examination, including the recording of body measurements and an evaluation of signs of sexual maturity, an X-ray examination of the left hand, and a dental examination which records dentition status and evaluates an orthopantomogram (Schmeling et al. 2001a). One major criterion for dental age estimation is the evaluation of third molar mineralization.

Various classifications have been devised for evaluating tooth mineralization (Gleiser and Hunt 1955; Nolla 1960; Haavikko 1970; Liliequist and Lundberg 1971; Demirjian et al. 1973; Gustafson and Koch 1974; Nortje 1983; Harris and Nortje 1984; Kullman et al. 1992; Köhler et al. 1994). Since the validity of an age estimate crucially depends on the classification method used, the most appropriate one should be selected. The present work examines the validity of five different stage-based systems.

Materials and methods

We analyzed 420 conventional orthopantomograms of German females aged 12–25 years old and each age group was represented by X-rays from at least 30 females. This study could be limited to female probands, as there are no statistically significant gender-related differences to the chronology of wisdom tooth mineralization (Olze et al. 2003). The mineralization status of the left mandibular third molar (tooth 38) was determined for all persons included in the study using five different methods, namely the classifications by Gleiser and Hunt (1955), Demirjian et al. (1973), Gustafson and Koch (1974), Harris and Nortje (1984) and Kullman et al. (1992).

The methods of Gleiser and Hunt (1955), Demirjian et al. (1973), Harris and Nortje (1984) and Kullman et al. (1992) are described in both text and in diagrammatic form (Figs. 1, 2, 3, and 4).
Fig. 1

Diagrammatic representation of the stage classification by Gleiser and Hunt (1955)

Fig. 2

Diagrammatic representation of the stage classification by Demirjian et al. (1973)

Fig. 3

Diagrammatic representation of the stage classification by Harris and Nortje (1984)

Fig. 4

Diagrammatic representation of the stage classification by Kullman et al. (1992)

Gustafson and Koch (1974) only provided a verbal description and defined the following 4-stage classification:
  • Stage 1: commencement of mineralization

  • Stage 2: completion of crown

  • Stage 3: eruption when the cusp(s) penetrate the gingiva

  • Stage 4: completion of root(s).

For reasons of space we have omitted the verbal descriptions of the remaining classifications and kindly ask readers to refer to the original works.

Each of the five methods included in the study was evaluated by two independent observers (A, and B), with observer A examining all X-ray images twice (A1 and A2) where 6 months elapsed between evaluations A1 and A2.

A 2-factor analysis of variance was used to test the differences between these individual evaluations and methods; in order to account for the fact that observer values Xm,i,j(k) were measured on a category scale, it was carried out non-parametrically (Brunner et al. 2002). Significance was assessed at p<0.05.

Both inter-observer and intra-observer agreement were determined using the weighted kappa coefficient (Fleiss 1981), taking two (scaled) measurements at a time (intra-class correlation) and the 95% confidence intervals were calculated for each kappa value. The method that yields the best inter-observer and intra-observer agreement, i.e. where assessment variability is low, is assumed to be the most appropriate method.

Correlation between true age and estimated age based on the results obtained from the various methods was evaluated by directly comparing the stage determined on a category scale with the age determined on an interval scale using the eta coefficient. Eta is a coefficient of categorical-by-interval association (association refers to measures of strength of relationship in which at least one of the variables is dichotomous, nominal, or ordinal). For linear relationships, eta equals the Pearson’s correlation coefficient (Siegel 1956). The method providing maximum correlation was considered to be the most appropriate.

Results

Table 1 shows the p-values of the analysis of variance for a comparison between all three observations across all five methods (test 1), a comparison of observer A1 versus observer B for all methods (test 2), a comparison of observer A2 versus observer B for all methods (test 3) and a comparison of observer A1 versus observer A2 for all five methods (test 4). No differences were observed between observers, but significant differences were observed between methods, and there was a significant interaction between methods and observers. In this context, significant interaction means that a difference between methods is not independent of the observers, i.e., it is particularly characteristic of certain observers.
Table 1

P-values of the analysis of variance for comparison between observers and methods

Analysis of variance

Test 1

Test 2

Test 3

Test 4

Systematic differences between observers

0.73958

0.63824

0.99388

0.62295

Systematic differences between methods

0.00000

0.00000

0.00000

0.00000

Interaction between observers and methods

0.00000

0.00012

0.01904

0.00000

For a quantification of these systematic influences, an intra-class coefficient (ICC) was determined together with 95% confidence intervals (Table 2).
Table 2

Intra-class coefficients (ICC) with 95% confidence intervals (CI)

Method

Observer

ICC

95% CI

Demirjian et al. 1973 (interind.)

A1-B

0.953

0.9440–0.9612

Demirjian et al. 1973 (interind.)

A2-B

0.989

0.9846–0.9926

Demirjian et al. 1973 (intraind.)

A1-A2

0.957

0.9488–0.9659

Gustafson and Koch 1974 (interind.)

A1-B

0.873

0.8380–0.9088

Gustafson and Koch 1974 (interind.)

A2-B

0.979

0.9662–0.9908

Gustafson and Koch 1974 (intraind.)

A1-A2

0.885

0.8505–0.9191

Gleiser and Hunt 1955 (interind.)

A1-B

0.954

0.9314–0.9772

Gleiser and Hunt 1955 (interind.)

A2-B

0.979

0.9569–0.9991

Gleiser and Hunt 1955 (intraind.)

A1-A2

0.975

0.9696–0.9810

Kullman et al. 1992 (interind.)

A1-B

0.923

0.9098–0.9426

Kullman et al. 1992 (interind.)

A2-B

0.979

0.9725–0.9862

Kullman et al. 1992 (intraind.)

A1-A2

0.941

0.8994–0.9593

Harris and Nortje 1984 (interind.)

A1-B

0.833

0.7831–0.8827

Harris and Nortje 1984 (interind.)

A2-B

0.931

0.9111–0.9505

Harris and Nortje 1984 (intraind.)

A1-A2

0.902

0.8727–0.9307

The highest intra-class coefficient was displayed by Demirjian et al.’s method for evaluation A2-B. An intra-class coefficient of >0.95 was obtained for the Gleiser and Hunt method and for Kullman et al. (A2-B). The methods developed by Gustafson and Koch (A1-B, A1-A2) and by Harris and Nortje (A1-B) yielded intra-class coefficients of <0.90.

In order to assess correlation between age estimated on the basis of the stages defined in the various methods and true age, eta coefficients were determined for all methods and all observers (Table 3).
Table 3

Eta coefficients for all methods and all observers

Method

Observer A1

Observer A2

Observer B

Demirjian et al. 1973

0.883

0.865

0.856

Gustafson and Koch 1974

0.745

0.770

0.773

Gleiser and Hunt 1955

0.879

0.871

0.855

Kullman et al. 1992

0.880

0.859

0.846

Harris and Nortje 1984

0.863

0.846

0.811

The method of Demirjian et al. shows the highest eta coefficient (eta coefficient for A1=0.883), followed by the methods of Kullman et al. (eta coefficient for A1=0.880) and of Gleiser and Hunt (eta coefficient for A1=0.879). An eta coefficient of <0.8 was calculated for the method developed by Gustafson and Koch.

Hence, the method developed by Demirjian et al. can be considered the best of the five methods reviewed here, producing the highest values for both observer agreement and correlation between true age and age estimated on the basis of the stages defined. Good results were also achieved with the methods developed by Gleiser and Hunt and by Kullman et al. The methods developed by Gustafson and Koch and by Harris and Nortje performed noticeably less well.

Discussion

Various classifications are available for evaluating tooth mineralization (Gleiser and Hunt 1955; Nolla 1960; Haavikko 1970; Liliequist and Lundberg 1971; Demirjian et al. 1973; Gustafson and Koch 1974; Nortje 1983; Harris and Nortje 1984; Kullman et al. 1992; Köhler et al. 1994). They differ with regard to the number of stages, the definition of each stage and the presentation.

Gustafson and Koch, like Harris and Nortje, defined 4–5 stages, Kullman et al. offered 7 stages, and Demirjian et al. as well as Nortje established 8 stages. The remaining classifications use 10–16 stages. Apart from Gustafson and Koch, where the method was only described verbally, all the authors described their stages in both text and diagram form. Apart from Demirjian et al., all the authors used stages based on fractions of the future length of crown or root. Demirjian et al.’s stages are defined by changes in form and do not depend on speculative estimates of length.

As some of these methods are very similar, our review was confined to five basic types of classification. We selected the methods by Gleiser and Hunt (1955), Demirjian et al. (1973), Gustafson and Koch (1974), Harris and Nortje (1984) and Kullman et al. (1992).

A number of authors have already compared the validity of different stage-based methods. Hägg and Matsson (1985) examined 300 Swedish children aged 3.5–12.5 years old using the multi-stage classifications developed by Liliequist and Lundberg (1971), Demirjian et al. (1973) and Gustafson and Koch (1974).

Accuracy tests were performed to establish mean deviations and the increase in linear correlation between estimated and real age. A high rate of agreement was obtained from Demirjian et al.’s method with children aged 3.5–6.5 years old, but for older children Demirjian et al.’s method displayed less accuracy. When applying the method by Liliequist and Lundberg (1971), accuracy was low across all age groups with a general tendency to underestimate. For the method described by Gustafson and Koch (1974), accuracy was only high for male subjects.

Gustafson and Koch’s method (1974) displayed greater inter-observer and intra-observer error than the methods devised by Demirjian et al. (1973) and Liliequist and Lundberg (1971).

Pöyry et al. (1986) investigated the mineralization of teeth 31–38 using 96 orthopantomograms from 48 Finnish boys aged 3–14 years old. They compared Haavikko’s (1970) modified version of the Gleiser and Hunt classification (1955) (method A) with the classification defined by Demirjian et al. (1973) (method B). The estimates were performed by two observers assessing each X-ray twice. For method A the mean intra-observer error was 20.1% and for method B, 10.2%. However, the inter-observer error was 31.0% for method A and 19.2% for method B. Stages which proved difficult to assess were Cco (coalescence of cups), Cr1/2 (crown 1/2 complete) and Crc (crown complete) in method A and stage G in method B.

Staaf at al. (1991) scrutinized the validity of age estimation methods based on the classifications set out by Demirjian et al. (1973), Haavikko (1970) and Liliequist and Lundberg (1971). They examined the orthopantomograms of 541 Swedish children of both genders aged 5.5–15.5 years old and 37 randomly selected cases were evaluated a second time. When using the method developed by Demirjian et al. (1973), the age of both boys and girls was overestimated by 6–10 months. The methods drawn up by Haavikko (1970) and by Liliequist and Lundberg (1971) resulted in a systematic underestimation or overestimation by 6–7 and 7 months, respectively. Intra-observer error was low for all three methods and not significant. In discussing why the method by Demirjian et al. (1973) should prove less valid, the authors proposed possible ethnic differences between the original reference population and the sample with which they were working.

A study by Mörnstad et al. (1995) considered 197 orthopantomograms from Swedish children aged 5, 6, 9 and 12 years old. A total of 13 independent observers assessed the mineralization status of the 7 mandibular teeth excluding the third molar using the classifications by Demirjian et al. (1973), Gustafson and Koch (1974), Liliequist and Lundberg (1971) and Haavikko (1970). As Hägg and Matsson (1985) and Staaf et al. (1991) had already reported overestimated age in Scandinavian subjects following the use of Demirjian et al.’s stages, the authors incorporated reference values from a study by Kataja et al. (1989), where the criteria defined by Demirjian et al. (1973) had been applied to a Finnish population. The highest accuracy compared with real age was obtained from the method by Demirjian et al. (1973), although applying the reference values gained from a Canadian sample to a Swedish population resulted in an overestimation across both genders and all age groups of 0.4–1.8 years. The overestimation was much smaller (0.1–0.8 years) when using the Scandinavian reference values reported by Kataja et al. (1989).

Reventlid et al. (1996) published intra-observer and interobserver error rates from the random sample investigated by Mörnstad et al. (1995). All 197 orthopantomograms were assessed by 13 independent observers with experience in estimating children’s ages: 28 orthopantomograms were analyzed again by 12 observers after an interval of about 2 years. Mean intra-observer error was 0.03–0.20 years. Statistically significant differences emerged in the 6–9 year age group when using the method by Haavikko (1970) and in the 12 year age group when using the classifications by Gustafson and Koch (1974) and Demirjian et al. (1973). No systematic overestimation or underestimation was displayed by one or more observers. For interobserver error there were significant differences between estimated age across all observers, age groups and classification systems. The scatter for mean estimated age was 0.8–1.2 years for all observers and methods. The range for individual estimated ages varied from 3.2 to 8 years, being particularly low using the method by Haavikko (1970) and particularly high using the method by Gustafson and Koch (1974).

Existing studies on the validity of different classification systems for assessing dental mineralization are of limited value in that the stages defined for each method were based either on the reference population used for the original work or on some other population dissimilar to the sample selected for the study. It is impossible, therefore, to dismiss the risk that differences in sample size, age group, age distribution across the group, ethnic origin or the health status of the subjects might have exerted a substantial influence on the comparative studies published to date. The work by Mörnstad et al. (1995) revealed, for example, that applying Demirjian et al.’s stages to different reference populations will result in different deviations between estimated and real age.

The procedure adopted by the present study resolves this methodological problem, to the authors’ knowledge for the first time, by using weighted kappa coefficients (Fleiss 1981) and eta coefficients (Siegel 1956) to analyze the various stages independently of reference populations.

Of the various methods examined, Demirjian et al.’s classification achieved the highest values for both observer agreement and for correlation between the stages as defined by the method and true age. It can, therefore, be regarded as the best method. Good results were also obtained by the classifications according to Gleiser and Hunt and Kullman et al. Noticeably poorer results were derived from the methods by Gustafson and Koch and Harris and Nortje. The small number of stages evidently exerts a negative impact here, because a large age interval between stages means that guessing a wrong stage will result in a correspondingly greater error in estimated age. One particular advantage of the classification by Demirjian et al. is that stages are not defined on the basis of speculative estimates of length.

The authors conclude that Demirjian’s stages should be used to evaluate third molar mineralization for forensic age estimates.

Acknowledgments

This study was supported by a grant from the Deutsche Forschungsgemeinschaft (GE 968/3–1).

Copyright information

© Springer-Verlag 2004