AI Reads Four Molars From CBCT Scans to Estimate Adult Age Within 2.5 Years

Forensic scientists in China have built an artificial intelligence framework that estimates a person’s age from the pulp chambers of their four first molars, and their best model lands within about two and a half years of the truth. The study, published in the Egyptian Journal of Forensic Sciences, analyzed cone-beam computed tomography (CBCT) scans from 2,000 Sichuan Han Chinese adults aged 18 to 65 and systematically pitted eleven traditional regression models, six machine learning algorithms, and a custom one-dimensional convolutional neural network against one another. The winner was an Extreme Gradient Boosting (XGBoost) model that treated each tooth individually rather than averaging them, achieving a mean absolute error of 2.51 years and a coefficient of determination of 0.907 on an independent test set.

The biological signal underlying the approach is the pulp-dentin index, or PDI, a ratio that captures how the pulp chamber shrinks over a lifetime as secondary dentin steadily deposits on its walls. On CBCT images, researchers measured the height and buccolingual width of each pulp chamber along with the surrounding dentin thicknesses, then combined these linear distances into a single percentage for each tooth. Because the index relies on linear measurements rather than laborious three-dimensional volume segmentation, it offers a streamlined surrogate for the volumetric pulp reduction that forensic odontologists have long tracked with two-dimensional pulp-to-tooth area ratios.

The correlation with chronological age was strong and negative across all four first molars, with Pearson coefficients ranging from −0.798 to −0.866 in the overall sample. The maxillary left first molar, tooth 26 in the FDI numbering system, showed the strongest association at r = −0.866, a pattern that held in both sexes and reached r = −0.881 in females. The team also found systematic anatomical and sex differences: maxillary molars had significantly lower PDI values than mandibular ones, and males showed significantly lower values than age-matched females for all four teeth, findings the authors link to functional masticatory load and potentially sex-related differences in dentin deposition.

Measurement reliability was assessed on a random subset of 200 cases, with two trained researchers working independently. The intraclass correlation coefficient for intra-observer agreement was 0.938, and for inter-observer agreement it was 0.875, both exceeding the 0.75 threshold the authors cite as indicating high repeatability. The scans themselves came from a Planmeca ProMax 3D Max device with a standardized 8 × 8 cm field of view, 200 μm voxels, and fixed exposure parameters, and were originally acquired for clinical indications such as implant planning, orthodontics, and impacted tooth evaluation rather than for age estimation.

That retrospective, clinically driven origin matters, because the authors acknowledge it may introduce selection bias and limit how well the sample represents the general population. After screening 2,680 cases against strict criteria—first molars free of caries, severe attrition, restorations, and pulp calcification, and no history of disease, trauma, or medication affecting tooth metabolism—the final cohort of 2,000 subjects spanned five age decades with balanced sex representation. The data were partitioned into training, validation, and test sets at an 8:1:1 ratio, with the test set held out entirely from model training and hyperparameter selection.

The modeling comparison produced a clear hierarchy. Traditional approaches came first in the lineage: a stepwise regression combining all four tooth positions reached an R² of 0.804 with a mean absolute error of 4.97 years, while the best cubic curve regression, built on the maxillary left first molar, achieved an R² of 0.792 and an error of 4.88 years. When the six machine learning algorithms were trained on a single averaged PDI value, Random Forest led with an error of 4.53 years—only modestly better than regression. The decisive change came from feature engineering: when the four tooth-specific PDI values replaced the average, XGBoost’s error dropped from 5.16 to 2.51 years, a 51.4% reduction, and Random Forest improved to 2.72 years, a 40.0% reduction.

Paired t-tests on the test set confirmed the differences were statistically significant, with the tooth-specific XGBoost model showing lower absolute prediction errors than Random Forest by a mean of 0.21 years (95% CI 0.09–0.33) and than Gradient Boosting by 1.12 years (95% CI 0.95–1.29). Sex-stratified analyses told a subtler story: XGBoost was the top performer in the male subsample with an error of 3.53 years after tooth-specific modeling, while Random Forest retained the lead in the female subsample at 3.59 years. The authors suggest this algorithmic divergence may reflect sex-based differences in the pattern of pulp chamber regression, though they frame it as an observation requiring further study rather than an established mechanism.

The deep learning arm of the study took a deliberately lightweight approach. Instead of feeding raw CBCT images into a large network, the one-dimensional CNN received only a length-four sequence of PDI values, one per first molar, and learned local correlations among adjacent teeth through convolutional kernels, average pooling, and fully connected layers. Trained for 500 epochs with the Adam optimizer, cosine annealing of the learning rate, and early stopping, it converged to a mean absolute error of 4.24 years on the test set. That figure beat the best traditional regression model and the best machine learning model without tooth-specific features, but fell short of the feature-engineered XGBoost. The authors attribute this gap to the very low input dimensionality, the moderate sample size relative to deep learning’s data appetite, and the loss of spatial information inherent in using derived measurements rather than raw images.

The error profile of the CNN also varied with age, decreasing for individuals under 20, remaining relatively stable between 20 and 35, and gradually increasing above 35—a pattern the authors connect to the slower and more variable pace of secondary dentin deposition in later adulthood. This age-related degradation is one of several limitations the team flags. The model was developed and validated exclusively on Sichuan Han Chinese individuals, and dental aging patterns vary across ethnic, geographic, and dietary populations, so applicability elsewhere remains untested. The current pipeline also depends on manual landmark placement, and the authors call for full automation and explainable AI methods such as Grad-CAM to strengthen transparency and forensic credibility.

Context matters when judging the headline accuracy figure. Methods based on tooth development, such as Demirjian staging, routinely achieve errors under one year in children and adolescents because tooth formation follows a rapid, programmed schedule; adult age estimation, by contrast, relies on regressive changes that accumulate slowly and unevenly, so errors around 2.5 years represent strong performance for adults. The authors position their framework as a pragmatic middle ground between traditional morphometrics and full-image deep learning, and they stress that external validation on diverse, multi-center datasets is an urgent priority before any clinical or forensic application can be recommended. Within those boundaries, the study’s central message stands: the age-related information distributed across four first molars is greater than the sum of its averaged parts, and machine learning models that respect tooth-by-tooth anatomy can extract it with notable precision.

Subject of Research: AI-assisted forensic adult age estimation from CBCT-derived pulp-dentin index values of first molars

Article Title: AI-assisted age estimation from first molars: an integrated evaluation of regression, machine learning, and 1D-CNN models using CBCT-derived pulp-dentin index

Article References: Ding, Y., He, Y., Chen, R., Zhong, T., Tao, W., Fu, Y., Wang, W., Zhang, S., Zhang, X., & Jin, B. (2026). AI-assisted age estimation from first molars: an integrated evaluation of regression, machine learning, and 1D-CNN models using CBCT-derived pulp-dentin index. Egyptian Journal of Forensic Sciences, 16(1), Article 80. https://doi.org/10.1186/s41935-026-00581-2

Image Credits: AI Generated

DOI: 10.1186/s41935-026-00581-2

Keywords: forensic age estimation, CBCT, pulp-dentin index, first molars, XGBoost, machine learning, 1D-CNN, secondary dentin, forensic odontology, Sichuan Han Chinese, regression models, deep learning