Trustworthy AI-based legal age estimation using orthopantomographs

In forensic anthropology, legal age estimation means estimating the age of a subject in order to determine if it is above or below a specific legal threshold (typically 18 years old). It is of great importance in many different scenarios such as migration of undocumented minors, child brides, exploitation or trafficking. The main goal of this work is to develop a legal age estimation method, based on artificial intelligence, that can be applied in day-to-day forensic practice. For this purpose, we use a sample of 10,739 orthopantomographs of individuals ranging from 14–26 years, across twelve countries and four continents. Our best model fine-tuned ResNeXt50 (adapted to regression) and obtained a mean absolute error of 1.12 years in testing. When thresholding the regression estimates at 18 years old, it achieved a classification accuracy of 88.38 %. In order to show the validity of the model in real forensic scenarios, we evaluated it with four samples from population origins not seen during training. Over these samples, we obtained mean absolute errors between 1.21 and 1.47 years, and accuracies between 83 and 92 % when setting the 18-year threshold. Moreover, we obtained prediction intervals of different coverage levels to address the ethical problem of overestimating the age of a minor. This allows the model to estimate the minimum of an interval with a certain coverage level, where the higher the coverage the fewer minors are estimated as adults. As an example, if estimating the minimum of a prediction interval of 95 % coverage, 98.2 % of minors in the test set are classified as such. Our robustness to geographical origin, state-of-the-art accuracy and the estimation of prediction intervals, which are the main contributions of this work, make the proposed method, to the best of our knowledge, the only one that has been proven to be applicable in real forensic scenarios. As a result, our method properly deals with the following principles of Ethics Guidelines for Trustworthy Artificial Intelligence: Human agency and oversight; Technical robustness and safety; Transparency; Diversity, non-discrimination and fairness; and Societal and environmental well-being. Furthermore, it provides both the most probable age and a prediction interval, which in turn allows the estimation of a minimum age with fixed statistical support. As a result, the developed method can be used in conjunction with other methods according to the European protocols of legal age estimation. • With a test MAE of 1.12 years, our method performs on par with human experts. • It yields the most probable age and a prediction interval for minimum age estimation. • This offers strong statistical support to avoid misclassifying minors as adults. • The method remains accurate regardless of the individual’s geographic origin. • Trustworthiness is supported by activation maps and three-continent external validation.