Exploring optimal methods for age-at-death estimation using pulp/tooth area ratios: a South African study
- Open Access
- 01.11.2024
- Original Article
Erschienen in:
Abstract
Introduction
Skeletal age-at-death estimation is fundamental to biological anthropology, providing valuable insights into past population demographics, health, lifeways, and evolutionary trends [1, 2]. In forensic medicine and science, it plays a crucial role in the identification of unknown human decedents, contributing significantly to an individual’s biological profile, alongside sex, stature, and ancestry analyses [3, 4]. This information aids law enforcement in refining missing persons lists, thereby improving the efficiency of medico-legal investigations and identification efforts [5, 6]. Identifying deceased individuals holds profound ethical and judicial value, helping resolve various legal issues, deliver justice, and provide closure to affected families and communities [7‐9].
Among various biological indicators, teeth emerge as reliable age markers due to their durability and resistance to taphonomic changes [10, 11]. Adult dental ageing methods rely on post-formation, regressive changes that typically exhibit greater variability and inaccuracy compared to subadult indicators [12]. These techniques encompass visual, histological, radiographic, biochemical and genetic/epigenetic approaches [13]. Radiographic methods, particularly those evaluating secondary dentine apposition, are valued for their efficiency, minimal invasiveness and relative accuracy [14]. Secondary dentine deposition occurs throughout life following root formation, gradually reducing a tooth’s pulp cavity, providing useful age-related information [15].
Anzeige
The pulp/tooth area ratio (PAR) method, pioneered by Cameriere et al. [16‐18], is a widely utilised adult ageing technique, which quantifies changes in pulp cavity size from radiographs as an indirect measure of secondary dentine apposition. The PAR has shown strong correlation with chronological age, offering reliable, sex-independent and relatively accurate age estimates, with canine mean absolute errors as low as 2.43 years in certain populations [14, 19]. Despite nearly two decades of evaluation, several research gaps persist for the PAR method. Notably, studies on pulp/tooth volume ratios have shown better age estimation accuracy when excluding enamel dimensions from analyses [20, 21]; however, such research for PARs is lacking. Moreover, while labiolingual radiographic views are predominantly used for PAR assessments, the relative utility of mesiodistal views is inadequately investigated [18, 22]. Additionally, limited research has compared age estimates from radiographs versus tooth section images, often relying on published equations developed for radiographic data, which may introduce biases and compromise comparison validity [23, 24].
This study aimed to evaluate the applicability of the PAR method for estimating adult age-at-death using maxillary canines from a South African cadaveric sample. Objectives included (1) assessing the reliability of the PAR method, (2) examining sex differences in PAR values, (3) comparing the performance of the PAR method across different image types (labiolingual periapical radiographs, mesiodistal periapical radiographs, labiolingual stereomicroscopic tooth section images), and (4) contrasting age estimation models derived from PAR values with and without consideration of enamel areas. Such validation is critical to account for population-specific variations with ageing [25], contributing valuable data to the limited research on dental ageing techniques for South African adults [26, 27]. This study also provides insights for global practitioners on optimal approaches when utilising the PAR method for maxillary canines and introduces a modified stereomicroscopic technique.
Materials & methods
Ethical clearance was granted by the Human Research Ethics Committee at the University of Cape Town (UCT) (HREC REF: 336/2022). Permission was obtained from UCT’s Cadaver Research Governance Committee (CRGC 2022/001), as well as the Division of Clinical Anatomy, Faculty of Medicine and Health Sciences, Stellenbosch University (SU), and the Provincial Inspector of Anatomy. This research, conducted during 2022 and 2023, complies with the STROBE protocol for observational, cross-sectional studies.
The sample included 52 individuals with known chronological age and sex (Table 1). The number of suitable teeth available during the study period determined the sample size. Maxillary canines were sourced from cadaveric human donor bodies at the Department of Human Biology, UCT (29 individuals) and the Division of Clinical Anatomy, SU (23 individuals). The sample comprised 25 males and 27 females, aged between 26 and 91 years (mean age ± standard deviation = 62.4 ± 19.9 years). According to age categories defined by Botha and Steyn [28], the sample included 32 older adults (≥ 60 years), 8 middle-aged adults (40–59 years) and 12 younger adults (16–39 years).
Anzeige
Table 1
Summary statistics for the sample demographic variables ─ sex and age (years)
Age category | Age interval | Males (n) | Females (n) | Pooled sexes (n) | Mean age ± SD |
|---|---|---|---|---|---|
Younger adults | 20–29 | 0 | 3 | 3 | 27.7 ± 1.5 |
30–39 | 7 | 2 | 9 | 35.1 ± 3.5 | |
Middle-aged adults | 40–49 | 1 | 1 | 2 | 42.5 ± 3.5 |
50–59 | 3 | 3 | 6 | 54.5 ± 2.6 | |
Older adults | 60–69 | 6 | 2 | 8 | 63.3 ± 2.9 |
70–79 | 3 | 8 | 11 | 74.9 ± 2.9 | |
80–89 | 5 | 6 | 11 | 83.8 ± 3.1 | |
90–99 | 0 | 2 | 2 | 90.5 ± 0.7 | |
Total | 25 | 27 | 52 | 62.4 ± 19.9 |
Tooth sampling & selection
Aligned with Merdietio Boedi [29], only adults (≥ 18 years) with at least one fully erupted, mature (closed apex), and healthy maxillary canine were included. Previous research found no significant PAR measurement differences between the left and right sides [14]. Therefore, the right tooth was prioritised for analysis; the left was used when the right was unavailable or unsuitable. Teeth were excluded if they showed dental pathologies (e.g. caries and traumatic lesions), developmental anomalies (e.g. impaction and multiple roots), artificial modifications (e.g. restorative procedures) and/or severe dental wear, characterised by at least Phase G of the Lovejoy [30] scoring system.
Radiography
Radiographic imaging was conducted using an Intra-oral X-ray Machine (ACTEON X-MIND), with exposure parameters: 70 kv; 8 mA; 0.50 ms. Periapical radiographs were obtained using indirect digital imaging, where the receptor was scanned (Intra-oral Scanner - Carestream CS7600) following exposure to X-radiation.
Each tooth was radiographed separately using the paralleling technique [31], capturing labiolingual and mesiodistal image projections. Positioned on a flat surface, each tooth was placed directly on a phosphor plate receptor and secured with Sellotape. This technique reduced the object-receptor distance, minimising magnification and loss of definition [31]. The cone (length = 167 mm) was centred over the flat surface, covering the phosphor plate and tooth, thereby standardising the target-receptor distance and limiting motion.
Sectioning & stereomicroscopy
Radiographed teeth were sectioned for stereomicroscopic analysis. The ideal cutting plane was marked on the mesial and distal tooth surfaces prior to embedding each tooth individually in an epoxy resin block (Kristal50 [AMT Composites, Maitland, Cape Town]). Embedded teeth were sectioned longitudinally along the mesiodistal plane (non-obliquely through the crown’s long axis) using a Buehler Isomet saw with a diamond wafering blade (200 RPM, 3 mm/min). This produced thick labial and lingual portions, exposing the labiolingual view of the pulp cavity and surrounding tissue. This plane of sectioning was chosen as the mesiodistal view presented minimal root curvature and large pulp cavity dimensions, ensuring the consistent acquisition of complete tooth sections.
Tooth sections were examined under a Zeiss Discovery V20 stereomicroscope with an Axiocam 503 colour camera at 10x magnification. Each tooth was imaged in three parts and digitally stitched together using Hugin® 2022.0.0 (open-source panorama photo stitcher, Pablo d’Angelo) to create a single composite image.
Image analyses
Radiographic and stereomicroscopic images were analysed using ImageJ 1.54b (open-source imaging processing program, National Institutes of Health, US). Following D’Ortenzio et al. [23], areas (mm²) were calculated using the polygon selection tool, with at least 50 points marked for each tooth outline (including and excluding the enamel) and 40 points for each pulp cavity outline (Fig. 1).
Due to the curvature and tapering of root canals towards the apex, observation of the entire pulp cavity in this region post-sectioning was not always feasible. Therefore, measurements were confined to the coronal half of each tooth section, systematically obtained as follows (Fig. 1): (1) Tooth sections were imaged with the crown’s long axis aligned approximately parallel to the vertical plane; (2) Two tangential horizontal lines (L1 and L2) were drawn along the most coronal and apical aspects of the tooth area outline; (3) A temporary vertical line (L3) connected L1 and L2; (4) A third horizontal line (L4) was drawn through the midpoint of L3, dividing the tooth into two portions; (5) Only the region coronal to L4 was considered for measurements.
Anzeige
Fig. 1
Labiolingual radiograph (top row), mesiodistal radiograph (middle row) and stereomicroscopic tooth section image (bottom row) measurements using ImageJ. a) unmeasured labiolingual radiograph, b) pulp cavity area measured from labiolingual radiograph, c) complete tooth area measured from labiolingual radiograph, d) tooth area (excluding enamel) measured from labiolingual radiograph, e) unmeasured mesiodistal radiograph, f) pulp cavity area measured from mesiodistal radiograph, g) complete tooth area measured from mesiodistal radiograph, h) tooth area (excluding enamel) measured from mesiodistal radiograph, i) unaltered stereomicroscopic tooth section image, j) lines (L1 – L4) drawn/positioned consecutively to systematically obtain the region of interest for conducting measurements, k) pulp cavity area measured from stereomicroscopic tooth section image, l) complete tooth area measured from stereomicroscopic tooth section image, m) tooth area (excluding enamel) measured from stereomicroscopic tooth section image
Statistical analyses
Unless noted otherwise, data analyses used the IBM® SPSS® Statistics software (version 28), with p ≤ 0.05 indicating statistical significance. PARs were calculated by dividing pulp cavity area by total tooth area (including and excluding enamel).
Intra- and inter-observer reliability were assessed by re-examination of 19 randomly selected images (of each type) two weeks after initial analysis. Inter-observer analyses were performed by a single individual with experience and training on the measurement technique. Agreement between raters was determined using the intraclass correlation coefficient (ICC) [32].
Analysis of covariance (ANCOVA) assessed sex effects on PARs, controlling for age. Pearson’s correlation was used to evaluate linearity between age and PARs based on criteria by Rowntree [33]. Moderation analysis via PROCESS macro (v4.2) in SPSS [34] explored interaction effects. Additionally, multicollinearity among predictors was assessed using the variance inflation factor (VIF > 5 indicates problematic multicollinearity) [35].
Anzeige
The relationship between age-at-death (response variable) and PARs (predictor variables) was evaluated through linear regression. Best-subsets regression, conducted in JMP® 17.1 (SAS Institute Inc., Cary, NC, 1989–2023), compared age estimation models using statistical criteria, including the adjusted coefficient of determination (R2-adjusted), standard error of the estimate (SEE), Akaike information criterion adjusted for sample size (AICc), Bayesian information criterion (BIC), and Mallows’ Cp statistic (Cp) [36]. Models were selected based on superior performance across these criteria, with only the most effective ones advancing. Additional parameters, including the mean absolute error (MAE), coefficient of determination (R2), standard deviation (SD) of the absolute residuals, and a 95% prediction interval, were also reported. Additionally, models underwent a post hoc power analysis using G*Power 3.1.9.7 [37]. to ensure adequate power (≥ 0.80).
Selected models underwent a leave-one-out cross-validation (LOOCV), performed in RStudio® 2023.03.1.446 using the trainControl function from the caret library [38]. Cross-validation parameters, including the coefficient of determination (R2CV), mean absolute error (MAECV) and standard error of the estimate (SEECV), were calculated to evaluate model performance. Optimism analysis, comparing cross-validation and in-sample results, gauged predictive bias.
A two-way repeated measures analysis of variance (ANOVA) (with two within-subjects factors) analysed effects of image type and enamel area on absolute estimate residuals/errors, followed by Fisher’s least significant difference procedure for multiple comparisons.
Results
All variables demonstrated ICC estimates above 0.9 during both intra- and inter-observer reliability assessments, indicating excellent data reproducibility (Table 2). Age significantly correlated with all PAR variables in the ANCOVA (p < 0.001), while sex did not exhibit a significant effect on these variables (Table 3). Consequently, sex-specific age estimation models were not developed for the sample. The moderation analysis did not reveal any significant first-order interaction effects among predictor variables (see Online Resource 1).
Anzeige
Table 2
Observer reliability assessment results for each variable
Inter-observer reliability | Intra-observer reliability | ||||||
|---|---|---|---|---|---|---|---|
Variables | ICC | 95% CI | Variables | ICC | 95% CI | ||
Lower bound | Upper bound | Lower bound | Upper bound | ||||
PA1 | 0.94 | 0.66 | 0.98 | PA1 | 0.99 | 0.98 | 1.00 |
TA1 | 1.00 | 1.00 | 1.00 | TA1 | 1.00 | 1.00 | 1.00 |
TA_EE1 | 0.99 | 0.97 | 1.00 | TA_EE1 | 1.00 | 0.99 | 1.00 |
PAR1 | 0.92 | 0.50 | 0.98 | PAR1 | 0.99 | 0.98 | 1.00 |
PAR_EE1 | 0.93 | 0.67 | 0.98 | PAR_EE1 | 0.99 | 0.98 | 1.00 |
PA2 | 0.95 | 0.62 | 0.99 | PA2 | 0.98 | 0.95 | 0.99 |
TA2 | 1.00 | 1.00 | 1.00 | TA2 | 0.95 | 0.88 | 0.98 |
TA_EE2 | 1.00 | 0.98 | 1.00 | TA_EE2 | 0.94 | 0.85 | 0.98 |
PAR2 | 0.92 | 0.47 | 0.98 | PAR2 | 0.96 | 0.89 | 0.98 |
PAR_EE2 | 0.94 | 0.60 | 0.98 | PAR_EE2 | 0.95 | 0.88 | 0.98 |
PA3 | 0.99 | 0.94 | 1.00 | PA3 | 1.00 | 1.00 | 1.00 |
TA3 | 0.98 | 0.95 | 0.99 | TA3 | 1.00 | 1.00 | 1.00 |
TA_EE3 | 0.97 | 0.91 | 0.99 | TA_EE3 | 1.00 | 1.00 | 1.00 |
PAR3 | 0.99 | 0.91 | 1.00 | PAR3 | 1.00 | 1.00 | 1.00 |
PAR_EE3 | 0.99 | 0.97 | 1.00 | PAR_EE3 | 1.00 | 1.00 | 1.00 |
Table 3
Analysis of covariance (ANCOVA) results for assessing the main effect of sex on the pulp/tooth area ratio (PAR) variables, with age as the covariate
Variables | df | SS | MS | F-statistic | p |
|---|---|---|---|---|---|
PAR1 | |||||
Age | 1 | 0.02 | 0.02 | 83.63 | < 0.001 |
Sex | 1 | 3.15e− 4 | 3.15e− 4 | 1.34 | 0.253 |
Residuals | 49 | 0.01 | 2.36e− 4 | ||
PAR_EE1 | |||||
Age | 1 | 0.03 | 0.03 | 86.88 | < 0.001 |
Sex | 1 | 9.67e− 5 | 9.67e− 5 | 0.31 | 0.579 |
Residuals | 49 | 0.02 | 3.10e− 4 | ||
PAR2 | |||||
Age | 1 | 0.03 | 0.03 | 32.79 | < 0.001 |
Sex | 1 | 3.68e− 3 | 3.68e− 3 | 3.74 | 0.059 |
Residuals | 49 | 0.05 | 9.86e− 4 | ||
PAR_EE2 | |||||
Age | 1 | 0.04 | 0.04 | 35.03 | < 0.001 |
Sex | 1 | 3.31e− 3 | 3.31e− 3 | 2.72 | 0.105 |
Residuals | 49 | 0.06 | 1.22e− 3 | ||
PAR3 | |||||
Age | 1 | 0.04 | 0.04 | 124.13 | < 0.001 |
Sex | 1 | 2.16e− 4 | 2.16e− 4 | 0.76 | 0.386 |
Residuals | 49 | 0.01 | 2.83e− 4 | ||
PAR_EE3 | |||||
Age | 1 | 0.06 | 0.06 | 148.35 | < 0.001 |
Sex | 1 | 1.86e− 5 | 1.86e− 5 | 0.05 | 0.828 |
Residuals | 49 | 0.02 | 3.92e− 4 |
Pearson correlation coefficients and scatter plots indicated a moderate to strong negative linear relationship between PAR variables and age (Fig. 2). Stereomicroscopic tooth section images showed the strongest correlation values (r = -0.85 and − 0.87), followed by labiolingual radiographs (r = -0.80 and − 0.81) and mesiodistal radiographs (r = -0.65 and − 0.66), respectively. Notably, for each image type, PAR variables excluding enamel area exhibited stronger correlations with age-at-death. Figure 2 illustrates age-related changes in maxillary canine PARs: as individuals age, their pulp cavities become increasingly obliterated, thereby reducing the ratio of pulp cavity to tooth area. Summary statistics for the PAR variables are provided in Online Resource 1.
Fig. 2
Age-related changes in maxillary canine pulp/tooth area ratios (PARs) observed from labiolingual periapical radiographs (a), mesiodistal periapical radiographs (b) and labiolingual stereomicroscopic tooth section images (c). For the dental images on the left side, each column shows three different image types from one individual, while each row shows three of the same image types from individuals of different ages. On the right side, scatter plots show the direction and strength of the linear relationship between age-at-death and the PAR variables. The Pearson correlation coefficient (r) and associated statistics are reported on each graph. As individuals age, their pulp cavities (indicated by the arrows) become increasingly obliterated, reducing the ratio of pulp cavity to tooth area. PAR1 = pulp/tooth area ratio obtained from labiolingual radiograph; PAR_EE1 = pulp/tooth area ratio (excluding enamel) obtained from labiolingual radiograph; PAR2 = pulp/tooth area ratio obtained from mesiodistal radiograph; PAR_EE2 = pulp/tooth area ratio (excluding enamel) obtained from mesiodistal radiograph; PAR3 = pulp/tooth area ratio obtained from stereomicroscopic tooth section image; PAR_EE3 = pulp/tooth area ratio (excluding enamel) obtained from stereomicroscopic tooth section image; n = sample size
Anzeige
Model selection
Employing a best-subsets regression approach, 63 candidate models were initially developed from the six PAR predictor variables. Models with high multicollinearity (VIF values > 5, see Online Resource 1) were first eliminated. Based on the selection criteria, 14 suitable models were chosen for age estimation (Table 4). These included six simple and eight multiple (incorporating two predictor variables) linear regression models (see Online Resource 1 for ANOVA and coefficients table outputs).
Among the simple regression models, those derived from stereomicroscopic tooth section images showed superior performance for age estimation, followed by models from labiolingual and mesiodistal radiographs, respectively (Table 4). Notably, models excluding enamel area consistently outperformed those including enamel within each image type group. Some multiple regression models combining PARs from labiolingual radiographs and stereomicroscopic tooth section images, as well as labiolingual and mesiodistal radiographs, demonstrated enhanced performance compared to simpler models.
Models based on stereomicroscopic tooth section images achieved the best MAE (7.21–7.47 years), SEE (9.88–10.53 years) and R2 (0.73–0.76) values. Models derived from labiolingual radiographs followed (MAE = 9.52–9.76 years; SEE = 11.90–12.03 years; R2 = 0.64–0.65), with those derived from mesiodistal radiographs performing less accurately (MAE = 11.87–11.93 years; SEE = 15.11–15.33 years; R2 = 0.42–0.44) (Tables 4 and 5). Within each image type group, the model excluding enamel area showed superior MAE, SEE and R2 values.
Most selected multiple regression models exhibited minor accuracy improvements (less than one year) over corresponding simpler models (Tables 4 and 5). Models combining data from stereomicroscopic tooth section images and labiolingual radiographs achieved MAEs and SEEs of 6.97–7.47 years and 9.65–10.20 years, respectively, with R2 ranging between 0.75 and 0.77. Those combining PARs from labiolingual and mesiodistal radiographs demonstrated MAEs and SEEs of 9.13–9.20 years and 11.76–11.92 years, respectively, while R2 ranged between 0.66 and 0.67.
Anzeige
The 95% prediction interval ranged from ± 19.85 years to ± 30.79 years for simple regression models and ± 19.40 years to ± 23.95 years for multiple regression models (Table 5). Residuals from all models were normally distributed with equal variance according to Shapiro-Wilk and Breusch-Pagan tests (see Online Resource 1). Additionally, all models demonstrated large effect sizes and sufficient statistical power (≥ 0.80) (Table 5).
Table 4
Best-subsets regression results for all suitable models chosen. Models are grouped and ranked separately based on the number of predictor variables (k) they contain. Within each size group (k), they are ordered based on performance – from top (best) to bottom (worst) – as determined by selection criteria values
k | Image type | Model | Performance/selection criteria | ||||
R2-adjusted | SEE | AICc | BIC | Cp | |||
1 | STSI | Age = 91.83 – 439.05(PAR_EE3) | 0.75 | 9.88 | 390.25 | 395.60 | 5.49 |
Age = 91.47 – 534.64(PAR3) | 0.72 | 10.53 | 396.86 | 402.22 | 12.74 | ||
LR | Age = 117.77 – 548.95(PAR_EE1) | 0.64 | 11.90 | 409.55 | 414.91 | 29.53 | |
Age = 118.29 – 625.44(PAR1) | 0.64 | 12.03 | 410.74 | 416.10 | 31.32 | ||
MR | Age = 100.82 – 281.03(PAR_EE2) | 0.42 | 15.11 | 434.46 | 439.81 | 77.16 | |
Age = 100.24 – 307.84(PAR2) | 0.41 | 15.33 | 435.91 | 441.27 | 80.72 | ||
2 | LR + STSI | Age = 101.32 – 161.82(PAR_EE1) – 337.20(PAR_EE3) | 0.77 | 9.65 | 389.11 | 396.06 | 4.01 |
Age = 100.84 – 171.53(PAR1) – 344.89(PAR_EE3) | 0.76 | 9.70 | 389.63 | 396.58 | 4.52 | ||
Age = 103.68 – 208.15(PAR_EE1) – 373.11(PAR3) | 0.74 | 10.09 | 393.77 | 400.73 | 8.71 | ||
Age = 102.88 – 218.35(PAR1) – 385.67(PAR3) | 0.74 | 10.20 | 394.88 | 401.84 | 9.89 | ||
LR + MR | Age = 119.39 – 464.89(PAR_EE1) – 73.82(PAR_EE2) | 0.65 | 11.76 | 409.65 | 416.61 | 28.25 | |
Age = 119.48 – 470.92(PAR_EE1) – 77.95(PAR2) | 0.65 | 11.78 | 409.82 | 416.77 | 28.48 | ||
Age = 119.84 – 524.76(PAR1) – 77.17(PAR_EE2) | 0.64 | 11.88 | 410.68 | 417.64 | 29.73 | ||
Age = 119.84 – 534.30(PAR1) – 78.91(PAR2) | 0.64 | 11.92 | 411.03 | 417.98 | 30.24 | ||
Table 5
Accuracy and other performance parameters associated with each age estimation model selected during the best-subsets regression analysis. Models are listed in the same order as they appear in table 4
k | Image type | Model | R2 | r | f2 | Power | Absolute residual parameters (years) | 95% PI (years) | ||
MAE | SD | MAE 95% CI | ||||||||
1 | STSI | Age = f(PAR_EE3) | 0.76 | -0.87 | 3.17 | 1.00 | 7.21 | 6.60 | (5.37, 9.05) | ± 19.85 |
Age = f(PAR3) | 0.73 | -0.85 | 2.70 | 1.00 | 7.47 | 7.27 | (5.45, 9.50) | ± 21.15 | ||
LR | Age = f(PAR_EE1) | 0.65 | -0.81 | 1.86 | 1.00 | 9.52 | 6.87 | (7.61, 11.44) | ± 23.89 | |
Age = f(PAR1) | 0.64 | -0.80 | 1.78 | 1.00 | 9.76 | 6.76 | (7.88, 11.64) | ± 24.17 | ||
MR | Age = f(PAR_EE2) | 0.44 | -0.66 | 0.79 | 1.00 | 11.87 | 9.05 | (9.35, 14.39) | ± 30.36 | |
Age = f(PAR2) | 0.42 | -0.65 | 0.72 | 1.00 | 11.93 | 9.33 | (9.33, 14.53) | ± 30.79 | ||
2 | STSI + LR | Age = f(PAR_EE1, PAR_EE3) | 0.77 | -0.88 | 3.35 | 1.00 | 6.97 | 6.45 | (5.18, 8.77) | ± 19.40 |
Age = f(PAR1, PAR_EE3) | 0.77 | -0.88 | 3.35 | 1.00 | 7.03 | 6.45 | (5.23, 8.83) | ± 19.49 | ||
Age = f(PAR_EE1, PAR3) | 0.75 | -0.87 | 3.00 | 1.00 | 7.40 | 6.62 | (5.56, 9.24) | ± 20.29 | ||
Age = f(PAR1, PAR3) | 0.75 | -0.86 | 3.00 | 1.00 | 7.47 | 6.70 | (5.60, 9.33) | ± 20.50 | ||
LR + MR | Age = f(PAR_EE1, PAR_EE2) | 0.67 | -0.82 | 2.03 | 1.00 | 9.13 | 7.06 | (7.17, 11.10) | ± 23.63 | |
Age = f(PAR_EE1, PAR2) | 0.66 | -0.81 | 1.94 | 1.00 | 9.14 | 7.07 | (7.17, 11.11) | ± 23.67 | ||
Age = f(PAR1, PAR_EE2) | 0.66 | -0.81 | 1.94 | 1.00 | 9.16 | 7.21 | (7.15, 11.17) | ± 23.87 | ||
Age = f(PAR1, PAR2) | 0.66 | -0.81 | 1.94 | 1.00 | 9.20 | 7.22 | (7.19, 11.21) | ± 23.95 | ||
Cross-validation
The cross-validation results are presented in Fig. 3, with detailed results provided in Online Resource 1. Minor differences in MAE and SEE values (below one year) were observed between in-sample and cross-validation assessments, indicating a relatively small prediction bias across all models. As before, models derived from stereomicroscopic tooth section images showed the best MAECV (7.45–7.72 years), SEECV (10.17–10.84 years) and R2CV (0.71–0.74) values. Models derived from labiolingual radiographs followed (MAECV = 9.89–10.13 years; SEECV = 12.32–12.45 years; R2CV = 0.62–0.63), while those derived from mesiodistal radiographs performed least effectively (MAECV = 12.31–12.37 years; SEECV = 15.64–15.85 years; R2CV = 0.38–0.40). Exclusion of enamel area consistently improved model performance across all image types.
Fig. 3
Box plots comparing the cross-validation absolute residual distributions associated with the simple linear regression age estimation models. Shaded box plots represent models excluding enamel. The cross-validation mean absolute error (MAECV) and standard error of the estimate (SEECV) (measured in years) associated with each model are shown as red circles and triangles for comparison. Each model’s cross-validation coefficient of determination (R2CV) is also indicated on the plots. Brackets with * indicate a significant difference (p ≤ 0.05) in MAECV between image type groups as determined by pairwise comparisons. PAR1 = pulp/tooth area ratio obtained from labiolingual radiograph; PAR_EE1 = pulp/tooth area ratio (excluding enamel) obtained from labiolingual radiograph; PAR2 = pulp/tooth area ratio obtained from mesiodistal radiograph; PAR_EE2 = pulp/tooth area ratio (excluding enamel) obtained from mesiodistal radiograph; PAR3 = pulp/tooth area ratio obtained from stereomicroscopic tooth section image; PAR_EE3 = pulp/tooth area ratio (excluding enamel) obtained from stereomicroscopic tooth section image
Only two of the selected multiple regression models, derived from stereomicroscopic tooth section images and labiolingual radiographs, demonstrated similar or improved MAECV, SEECV and R2CV parameters compared to their simpler subset models (see Online Resource 1). However, improvements were marginal.
A complementary two-way repeated measures ANOVA evaluated the main and interaction effects of image type and enamel area on absolute estimate residuals (Table 6). Sphericity was observed for both image type and the interaction term (Mauchly’s test, p > 0.05). There was no significant interaction effect between image type and enamel area on absolute residuals (F = 0.44, df = 2, p = 0.643). A significant main effect was observed for image type (F = 7.48, df = 2, p < 0.001) but not for enamel area (F = 1.50, df = 1, p = 0.227). Pairwise comparisons (Table 7) showed that models derived from stereomicroscopic tooth section images had significantly lower mean absolute residuals compared to those derived from labiolingual (p = 0.026, 95% CI = [0.31, 4.55]) and mesiodistal radiographs (p = 0.001, 95% CI = [1.94, 7.57]) (Fig. 3). Models derived from labiolingual radiographs exhibited notably lower mean absolute residuals than those from mesiodistal radiographs, although this difference did not reach significance (p = 0.060, 95% CI = [-0.10, 4.76]).
Table 6
Tests of within-subjects effects results (sphericity assumed) for the two-way repeated measures analysis of variance (with two within-subjects factors)
Source | df | SS | MS | F-statistic | p |
|---|---|---|---|---|---|
Image type | 2 | 1176.13 | 588.06 | 7.48 | < 0.001* |
Error (image type) | 102 | 8024.14 | 87.01 | - | - |
Enamel area | 1 | 2.86 | 2.86 | 1.50 | 0.227 |
Error (enamel area) | 51 | 97.16 | 1.91 | - | - |
Interaction | 2 | 0.67 | 0.34 | 0.44 | 0.643 |
Error (interaction) | 102 | 77.18 | 0.79 | - | - |
Table 7
Pairwise comparisons (using Fisher’s least significant difference procedure) results following the two-way repeated measures analysis of variance (with two within-subjects factors)
(I) Image type | (J) Image type | Mean difference (I – J) | SE | p | 95% CI for difference |
|---|---|---|---|---|---|
LR | MR | -2.33 | 1.21 | 0.060 | (-4.76, 0.10) |
STSI | 2.43 | 1.06 | 0.026* | (0.31, 4.55) | |
MR | LR | 2.33 | 1.21 | 0.060 | (-0.10, 4.76) |
STSI | 4.76 | 1.40 | 0.001* | (1.94, 7.57) | |
STSI | LR | -2.43 | 1.06 | 0.026* | (-4.55, -0.31) |
MR | -4.76 | 1.40 | 0.001* | (-7.57, -1.94) |
Discussion
This study evaluated the PAR method’s applicability for predicting adult age-at-death using maxillary canines from a South African cadaveric sample. We assessed the method’s reliability, sex bias, and accuracy across different image types, with and without consideration of enamel.
The PAR method demonstrated excellent intra- and inter-observer reliability. Its simplicity and reliance on direct observation quantification and computer-assisted image analysis reduces observer variability, distinguishing it from conventional, more subjective skeletal ageing techniques [6].
Our findings indicated non-significant sex differences in PARs, suggesting the potential for developing unified age estimation models across sexes. This aligns with prior research [14], supporting the method’s applicability for cases where skeletal sex estimation is challenging/impossible [10].
PAR models derived from labiolingual stereomicroscopic tooth section images were most effective for age estimation and produced significantly lower absolute estimate residuals/errors than models derived from radiographs. Aligned with D’Ortenzio et al. [23], this highlights the superior image resolution/sharpness and direct observation capabilities of stereomicroscopic images. Conversely, radiograph resolution and contrast were limited by indirect observations (via X-radiation) and other potential factors such as the machine focal spot size and pixel/crystal size of the receptor [39]. Additionally, two-dimensional radiographs are susceptible to dimensional distortion [31]. The non-linear morphology of maxillary canines introduces an angular relationship between their surfaces and the receptor, likely exacerbating image distortion [31].
Couoh and Bautista [24] and Keerthi Priyadharshini [40] observed non-significant differences between age estimates from radiographs and tooth section images. They suggested that incomplete dental sections resulting from root curvature (asymmetry) may have affected the accuracy of their results. We addressed this limitation by developing a modified stereomicroscopic technique focused on mesiodistal sectioning and measurements confined to the coronal half of each tooth section (discussed in the methods section). This adjusted approach allowed for superior age estimation accuracy from maxillary canine tooth section images. One limitation is that older adults or those with premature dentine apposition may exhibit pulp cavities that have receded (apically) beyond the coronal half of the tooth. However, this was only observed for the oldest individual in our sample (91 years), who was assigned a PAR value of zero. This suggests that the method can still discriminate age-at-death well into old age, likely due to relatively slow rates of secondary dentine deposition in canines [41, 42].
PAR models derived from labiolingual radiographs performed better than those based on mesiodistal images. Two factors may explain these results. Firstly, maxillary canine roots typically angle distally relative to the vertical crown axis [43], likely causing greater angular distortion during the paralleling procedure for mesiodistal radiographs [31]. Secondly, reductions in pulp cavity size primarily occur in a mesial-distal direction [44], visible in labiolingual views. In contrast, reductions in labial-lingual/vestibular-oral directions (observed in mesiodistal views) occur later in life (around 60–70 years) due to fibrous dentine formation [44], thereby weakening the correlation between age-at-death and PARs from mesiodistal radiographs.
While enamel area did not significantly affect estimate residuals, its exclusion consistently improved model performance across all image types, supporting similar studies on pulp/tooth volume ratios [20, 21]. Dental wear is a multifaceted and variable process [45, 46], likely introducing errors for age analyses dependent on tooth size. Our findings may suggest that excluding enamel area minimises this variability, enhancing accuracy in age estimation models. However, further research is required to assess the complex effect of tooth wear on PAR age estimates [24].
Overall, this research suggests that the PAR method is a reliable, sex-independent technique, providing relatively accurate age-at-death estimates for South African adults. It demonstrates superior accuracy compared to other dental ageing methods validated in this population, such as those developed by Gustafson [47] and Lamendin et al. [48], which have shown complete-sample MAEs ranging from 11.6 to 15.10 years [26, 27]. Age estimation of decedents in South Africa is problematic as most available methods were created using populations from the global north and are known to be relatively inaccurate locally [49‐51]. South Africa is considered an epicentre for understanding human origins and has an exceptionally high rate of unidentified medico-legal decedents [9, 52‐54]; therefore, local forensic and bioarchaeological applications for this method are extensive.
Our findings underscore the importance of image type and enamel area considerations for optimising method accuracy. We recommend excluding enamel area from PAR calculations to minimise errors from dental wear and enhance formula applicability across diverse population groups. Both labiolingual stereomicroscopic tooth section images and labiolingual radiographs proved effective for PAR-based age estimation, producing error values (MAE and/or SEE) of approximately 10 years or less – a recommended benchmark for age estimation methods [11, 15].
While relatively accurate, the PAR method produced broad 95% prediction intervals (± 20 years at best), a common limitation in adult age estimation techniques [6]. This imprecision reflects the complex relationship between biological and chronological age, shaped by individual and population differences in genetics, environment, health and lifestyle [52, 55]. As individuals age, variability in degenerative skeletal indicators increases, making precise predictions challenging for older adults and limiting the utility of biological profiles in forensic investigations. To improve precision, the PAR method should be used alongside other skeletal techniques and holistic analyses that incorporate a wider range of age-informative traits. Future research should also explore whether this method is more precise in younger adults and if separating analyses into life phases could enhance its utility.
While the stereomicroscopic approach delivered superior results, its invasive and time-consuming nature may limit practical application. Destructive sampling of human remains in forensic and archaeological sciences poses ethical and practical challenges, particularly regarding skeletal preservation [56‐59]. This can impact future research opportunities, curation, the integrity of medicolegal evidence and the cultural/religious concerns of affected communities [56‐59]. Practitioners should weigh the value of information obtainable from such techniques against feasibility on a case-by-case basis, considering factors like finances, resource availability, timeframe, procedural viability, and the ability to record and preserve other skeletal information prior to sampling [57, 59]. It is also important to explore alternative, minimally destructive methods that can provide comparable data.
Validating our sample-specific formulae across other South African cohorts is required to assess their generalisability. Multiple regression models were explored but marginally improved estimates and involved multiple procedures; hence, their application is not recommended. The sample age distribution was skewed towards older adults, which may diminish the predictive power of regression models for younger age groups. Future studies should increase the reference sample size to enhance variation and ensure balanced representation across age groups, thereby reducing prediction biases and optimising regression model performance [60]. Evaluation of the method, including the modified stereomicroscopic approach, across other tooth types and population groups is encouraged. The development of sample-specific formulae for non-South African population groups is recommended to account for population differences in skeletal ageing [25].
Acknowledgements
A special note of thanks to all the donors, without whom, this research would not be possible. Thank you to the staff in the Division of Clinical Anatomy and Biological Anthropology, Department of Human Biology, UCT, who assisted during the data collection process: Mr Michael Cassar, Ms Shirees Benjamin, Mr Jacques Jacobs and Mr Wilber Van Niekerk. Thank you to Mr Adriaan Geldenhuys and Mr Abduraghmaan Fisher for providing training on dental extraction. The Division of Clinical Anatomy (Faculty of Medicine and Health Sciences, Stellenbosch University) is acknowledged for use of cadaveric material for the study. Specifically, the following staff members are gratefully acknowledged for assistance with permission application and/or data collection: A/Prof. Nanette Briers, Mrs Jodie Lemphane, Ms Ayesha Jeffery, Mr Paul Pretorius, Mr Logan Williams, Mr Jonathan Goliath, Mrs Joyce Cloete, Mr Anthony Marthie and Mr Warren Uithaler. We acknowledge the Department of Oral and Maxillofacial Radiology (Faculty of Dentistry, University of the Western Cape) for assisting us with their expertise and resources. Thank you to Dr Emile Prince and Dr Tineke van Zyl for permitting us to utilise the department’s radiology equipment. A special thank you to Mrs Moreza Samuels and Mrs Candice Festus for assisting with the exposure and confirmation of diagnostically acceptable radiographic images. We would also like to thank Dr Charlene Saayman at the faculty for her guidance and assistance during the initial stages of the project.
Declarations
Conflict of interest
The authors declare no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Ethics approval
This study was performed in line with the principles of the Declaration of Helsinki. Ethical clearance was granted by the Human Research Ethics Committee at the University of Cape Town (UCT) (HREC REF: 336/2022). Permission was obtained from UCT’s Cadaver Research Governance Committee (CRGC 2022/001), as well as the Division of Clinical Anatomy, Faculty of Medicine and Health Sciences, Stellenbosch University (SU), and the Provincial Inspector of Anatomy.
Consent to participate
This study utilised cadaveric human bodies donated by individuals who consented to their use for teaching and research purposes.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
