INTRODUCTION

Anatomical variability can exist between different populations around the world. If different ethnic types are all treated the same, a serious error could be made.¹ The issue of craniofacial growth in differing ethnicities may be of interest in diverse countries, particularly where heterogeneity of ethnicities exists, such as in the US.

Evaluation and prediction of craniofacial growth are of great importance in orthodontics. However, this topic is only considered in a few publications.^2–4 Since growth is influenced by genetic and circumstantial factors, various growth evaluation materials and methods have been tailored to specific populations.^2–7 While the importance of considering population differences in predicting growth has been emphasized traditionally, the consequences of overlooking these differences in a growing child have never been provided or visualized.

A major obstacle when conducting growth studies may be the need to collect longitudinal serial growth data meticulously. Due to funding and ethical reasons, it has become challenging to take repeated photographic and radiographic records of children regularly. Recently, Moon et al.^2,3 reported individualized craniofacial growth prediction models using longitudinal serial cephalograms of 410 Korean children. In addition, Roseth⁴ developed growth prediction models using longitudinal serial growth datasets of American children of northern European origin collected by the Mathews Growth Study. These two datasets provide an opportunity for comparison of how different ethnicities affect growth prediction results.

Development of a growth prediction model requires a sophisticated method that can handle multiple input and output variables that are significantly correlated. The partial least squares (PLS) method is a machine learning method with the advantages of multivariate multiple linear regression and principal component analysis through dimensional reduction and modeling with latent variables.^8,9 Since data characteristics commonly used in orthodontic research typically include relatively large numbers of skeletal, dental, and soft tissue variables that are correlated to each other, PLS has mostly been applied for predicting soft tissue changes resulting from combined surgical-orthodontic treatment,^10–14 orthodontic treatment outcomes,¹⁵ and in predicting craniofacial growth,^2,3 where multiple input and output variables are prevalent. One of the significant advantages of the PLS algorithm is its ability to account for nearly all skeletal and soft tissue cephalometric variables for each individual subject. This capability is achieved through linear combination and iterative computations involving the topology of 78 cephalometric landmarks during the PLS algorithm procedures. In other words, even if specific vertical and anteroposterior cephalometric variables are not explicitly included among the input variables, the algorithm still effectively considers factors such as whether an individual subject has a hypodivergent or hyperdivergent facial growth pattern or Class II or III relationship, among others.^8,9

The purpose of this study was to compare craniofacial growth prediction results for Korean and American children according to growth prediction models developed using Korean and American longitudinal growth data, by applying the PLS method.

MATERIALS AND METHODS

The Institutional Review Board for the Protection of Human Subjects of the Seoul National University School of Dentistry reviewed and approved the data collection protocol (S-D20240019). The University of the Pacific (UOP) Human Subjects Protection Office of Research and Sponsored Programs approved the research protocol, and the project received an exempt review (UOP IRB 2023-28).

Longitudinal Serial Growth Data Collection From Two Different Population Data Sources

The longitudinal serial growth datasets of Korean children were presented by Moon et al.² The study included 410 children, comprising 236 girls and 174 boys, with an average age of 11.5 years. The children were selected through the collection and collation of dental records from 25,810 patients who sought orthodontic treatment between January 2002 and December 2022 at the Department of Orthodontics, Seoul National University Dental Hospital. These children did not undergo any orthodontic or orthopedic treatment but had serial cephalograms taken. Among the 410 subjects, 90 (22%) subjects had cephalograms taken more than three times. Consequently, 935 cephalograms were taken for 679 serial growth datasets. On each of the 935 images, 78 cephalometric landmarks, consisting of 46 hard tissue and 32 soft tissue landmarks, were manually identified by a single examiner (SJL).

The longitudinal serial growth datasets of American children were collected from the American Association of Orthodontist Foundation Craniofacial Growth Legacy Collection homepage. Among the various datasets, lateral cephalograms collected by the UOP Mathews Growth Study were used. The American children were of northern European origin. On the 296 lateral cephalogram images from 33 subjects, comprising 21 girls and 12 boys aged 7.4–19.8 years, 78 cephalometric landmarks were manually identified by the same examiner (SJL). These UOP Mathews Growth Study data included 1257 serial growth datasets, which is twice as much data as the 679 Korean serial growth datasets. Table 1 summarizes the characteristics of the Korean and American longitudinal serial growth data.

Table 1. Characteristics of the Korean and American Subjects Whose Longitudinal Serial Craniofacial Growth Data Were Used in the Present Study

View Fullscreen

RESULTS

The growth prediction errors for Korean and American subjects with prediction models developed using data from the same and different populations are summarized in Table 2. In general, the growth prediction results of the prediction model developed using data from the same population were more accurate than those of the prediction model developed using data from a different population (P < .0001).

Table 2. Growth Prediction Errors of Korean and American Subjects With Prediction Models Developed Using Data From the Same and Different Populations^a

View Fullscreen

Figure 1 illustrates several representative scatter plots of prediction errors for the posterior nasal spine, nasal bone tip, nose tip (pronasale), and upper lip. In general, the 95% confidence ellipses of the growth prediction errors for Korean children were larger than those of the growth prediction errors for American children, indicating that the prediction results were more accurate in American children than in Korean children.

View Figure

• Download as Powerpoint
• Download Figure

In general, the prediction errors were greater when growth of Korean children was predicted by the prediction model developed using American data than when growth of American children was predicted by the prediction model developed using Korean data. When Korean children’s growth was predicted by the prediction model developed using American data, the bias indicated by the black circle was positioned downward and anteriorly, which implied that craniofacial growth was overestimated. On the contrary, when American children’s growth was predicted by the prediction model developed using Korean data, the bias marked by the red diamond was located in the three-quarter plane, which implied that craniofacial growth was underestimated (Figure 1).

Real-case examples comparing real growth with the prediction results generated by prediction models developed using data from the same and different populations are shown in Figure 2. The prediction results generated by prediction models developed using data from the same population (blue dotted lines) were more accurate and closer to the actual growth than the prediction results generated by prediction models developed using data from a different population (red dotted lines). The red dotted lines show that the growth of American children was underestimated, and the growth of Korean children was overestimated.

View Figure

• Download as Powerpoint
• Download Figure

DISCUSSION

In the present study, we compare growth prediction results generated by prediction models developed using data from the same and different populations. Our specific aim was to highlight the importance of considering population differences when predicting craniofacial growth. The growth prediction results for Korean children were more accurate when the prediction model developed using Korean data was applied than when the prediction model developed using American data was applied. Likewise, the growth prediction results for American children were more accurate when the prediction model developed using American data was applied than when the prediction model developed using Korean data was applied. These results were unsurprising. Nonetheless, we are the first to confirm and visualize the discrepancies in craniofacial growth prediction results when ethnic differences are ignored or neglected.

The prediction model developed using American data yielded more accurate results than that developed using Korean data. The prediction model developed using Korean data underestimated the growth of American children, specifically in the upper-posterior direction. However, an ethnic difference should be inferred with caution because unknown or confounding factors may exist. For example, the results might have been influenced by the data characteristics. PLS is a machine-learning method, and its accuracy is dependent on the number of training or learning sample sizes.^27,28 The American data were collected from 33 children, while the Korean data were collected from 410 children (Table 1). In contrast, 1257 American datasets and 679 Korean datasets were collected. In other words, the American data were more homogeneous and cohesive, with a greater sample size than the Korean data, which included many more subjects and had a smaller sample size, thereby increasing heterogeneity. Recently, Lee et al.²⁷ investigated factors influencing the performance of artificial intelligence and found that growth prediction errors increased with an increasing number of subjects.

In addition, 51.5% and 48.5% of American subjects had Class I and II malocclusion, respectively, and none had Class III malocclusion. However, 43.2% and 34.1% of Korean subjects had Class II and III malocclusion, respectively. Authors of epidemiological studies have consistently shown that Koreans have significantly higher rates of Class III malocclusion than other ethnicities.^29,30 Authors of previous studies have reported that children with Class III malocclusion showed greater growth prediction errors than children with Class II malocclusions.^2,3

The two population datasets exhibited significant disparities in sample sizes, particularly regarding the number of subjects, longitudinal serial datasets, malocclusion proportions, and age range of the subjects. For example, the ages (minimum–maximum years) were 3.4–31.3 and 7.4–19.8 for Koreans and Americans, respectively. The difference in prediction accuracy might have partly originated from such subject characteristics, ie, a possible influence of malocclusion proportions and the subject age range on growth prediction accuracy.

The two-dimensional scatter plots for selected landmarks in Figure 1 show that the growth prediction errors increased when the landmark position was further away from the cranial base, and soft tissue landmarks demonstrated greater prediction errors than hard tissue landmarks. In addition, girls and subjects with Class II malocclusion had more accurate growth prediction results than boys and subjects with Class I malocclusion. All these results were in line with the growth prediction results of Moon et al.^2,3 Determining how much latent growth exists for patients with Class II and III malocclusions employing different cephalometric values for diverse ethnic groups is critical for determining an appropriate treatment plan. A decade ago, Haskell and Segal¹ strongly recommended the application of computer programs that can visualize treatment objectives. The drawn illustrations of growth prediction results, as shown in Figure 2, might help clinicians offer treatment choices for a growing child.

The results of the present study do not imply that craniofacial growth prediction results could be perfect or satisfactory. Nonetheless, some methods might be better than others. It is advisable to apply a growth prediction model developed using data from the same population that represents a specific ethnicity.

CONCLUSIONS

• This study confirmed and highlighted the importance of considering population differences when predicting craniofacial growth. Although growth cannot be perfectly predicted, some methods may be better than others. Therefore, it may be desirable to apply a growth prediction model generated using data from the same population.