The present study was performed with the approval of the Medical Ethics Committee of The First Affiliated Hospital of Henan University (2019LCSY-002). Signed informed consent was obtained from the caregivers and children prior to their participation in the study.

A cross-sectional study was conducted from April to June 2021 with a representative sample of schoolchildren aged 8 and 10 years in Lankao County, which is located on the eastern boundary of Henan Province. This county has endemic fluorosis, with fluoride concentrations ranging from 1.22 to 3.90 mg/L (Wang et al., 2008; Wang and Cao, 2013), which exceed the standard for drinking water quality in China (1.0 mg/L; GB5749-2006).

The formula for calculating a minimum number of randomly selected children was as follows: sample size (n) = [Z²×P(1-P)]/d², where Z is the statistical level of confidence for a 95% confidence interval (CI; Z = 1.96), P is the expected prevalence, and d is the precision (Naing et al., 2006). Recent studies have revealed that the global average prevalence of MIH is 12.9% (11.7%–14.3%) (Schwendicke et al., 2018). According to the formula, this study required 169 participants. Schools were selected randomly according to the number of schools in each town, and a stratified sample of pupils was selected from each school according to the total number of pupils in the school. The inclusion criteria were as follows: residents of both sexes, aged 8–10 years, born and raised locally, with all four permanent first molars and incisors fully erupted. The exclusion criteria were as follows: having no erupted permanent first molar and incisor, undergoing fixed orthodontic treatment with brackets or bands on permanent first molars, and defects less than 1 mm in diameter.

The European Academy of Paediatric Dentistry (EAPD) (Weerheijm et al., 2003) criteria for MIH were used in this study. Calibration exercises were conducted among three MIH investigators using clinical photographs of 26 patients. The tooth defects of patients covered all the degrees of MIH and other enamel defects, such as dental fluorosis, hypoplasia, and amelogenesis. The validity of using clinical photographs to study enamel defects was previously confirmed by Sabieha and Rock (1998), Wong et al. (2006), and Yi et al. (2021). After 1 month of training, three examiners were able to correctly diagnose all cases independently. Cohen’s kappa coefficients for inter- and intra-rater reliability were 0.92 and 0.89 for dental fluorosis, 0.86 and 0.75 for dental caries, and 0.65 and 0.77 for MIH, respectively. Furthermore, during this month, three investigators examined 10 enamel defect patients who visited the Department of Stomatology, which guaranteed that the three investigators were familiar with the diagnosis and management of children with MIH.

Participants were advised to brush their teeth before the exam, and the teeth stayed slightly wet during the process of inspection. The items to be prepared included a simple dental chair with a dental light source (DYNAMIC, China), disposable oral treatment plates, disposable gloves, and cotton balls.

MIH: Clinical examinations were carried out by unified trained specialist dentists and comprised examination for developmental enamel defects and dental caries using the EAPD criteria (Weerheijm et al., 2003). To guarantee between-examiner reproducibility, the examinations were performed jointly by two dentists. A specially designed chart was used to record sex, the year of birth, the presence of MIH, the number of affected incisors and molars, and the maximum degree of severity. Severity was quantified according to clinical appearance (Ghanim et al., 2017) and was classed as 1) mild, including white and yellow demarcated opacities; or 2) severe, including posteruptive enamel breakdown (PEB), atypical restoration, atypical carious lesions, and missing due to MIH.

Dental fluorosis: The Thysltrup–Fejerskov (TF) criteria were used to determine the occurrence of dental fluorosis with an ordinal scale from 0 to 9 (Thylstrup and Fejerskov, 1978). Based on the loss of structure, teeth with a TF of 0 were classified as normal, those with a TF of 1-4 were classified as mild, and those with a TF greater than 5 were classified as severe.

Dental caries: Clinical dental caries (manifest caries) were recorded as decayed, missing, or filled teeth (DMF). Decay was defined as visible tooth substance loss without the characteristics of developmental defects, pits, or fissures.

Oral hygiene status: Oral hygiene was recorded using the simplified oral hygiene index (OHI-S) described by Greene and Vermillion (1964). The level of oral hygiene was evaluated according to the debris index and was classed as fair (0) or poor (1).

To develop the machine learning models, we used a derivation cohort of children who met the inclusion criteria. The raw dataset contained the study subjects’ demographics and the results of the comprehensive oral examination. Initially, the dataset used for preprocessing and classification was collected. The main characteristics of this dataset included MIH, MIH severity, DF, DF severity, dental decay, oral hygiene, water quality, and sex.

The derivation cohort was randomly split into two datasets: a training cohort (70%) used to train the machine learning model and tune the parameters, and an internal validation cohort (30%) used to test the developed model on unseen data and to fine-tune the hyperparameters. For training, the original data space was balanced by oversampling using SMOTE. It worked by adding small samples from the data space to diminish the biased behavior of imbalanced data, thus changing the size of the training data space. In this study, supervised machine learning (e.g., logistic regression) and correlation analysis (e.g., Spearman correlation analysis) were used for classification and prediction. Once the number of models was considered for this particular study, we used accuracy (ACC), specificity (SPEC), and the ROC curve and area under the curve (AUC) to validate the prediction performance for binary classes (Figure 1).

The completed examination records were analyzed by the SPSS Statistics 22.0 program (IBM SPSS, Chicago, IL, United States of America), Python (version 3.9.7), and R (version 4.1.3). The presence of MIH was considered a dependent variable. Dental caries and dental fluorosis were considered independent variables. Confidence intervals of 95% were calculated for prevalence. Chi-squared tests and Fisher’s exact tests were used for comparisons and correlations. Logistic regression was conducted to analyze factors that could affect MIH. Significance was set at a p-value of < 0.05.

A total of 1,586 children were invited to participate. Of these, 18 were not in accordance with the inclusion criteria. The clinical characteristics and demographics of the study population are shown in Table 1. The dataset consisted of 1,568 samples with two types: MIH (215, 13.7%) and non-MIH. This record included 755 male and 813 female. Patients with MIH consisted of a greater proportion of males—;109 (14.4%) were male and 106 (13.0%) were female, —but the difference was not statistically significant. For permanent teeth, caries activity in Lankao City was low, at 5.6% in only permanent teeth and 14.2% in permanent and deciduous teeth. As expected, the percentage of dental fluorosis was as high as 54.8% (Table 1). There was a higher percentage of children with poor oral hygiene in this region (91.5%).

A supervised machine learning algorithm (e.g., logistic regression) was used to check for the occurrence of MIH. A total of 1,568 samples were analyzed, and six variables were included. During data preprocessing, the diagnosis of non-MIH or MIH was encoded using a binary encoder as 0 and 1, respectively. The ROC curve analysis was significant for this model, displaying an area under the curve of 0.72 (Figure 2). This model showed an accuracy of 70% and a specificity of 72%.

The typical clinical phenotype of MIH is shown in Figure 3, including mild and severe types. With respect to the distribution of MIH severity among affected index teeth, demarcated opacities comprised the predominant type of defects (Table 2).

The examined population presented a mixed dentition, and data from primary and permanent dentitions were presented separately (18,714 deciduous and 18,918 permanent teeth) (Table 3).

Based on a supervised machine learning algorithm, we constructed nomograms for predicting the occurrence of MIH, as shown in Figure 4. Non-DF had a great influence on the early occurrence of MIH. Significantly, the more severe the DF, the lower the occurrence.

Dental decay of permanent teeth was related to the occurrence of MIH. However, exposure to poor-quality shallow underground water did not increase the likelihood of developing MIH. Gender and oral hygiene had no influence on the occurrence.

A correlation heatmap of all parameters was generated using Spearman correlation coefficients. Figure 5 illustrates the correlation heatmaps of the core data features. MIH and MIH severity were negatively correlated with DF and DF severity. As in the logistic regression analysis, DF had a protective correlation with MIH, and the protective effect became stronger as DF severity increased. However, dental caries were positively correlated with MIH (OR = 1.843; 95% CI: 1.260–2.694) (Table 4), which was inconsistent with the heatmap. The heatmap showed a negligible correlation between MIH and dental caries.

Both the correlation heatmap and logistic regression analysis showed that the quality of underground water was not robustly correlated with MIH or MIH severity. Not surprisingly, the quality of underground water was correlated with DF and DF severity. A similar correlation was found between dental decay and oral hygiene.

To find the relationship between MIH and DF, we used a pair plot. In Python, the pair plot showed that MIH (Figure 6A) was strongly negatively correlated with DF and DF severity. The severity of MIH was reduced, followed by the incidence of DF and the strength of DF severity (Figure 6B).