Introduction
This article presents a formal correction to the study titled "With or without human interference for precise age estimation based on machine learning?", which was published in the International Journal of Legal Medicine in 2022. Following the discovery of calculation errors in the original analysis, the authors have provided updated performance metrics for their automated dental age estimation (ADAE) model. This correction is vital for the scientific community, as it ensures that forensic and clinical practitioners have an accurate understanding of the current capabilities and limitations of fully automated AI systems in biological age assessment.
Problem Statement and Objectives
Age estimation is a fundamental task in forensic science, essential for identifying human remains, determining the legal age of individuals without documentation, and assisting in clinical dentistry. The original research sought to investigate whether fully automated machine learning models—those operating without human intervention—could match or surpass the accuracy of manual or semi-automated methods.
However, the authors identified significant "calculation mistakes" in the final five columns of their primary data table (Table 4) and the associated visualization (Figure 3). These errors led to an overestimation of the model's accuracy. This correction recalibrates our understanding of "human-out-of-the-loop" AI performance, providing a more realistic benchmark for future research in the field.
Methodology and Dataset
The study utilized a comprehensive dataset consisting of 1,017 individuals, categorized into age groups spanning from 5.00 to 24.99 years. This broad range is crucial for evaluating model performance across various stages of dental development, from the eruption of permanent teeth in childhood to the maturation of third molars in early adulthood.
While the underlying machine learning architecture remains the same as in the original publication, this document provides the corrected statistical outputs for the Automated Dental Age Estimation (ADAE) model. The authors re-calculated several key performance indicators:
- Mean Absolute Error (MAE): This metric represents the average magnitude of the errors in age prediction, providing a clear measure of overall accuracy.
- Mean Age Difference: This indicates the systematic bias of the model, revealing whether it tends to overestimate or underestimate chronological age.
- 95% Confidence Intervals and p-values: These statistical measures are used to determine the significance of the differences between the actual chronological age and the predicted dental age.
Key Results and Performance Recalibration
The corrected analysis provides a more grounded and less optimistic view of the ADAE model’s capabilities. One of the most significant shifts in this correction is the realization that the ADAE model systematically underestimates chronological age across all age groups from 5 to 24 years. The original paper had incorrectly suggested that the model overestimated age in the 5–6 and 15–21 year brackets.
The recalibrated results include several key adjustments:
- Overall MAE Increase: The overall MAE for the ADAE model increased from the initially reported 0.83 years to 1.03 years.
- Systematic Bias: The overall mean difference was adjusted from -0.10 years to -0.59 years, indicating a much stronger trend toward underestimation (predicting individuals to be younger than they actually are).
- Accuracy Thresholds: The number of age groups achieving an MAE of less than 1.00 year dropped from 14 groups to 11 groups.
- Statistical Significance: Significant differences between predicted dental age and actual chronological age are now noted in age groups 5–14 and 20–24, a different distribution than previously reported.
As illustrated in the corrected Figure 3, the "performance gap" between the fully automated ADAE model and the semi-automated (MDAE) or manual (ADSE) methods has narrowed. This suggests that while automation offers efficiency, it has not yet achieved the significant leap in accuracy that was previously claimed.
Discussion and Future Work
The correction highlights a common and persistent challenge in forensic AI: maintaining high precision across the entire human developmental arc. The fact that the model consistently underestimates age suggests a potential bias in the training data or the feature extraction process, perhaps favoring dental markers that are more characteristic of younger individuals.
Error rates were not uniform across the dataset. The model performed best in younger cohorts but struggled as individuals reached their mid-20s, with the MAE peaking at 2.05 years for the 24-year-old group. Future work in this area will likely need to address this systematic underestimation. Potential solutions include implementing age-specific weighting in the model's loss function or refining how the system interprets the final stages of third-molar development. This correction ensures that the field of automated age estimation continues to advance on a foundation of accurate and transparent performance data.
Summary updated