Development and validation of an age estimation model based on dental characteristics using panoramic radiographs

Abstract

Dental characteristics have considerable potential as indicators for estimating chronological age. This study developed a regression model for age estimation using dental characteristics observed in panoramic radiographs. A total of 2,391 radiographs from individuals aged 20 to 89 years were analyzed, with a focus on five treatment-induced characteristics. Analyses revealed statistically significant correlations between all observed characteristics and chronological age, supporting the potential of dental characteristics as novel age indicators. A model incorporating only posterior teeth from both jaws achieved an adjusted R-squared value of 0.564 and a root mean square error (RMSE) of 13.144 years, closely comparable to the full-dentition model, which had values of 0.558 and 13.235 years, respectively, and is regarded as the most recommendable model for practical application. On the same test set, the developed model had an RMSE that was 2.651 years higher than that of a non-destructive method widely used in Korean forensic practice, indicating slightly lower accuracy. Nevertheless, given its convenience in forensic practice, it should be considered a supplementary tool used alongside conventional methods rather than a direct replacement. Future research should leverage diverse datasets from multiple institutions and apply advanced technologies, such as machine learning, to enhance the applicability and robustness of dental characteristic-based models.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-19164-7.

Introduction

Dental age estimation is widely employed in forensic and anthropological applications due to its superior accuracy compared to other physiological age indicators^1–5. Since Gustafson’s (1950) study⁶ on histological factors associated with aging, numerous techniques to estimate dental age based on these factors have been developed. However, as most of these factors are observable within the internal structure of the tooth, their assessment typically requires tooth destruction, raising ethical and biological concerns for living individuals. Hence, non-destructive approaches have been recommended for age estimation in living persons⁷.

Among the histological factors used in non-destructive age estimation, age estimation based on attrition and secondary dentine—which include observing morphological changes in the tooth crown caused by attrition through visual inspection^8–11 or evaluating reductions in pulpal volume due to secondary dentine deposition on radiographs^12–15, respectively—are the most prominent. However, as single indicator approaches often yield lower accuracy than regression models that consider multiple correlated factors¹⁶, the International Organization for Forensic Odonto-Stomatology recommends combining age estimates derived independently from different single indicators¹⁷. Similarly, the European Asylum Support Office¹⁸ and the Study Group on Forensic Age Diagnostics¹⁹ advocate for combining age estimates from multiple anatomical regions, such as the wrist, clavicle, and third molar, to enhance the determination of legal adulthood or minority. Therefore, age estimation methods using novel indicators beyond attrition and secondary dentine can further improve accuracy when combined with conventional methods.

Permanent teeth, once altered by dental treatments or biological aging, do not regenerate, making them potential indicators of age. A few studies have investigated the correlation between dental characteristics—such as tooth loss and restorative or prosthetic treatments—and chronological age, reflecting both physiological and treatment-induced changes within the oral cavity. Azrak et al.²⁰ explored the correlation between clinical and radiographic dental findings and age in a Central European population, demonstrating that these findings provide valuable insights into dental age estimation. However, their study was limited by the narrow age range of its sample, from the 20 s to the 50 s, which constrained its applicability to broader age groups. Yamashita et al.²¹ developed an age estimation formula based on intraoral findings in Japanese adults, reporting its efficacy for middle-aged and older individuals. Nonetheless, the study’s small sample size and exclusion of participants under 40 years old highlighted the need for further research with larger sample sizes and the development of age estimation formulas applicable to wider age ranges. Cheong et al.²² devised an age estimation formula using nine dental characteristics observed in panoramic radiographs of Koreans, demonstrating the potential of the developed model as an adjunct tool in forensic practice. However, they emphasized the necessity for large-scale validation studies to confirm the applicability of the model. Most previous studies require additional oral findings, highlighting the need for research to assess the practical applicability of age estimation methods based on dental characteristics.

In this study, we aimed to develop a non-destructive regression model for age estimation in living adults based on dental characteristics observed in panoramic radiographs. To validate the practical applicability of the developed model, its performance was compared with Roh’s method¹⁵, a non-destructive method commonly used in Korean forensic practice, by applying both methods to a test set.

Results

Reliability of observer assessments

The kappa coefficients for dental codes, including X (missing tooth), T (root canal treatment), F (filling), P (prosthesis), and L (dental implant), were 0.964 for intra-observer and 0.950 for inter-observer reliability, indicating an almost perfect agreement (P < 0.001)²³.

Distribution of dental codes by sex, side, and jaw

The Mann–Whitney U test revealed significant differences in the distributions of dental codes F and L between radiographs of males and females (Supplementary Table S1). The Wilcoxon signed-rank test found no significant differences in the distribution of dental codes between the left and right sides (Supplementary Table S2). In contrast, significant differences in the distribution of dental codes X, T, F, and P were observed between the upper and lower jaws (Supplementary Table S3).

Frequency of dental codes by age group

The frequency of dental codes observed across all radiographs is presented as the number and percentage of each code by age group in Table 1 and Supplementary Fig. S1. As age increased, the number of teeth assigned codes X, T, P, and L tended to increase. The frequency of teeth assigned the code F increased from the 20 s to the 40 s and decreased from the 50 s onward. The frequency of teeth assigned codes X and P showed the most significant differences between the 20 s and 80 s. For teeth assigned the code X, the frequency was 2.8% in the 20 s, rising to 41.3% in the 80 s. Similarly, the frequency of teeth assigned the code P was 3.0% in the 20 s and 33.3% in the 80 s. When analyzed by sex, radiographs of males and females exhibited similar distributions of dental codes (Supplementary Table S4). The descriptions of the dental codes are presented in Table 2.

Table 1.

Frequency of dental codes in the training set for both sexes.

Age group (years)	Percentages (number of dental codes)
	V	X	T	F	P	L
20s	80.6 (6548)	2.8 (224)	2.7 (221)	13.5 (1094)	3.0 (241)	0.2 (18)
30s	72.0 (5850)	3.2 (262)	6.5 (525)	16.5 (1337)	8.1 (660)	0.8 (63)
40s	63.1 (5120)	3.3 (271)	8.4 (681)	18.8 (1528)	14.9 (1209)	2.4 (196)
50s	59.2 (4805)	5.4 (435)	8.9 (719)	14.9 (1213)	21.8 (1770)	5.2 (420)
60s	50.3 (4084)	11.4 (925)	8.9 (723)	10.0 (815)	30.9 (2505)	10.3 (840)
70s	38.0 (3082)	21.6 (1751)	9.9 (805)	7.5 (605)	35.5 (2883)	10.8 (879)
80s	21.7 (1764)	41.3 (3356)	9.8 (794)	5.4 (441)	33.3 (2706)	6.0 (483)
Total	55.0 (31,253)	12.7 (7224)	7.9 (4468)	12.4 (7033)	21.1 (11,974)	5.1 (2899)

The description of dental codes is presented in Table 2. Percentage represents the proportion of observations of each code relative to the total number of codes observed in each age group.

Table 2.

Dental characteristics and associated codes.

Code	Description
V (Virgin tooth)	No evidence of dental disease, treatment, or anatomical abnormality
X (Missing tooth)	Extracted or congenital missing tooth
I (Impacted tooth)	Unerupted or impacted tooth
D (Defect)	Defect by dental caries, tooth fracture, or fallen-out fillings
R (Residual root)	Remaining root after severe dental caries or trauma
T (Root canal treatment)	Root canal-filled tooth by endodontic treatment
F (Filling)	Filled tooth with any kind of restoration
P (Prosthesis)	Tooth with full-veneer crown
L (Dental implant)	Dental implant inserted into an edentulous area

The system for coding dental characteristics used in this study was a modification by Cheong et al.²² of the original system proposed by Lee et al.⁴⁹.

Correlation between dental codes and chronological age

The Spearman correlation coefficients between the number of dental codes and chronological age are presented in Table 3 and Supplementary Table S5. Across all age groups, codes X, T, P, and L demonstrated positive correlations with chronological age, whereas code F exhibited a negative correlation. Similar correlation coefficients were observed for radiographs of each sex. The codes exhibiting moderate positive correlations were P (0.620) and X (0.602). Code L displayed a weak positive correlation (0.362), whereas code T exhibited a very weak positive correlation (0.257). Code F demonstrated a weak negative correlation with a coefficient of − 0.371.

Table 3.

Spearman correlation coefficients between the number of dental codes and chronological age in the training set for both sexes.

Age group (years)	V	X	T	F	P	L
20s	− 0.111	0.057	0.049	0.077	0.090	0.035
30s	− 0.185**	− 0.045	0.169	0.146*	0.188**	0.154**
40s	0.081	− 0.074	− 0.116	− 0.049	− 0.009	0.065
50s	− 0.074	0.112	0.012	− 0.069	0.066	− 0.021
60s	− 0.182**	0.190**	0.067	− 0.090	0.186**	0.127*
70s	− 0.224**	0.200**	− 0.013	− 0.024	− 0.011	− 0.187**
80s	− 0.111	0.178**	− 0.085	− 0.086	− 0.134*	− 0.143*
Total	− 0.727**	0.602**	0.257	− 0.371**	0.620**	0.362**

*P < 0.05, **P < 0.01. The description of dental codes is presented in Table 2.

Performance of regression models

The performance of regression models developed from various combinations of sextants in each dataset is presented in Fig. 1, Table 4, and Supplementary Table S6. The models were established using a training set of radiographs, and a separate test set, drawn from the same population, was employed to verify their performance. Model 2—incorporating the posterior teeth on both sides of both jaws—demonstrated the best performance, with a root mean square error (RMSE) of 13.144 years and adjusted coefficient of determination (adjusted R²) of 0.564. Model 1, which included all teeth except the third molars, was the second-best model, achieving an RMSE of 13.235 years and adjusted R² of 0.558. Model 3—which included the posterior teeth on both sides of the upper jaw—achieved an RMSE of 14.282 years and adjusted R² of 0.485, whereas Model 4—which included the posterior teeth on both sides of the lower jaw—achieved an RMSE of 14.490 years and adjusted R² of 0.470. When comparing performance using the training set to that using the test set, the RMSE values of models 1 to 4 increased by 2.515, 2.459, 2.118, and 2.031 years, respectively, from the values obtained when using the test set (Fig. 1 and Table 4). Despite these increases, the overall performance trends remained consistent between the training and test sets. Roh’s method had an RMSE of 12.952 years on the test set, whereas Model 2, the best-performing model in this study, had an increased RMSE of 15.603 years, showing a difference of 2.651 years (Fig. 2). Corresponding MAEs for each model are provided in Table 4 and Supplementary Table S6. Significant differences in MAE and RMSE between each developed model and Roh’s method were confirmed by the Wilcoxon signed-rank test (p < 0.05). Fisher Z-transformation analysis²⁴ revealed no statistically significant differences in adjusted R² between both sexes and sex-specific models (p > 0.05).

Fig. 1.

Performance comparison of developed models on training and test sets for both sexes. MAE mean absolute error, RMSE root mean square error, Tr training set, Te test set, Green represents mean absolute error (MAE) in years. Blue represents root mean square error (RMSE) in years. Error bars represent 95% confidence intervals.

Table 4.

Regression coefficients, intercepts, and performance of developed models for both sexes.

	Regression coefficient					Intercept	Training set		Test set		Adj. R2
Model	X	T	F	P	L		RMSE	MAE	RMSE	MAE
	(95% CI)	(95% CI)	(95% CI)	(95% CI)	(95% CI)	(95% CI)	(95% CI)	(95% CI)	(95% CI)	(95% CI)	(95% CI)
Model 1	1.621**	− 0.025	0.269**	1.920**	− 0.013	37.073	13.235	10.912	15.750	12.743	0.558
	(0.107)	(0.319)	(0.202)	(0.173)	(0.299)	(1.367)	(0.349)	(0.326)	(1.094)	(0.959)	(0.029)
Model 2	2.889**	− 0.073	0.111	2.632**	0.348	35.994	13.144	10.640	15.603	12.548	0.564
	(0.170)	(0.427)	(0.234)	(0.252)	(0.369)	(1.481)	(0.375)	(0.336)	(1.131)	(0.961)	(0.028)
Model 3	5.255**	0.047	0.585**	4.001**	1.004**	38.030	14.282	11.695	16.400	13.102	0.485
	(0.323)	(0.720)	(0.416)	(0.447)	(0.646)	(1.396)	(0.393)	(0.357)	(1.147)	(1.022)	(0.031)
Model 4	5.104**	0.354	− 0.418	3.579**	1.460**	40.935	14.490	11.760	16.521	13.348	0.470
	(0.362)	(0.723)	(0.462)	(0.442)	(0.661)	(1.497)	(0.407)	(0.369)	(1.089)	(1.009)	(0.032)

*P < 0.05, **P < 0.01. MAE mean absolute error, RMSE root mean square error, Adj. R² adjusted R², CI confidence interval.

The description of dental codes is presented in Table 2.

Fig. 2.

Comparison of accuracy between developed models and the conventional method on the test set for both sexes. MAE mean absolute error, RMSE root mean square error, CM conventional method, Conventional method represents Roh’s method¹⁵ and is shown in yellow, while the developed models are shown in pink. Error bars represent 95% confidence intervals.

Methods

Data collection

The required number of radiographs for this study was calculated using the statistical software G^*Power version 3.1.9.7 according to previous study²². An effect size of 0.15, a significance level of 0.05, a power of 0.80, and five variables were specified, indicating that a minimum sample size of 92 was required for regression analysis. According to Riley et al.²⁵, the most conservative criterion requires at least 239 subjects, and Schneider et al.¹⁶ recommend securing at least 20 observations per predictor variable in multiple linear regression, which corresponds to 100 subjects in this study. The training set of 2,030 radiographs greatly exceeded all these thresholds, enhancing model robustness and reducing the risk of overfitting. The inclusion criterion for the sample was dental panoramic radiographs of male and female patients aged between 20.00 and 89.99 years, which were retrospectively collected at Seoul St. Mary’s Hospital, The Catholic University of Korea. The exclusion criterion was radiographs displaying only sound teeth, without any observable dental characteristics. A total of 2,749 radiographs were collected, of which 358 were excluded, resulting in 2,391 radiographs (1,198 from males and 1,193 from females, mean age 55.02 ± 19.89 years) included in the analysis. The age distribution of the excluded radiographs is presented in Supplementary Table S7. The dataset was divided into a training set of 2,030 radiographs (1,015 from males and 1,015 from females, mean age 55.02 ± 19.93 years) for developing the regression model and a test set of 361 radiographs (183 from males and 178 from females, mean age 54.98 ± 19.70 years) for validating the model. The chronological age of each patient at the time of radiograph collection was calculated by subtracting the date of birth from the date the radiograph was taken. The radiographs were divided into seven decade-based age groups with nearly even distributions of age and sex. The age and sex distributions for each dataset are presented in Table 5, Supplementary Fig. S2, and Supplementary Table S8. Panoramic radiographs analyzed in this study were taken using either HDX WILL Q-Face (HDX Corp., Seoul, Republic of Korea) or Planmeca ProMax 2D (Planmeca Oy, Helsinki, Finland).

Table 5.

Age and sex distribution of the samples.

Age group (years)	Training set			Test set
	M	F	Total	M	F	Total
20.00–29.99	145	145	290	23	20	43
30.00–39.99	145	145	290	31	27	58
40.00–49.99	145	145	290	29	28	57
50.00–59.99	145	145	290	26	24	50
60.00–69.99	145	145	290	25	24	49
70.00–79.99	145	145	290	29	29	58
80.00–89.99	145	145	290	20	26	46
Total	1015	1015	2030	183	178	361

M radiographs of males, F radiographs of females.

Ethical considerations

This study was conducted in compliance with the Declaration of Helsinki and approved by the Institutional Review Board (IRB) of Seoul St. Mary’s Hospital, The Catholic University of Korea (approval no. KC24WISI0328). In this retrospective study, only the sex, date of birth, and imaging date of anonymized records were collected from the Clinical Data Warehouse (CDW) and used for research purposes. As no identifiable personal information was obtained, the IRB of Seoul St. Mary’s Hospital, The Catholic University of Korea waived the requirement for informed consent.

Categorization of dental characteristics

For statistical convenience, the dental characteristics observed in the radiographs were categorized into nine codes based on the system proposed by Cheong et al.²² (Table 2). These codes included sound tooth (V), missing tooth (X), impacted tooth (I), defect (D), residual root (R), root canal treatment (T), filling (F), prosthesis (P), and dental implant (L). Due to the high individual variability of the four third molars, only 28 teeth per radiograph were included in the analysis. Codes I, D, and R were excluded from the statistical analysis as they were observed at frequencies lower than 1%²⁶.

Design of age estimation models

Cheong et al.²² concluded that among various approaches, models employing either the entire dentition or only the posterior teeth from both jaws were the most suitable for practical application. Building on these findings, the present study analyzed similar models using the same teeth. The Wilcoxon signed-rank test was conducted to evaluate differences in the distribution of dental codes across tooth positions (upper, lower, left, and right). Statistically significant differences were observed between the upper and lower jaws, leading to separate statistical analyses for the posterior teeth on both sides of the upper and lower jaws. The resulting age estimation models are structured as follows:

• Model 1: All six sextants (all teeth excluding the four third molars)

• Model 2: Maxillary and mandibular right and left posterior sextants

• Model 3: Maxillary right and left posterior sextants

• Model 4: Mandibular right and left posterior sextants

Statistical analysis

All data underwent normality testing using the Shapiro–Wilk test, which indicated that the distributions were not normally distributed across all sexes and age groups. Intra- and inter-observer reliability were calculated using Cohen’s kappa statistics. A total of 361 radiographs, representing approximately 15% of all radiographs, were analyzed for observer reliability. Intra-observer reliability calculations were based on evaluations conducted by a single observer with a two-month interval between assessments. Spearman’s correlation coefficients were calculated to elucidate the relationship between the number of codes observed in a tooth and chronological age. Multiple regression analysis was conducted for each combination of sextant-based dental characteristics using best-subsets regression procedures. To assess multicollinearity among the six dental codes (V, X, T, F, P, and L), variance inflation factors (VIFs) were calculated. The V code exhibited a notably high VIF exceeding 40 and was therefore excluded from the regression analysis as part of the variable selection process to reduce model instability and overfitting risk. The remaining five predictors (X, T, F, P, and L) all had VIFs approximately below 3, which is well below the commonly used threshold of 5 for acceptable multicollinearity, as presented in Supplementary Table S9. Given the statistically significant distribution of the number of dental codes and sex observed in the Mann–Whitney U test, the regression analysis was performed with male, female, and pooled sex data, separately. Model performance was evaluated using the adjusted R² calculated from the training set. Statistical significance of differences in adjusted R² between both sexes and sex-specific models was assessed using Fisher’s Z-transformation method²⁴. The statistical significance and 95% confidence intervals (CI) for each regression coefficient were evaluated. Although some variables showed statistically insignificant results (p > 0.05), they were retained in the model based on their established clinical relevance and practical interpretability, in accordance with the recommendations of Chowdhury et al.²⁷, which advise against excluding clinically important predictors only on the basis of statistical thresholds. To assess model performance, the MAE and RMSE were calculated for each model on both the training and test sets. Additionally, these metrics were compared between the developed models and Roh’s method¹⁵, for which MAE and RMSE were calculated using the test set, to evaluate practical applicability. The statistical significance of differences in MAE and RMSE between the developed models and Roh’s method was assessed using the Wilcoxon signed-rank test. The metrics of model performance are included with their 95% CI. Data were expressed as means ± standard deviations or numbers (%), contingent upon the data characteristics. Statistical significance was defined by a two-tailed P-value < 0.05. All statistical analyses were conducted using SPSS Statistics version 29.0 (IBM Co., Armonk, NY, USA).

Discussion

In this study, we developed a non-destructive age estimation model for Korean adults using regression analysis based on the number of dental codes observed in each combination of sextants in panoramic radiographs. The findings demonstrated significant correlations between certain dental codes and chronological age, providing a basis for dental feature-based age estimation. Model 2, derived from the dental characteristics of the posterior teeth on both sides of both jaws, showed the best performance, with an RMSE of 13.144 years and adjusted R² of 0.564. Model 1, which included all sextants, exhibited performance closely comparable to that of Model 2, with an RMSE of 13.235 years and adjusted R² of 0.558. These findings demonstrated an improvement when compared with the study by Cheong et al.²², which reported an RMSE of 14.78 years and an adjusted R² of 0.461 for the posterior sextant model. The observed improvement in model performance can be attributed to several factors. First, excluding the anterior teeth, which were less associated with dental codes such as T, F, and P and thus reflect fewer age-related treatments, likely improved the accuracy of the age estimation model by focusing on more relevant dental changes, as proposed by Cheong et al.²². Second, radiographs containing only sound teeth were excluded, mitigating potential data distortion caused by the absence of age-related dental changes and enabling the model to more effectively represent dental characteristics across different age groups. Third, the sample size in this study was approximately 5.7 times larger than that in the research conducted by Cheong et al.²², which allowed for more comprehensive model training. However, across models, the RMSE on the test set exceeded that on the training set by 2.031–2.515 years, indicating possible overfitting or sample imbalance. Therefore, model robustness should be interpreted with caution.

To validate the applicability of the developed model in Korean forensic practice, its performance was compared to Roh’s method¹⁵, a non-destructive approach commonly used in forensics, by applying both methods to the same test set. Roh’s method, which employs quantitative measurements of six teeth observable on panoramic radiographs, rely on structural changes caused by secondary dentine deposition and thus reduces variability, potentially enhancing correlations with chronological age. In contrast, dental characteristics influenced by treatments and aging may show remarkable variations among individuals due to differences in dental care and lifestyle habits, introducing variability and potentially limiting their correlation with age. Applied to the test set, Roh’s method achieved an RMSE of 12.952 years, approximately 2.651 years superior to the RMSE of 15.603 years observed in Model 2. Statistically significant differences in RMSE were observed between the developed model and Roh’s method. However, the adjusted R² of 0.564 for Model 2 was higher than the R² of 0.47 reported by Roh et al.¹⁵, indicating that the fit of the developed model demonstrated acceptable performance. Moreover, considering that Roh’s method requires time-consuming, quantitative measurements of histological structures using software such as Photoshop, the model proposed in this study provides an advantage in convenience, as it utilizes dental characteristics directly observable on panoramic radiographs. While the developed model may not be sufficient as a standalone method due to the abovementioned individual variability, it could serve as an adjunct tool with conventional methods like Roh’s and enhance the accuracy and efficiency of age estimation.

Statistically significant differences in the distribution of the number of specific dental codes by sex and between the upper and lower jaws were observed in this study, prompting the development of separate regression models to reflect the relationship between the number of dental codes and chronological age. As presented in Table 4 and Supplementary Table S6, separate regression models were established for each sex, along with models based on upper jaw data (Model 3) and lower jaw data (Model 4). Azrak et al.²⁰ and Lee & An²⁸ had demonstrated improved model performance when incorporating both dental characteristics and additional oral findings as predictive variables in sex-specific models. However, despite significant differences in the distributions of the number of codes F and L by sex in this study, the performances of the separate models for each sex in Model 2 were similar, with adjusted R² of 0.558 and 0.570 and RMSEs of 13.147 and 13.096 years for males and females, respectively. These results suggest that the performance of the models remained consistent across sexes, likely due to the predominant influence of codes X and P, which constituted a substantial proportion of the overall code distribution for each sex. Furthermore, no statistically significant differences in adjusted R² were found between the both sexes and sex-specific models. Nevertheless, these should be interpreted with caution, as they may reflect potential imbalances in the training sample or inherent characteristics of the source data. Significant differences were also noted in the distributions of the number of codes X, T, F, and P between the upper and lower jaws. However, comparisons of Models 3 and 4 across all radiographs revealed similar performance, with adjusted R² of 0.485 and 0.470 and RMSEs of 14.282 and 14.490 years, respectively. Despite differences in code distribution, if the overall codes were observed with high frequency in both the upper and lower jaws, these may have mitigated performance discrepancies. The similar performance observed in both sex- and jaw-specific models suggests that developing separate models may not be necessary and that using a unified model could be more efficient.

Previous studies have demonstrated correlations between dental characteristics and chronological age. Azrak et al.²⁰ developed an age estimation regression equation based on dental findings, including features associated with codes X, T, F, and P, as well as the alveolar bone level. These features exhibited significant positive correlations with age. However, as their method used a total score derived by summing the scores of each feature as the variable, the specific correlations between individual features and age were not analyzed. Yamashita et al.²¹ reported significant correlations between age and features related to codes X and P in Japanese adults, highlighting the number of teeth, dental prostheses, and attrition as valuable age indicators. Kawashima et al.²⁹ demonstrated that incorporating the number of dental prostheses into regression models improved age estimation accuracy. In Korean populations, Lee & An²⁸ and Lee et al.³⁰ observed that the number of dental features related to codes X, T, P, and L increased with age from the 10 s to the 60 s, suggesting the potential of dental characteristics as age markers. However, in Lee & An’s study²⁸, the correlation between age and features associated with code X was notably weak (0.108), which may be caused by the inclusion of teenagers, who predominantly have sound teeth, and the exclusion of individuals over 70 years of age, who are more likely to exhibit tooth loss. As a result, the frequency of code X-related features in their dataset was lower, leading to a weaker correlation compared to the moderate correlation of 0.602 observed in this study. For features associated with code P, similar correlations with age—0.606 in Lee & An’s study²⁸ and 0.620 in the present study—were reported. These comparable results suggest that prosthetic treatment patterns in overlapping age ranges from the 20 s to 60 s were similar across the studies. Lee & An²⁸ reported an adjusted R² of 0.623 for their model, slightly exceeding the current study model’s performance of 0.564, which may be attributable to the inclusion of additional oral findings, such as periodontitis, pulp area, and tooth length, as predictive variables in Lee & An’s study²⁸. In contrast, the present study relied on the number of dental characteristics directly observable on panoramic radiographs, offering a faster and more efficient approach without requiring additional measurements.

Human aging is accompanied by irreversible dental treatments that typically follow a unidirectional trajectory of changes in dental characteristics. In the early stages of treatment, dental decay is removed, and the cavity is filled with materials such as resin or gold (F). As decay progresses to involve larger portions of the crown, the affected area is replaced with prosthetics (P). When decay reaches the pulp, root canal treatment (T) is performed by removing pulp and nerve tissues and sealing the canal. If the tooth becomes too compromised for restoration, it is extracted and replaced with an implant (L) to address the missing tooth (X). However, root canal treatment may be performed independently, meaning the progression from filling to prosthetic to root canal treatment, extraction, and implant does not always follow a strict sequence, requiring careful interpretation of the treatment process. In this study, code F was more frequently observed in younger individuals; code P became prevalent starting in middle age and continuing into older age, whereas code T showed similar frequencies from middle age to older age. Codes X and L were more commonly observed in older individuals (Supplementary Fig. S1). Notably, codes X and P exhibited moderate positive correlations with age, suggesting an increase in tooth loss and prosthetic treatments as individuals age. This finding aligns with the understanding that dental health deteriorates over time due to factors such as decay, periodontal disease, and attrition^{11,20,28–35}.

Table 3 shows that, while moderate to weak correlations were observed between age and the number of dental characteristics across the total age range, the correlations within each 10-year age group were markedly weaker. This discrepancy may be explained by the gradual progression of dental changes, such as the number of treated or missing teeth, which may not be sufficiently pronounced within a limited 10-year age interval. As a result, the narrow age range of each decennial group may be limited in its ability to capture the cumulative changes in dental characteristics across the full adult lifespan, thereby restricting the detection of broader age-related trends. Previous studies^16,27 have also shown that predictors with weak univariable correlations may improve predictive accuracy in multivariable regression when combined, owing to complementary and interactive effects among variables. Consistent with these findings, the regression model trained on the full age range in this study achieved a moderate adjusted R² of 0.564 in Model 2 for both sexes. However, given the weak correlations between the number of dental characteristics and age within each 10-year group, such age estimation models should be applied with careful consideration of the specific age range under investigation. Importantly, the observed model performance was obtained using a large sample size, exclusion of highly collinear variables, and independent test set validation, which together reduce the likelihood that the adjusted R² is inflated by overfitting.

The negative regression coefficients for code T, L and F do not indicate that individuals with these treatments are biologically younger. Rather, they reflect adjusted relationships within multivariable models, influenced by treatment trajectories and statistical adjustments. For code F, prevalence decreases after middle age (Supplementary Table S4) as restorative needs shift toward more invasive treatments, such as P, L, and X. This decline is greater in males, and, because equal numbers of cases were allocated to each decade, the turning point in males may have placed greater weight on decreasing segments across the full age range, yielding a negative coefficient. In contrast, the turning point in females likely weighted increasing segments more heavily due to the relatively less pronounced decline, resulting in a positive coefficient. This sex-specific pattern is consistent with previous reports of higher caries prevalence in males³⁶ and greater restorative intervention in mandibular molars³⁷. For code T, its low overall prevalence (7.9%, Table 1) and the tendency for root canal–treated teeth to be extracted in older adults³⁸ account for its decline in later life stages. Similarly, code L, which typically follows extractions, becomes less common in the oldest patients due to the shift toward conservative treatment preferences, such as opting for removable dentures instead of implant placement³⁹. From a statistical perspective, all three predictors (T, L, and F) exhibited variance inflation factor (VIF) values approximately below 3, well under the conventional threshold of 5, indicating that the sign changes are unlikely to result from problematic multicollinearity. Rather, they represent adjusted relationships that emerge after accounting for other correlated dental codes in a multivariable context. In line with Chowdhury et al.’s²⁷ recommendation that clinically important predictors should not be excluded only on the basis of statistical non-significance, these variables were retained, supported by previous studies^20,28,30 reporting their significant correlations with chronological age. Nevertheless, these non-significant coefficients should be interpreted cautiously, as they may reflect complex, non-linear relationships rather than simple monotonic trends with age. Analysis of the regression coefficients and their 95% confidence intervals revealed that codes X and P consistently had the largest positive coefficients across models, with relatively narrow confidence intervals, indicating stable and substantial contributions to age estimation. In contrast, codes T, L, and F generally exhibited smaller absolute coefficients with wider confidence intervals, suggesting higher variability and less consistent associations with age. These patterns were similar in both sexes and sex-specific models (Table 4 and Supplementary Table S6). Future studies should include effect size comparisons to more accurately delineate the influence of each code in multivariable regression models and to better identify the impact of complex variables.

Previous studies on age estimation methods are limited by small sample sizes^{21,22,28–30,40} or the absence of regression Eqs. ^30,40. Although our study addressed these limitations, it has some limitations that need to be acknowledged. First, a common limitation of previous studies has been the potential for data bias and limited generalizability due to reliance on data from a single institution. In this study, the dataset was also restricted to a single institution, with radiographs predominantly obtained from university hospital visitors, who may not represent the general population^41–44. Future research should incorporate external validation using data from multiple institutions with varied clinical settings and treatment patterns to determine whether the effects observed are consistent across different sources. Additional overfitting mitigation strategies, such as k-fold cross-validation, should be employed to rigorously confirm the robustness and generalizability of the model. Second, this study focused on five dental codes for regression analysis, potentially overlooking socio-environmental factors such as oral hygiene, dietary habits, and lifestyle^{20,30,41,44–46}. While collecting such comprehensive data would require substantial effort and cost and may be impractical for forensic applications, exploring the impact of these variables could provide valuable insights into age estimation. Third, the developed regression model in this study relied on the total number of dental codes as the variable, making it challenging to identify correlations between age and specific dental codes. Future studies should aim to construct age estimation models using variables encoded for individual teeth, enabling the assessment of tooth position-specific contributions to age estimation. Incorporating variables for specific teeth and codes would create numerous variables, complicating manual analyses. However, applying machine learning techniques, which are already widely used in dental and medical data analysis, could not only address these challenges by efficiently handling the large number of variables and their complex interactions, but also enable more accurate modeling of dental characteristics that follow nonlinear or piecewise age related patterns^40,47,48.

In conclusion, this study analyzed the correlations between dental characteristics observable in panoramic radiographs and chronological age. The results confirmed the potential of dental characteristics as novel age indicators and led to the development of a non-destructive regression model for age estimation in Korean adults. Based on its practical feasibility, Model 2 is recommended as the default approach for future dental characteristic–based age estimation in clinical and forensic settings. Although the developed models are convenient to use, in this study they showed lower accuracy than conventional method used for comparison. Therefore, they should be considered supplementary tools used alongside conventional methods rather than direct replacements.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

S.O., J.P., B.Y.R., and S.S.L. performed the experiments and wrote the manuscript. A.K. and S.S.L. designed the study and reviewed the manuscript. (S.S.L.: https://orcid.org/0000–0002-0171-561X).

Funding

This study was supported by a National Research Foundation of Korea (NRF) grant (No. 2022R1F1A1063719).

Data availability

All data supporting the findings of this study are available within the article and its supplementary files.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

This study was approved by the Institutional Review Board (IRB) of Seoul St. Mary’s Hospital, The Catholic University of Korea (approval no. KC24WISI0328). This retrospective study used medical records, and the IRB of Seoul St. Mary’s Hospital, The Catholic University of Korea waived the requirement for informed consent. This material was based on “CMC nu CDW” and “CMC integrated DATA” at the Catholic University of Korea, Catholic Medical Center. Thus, any research results obtained from them do not reflect relationships or a perspective from the Catholic University of Korea, Catholic Medical Center, and its affiliated hospitals.