Optimal age classification of young individuals based on dental evidence in civil and criminal proceedings

Criminal cases involving young people, irregular immigration and the many issues related to asylum seekers has increased the judicial demand of age estimation. Calcification of teeth and specifically of third molar has demonstrated to be reliable evidence to estimate age respect to 18 years threshold of age. As prosecution of research of Pinchi et al. (2010) and Corradi et al. (J Forensic Sci 58:51-59, 2013), the study aims to evaluate if tuning the size of the zone of indifference posed around the age threshold improves the performances of the age classification model. The sample was composed of 1,560 OPGs of Italian subjects aged between 15 and 22 years. Third molar calcification stage was assessed according to Demirjian's scale by three different experts. Intra- and inter-operator variability has been calculated. The statistical analysis was provided by a Modified Na < ve Bayesian allowing the use of soft evidence. Rate of in/correct classification was provided for individuals classified at a very high level of probability (90 %), as needed for criminal cases, and for a lower probability level (51 %) as it suffices for civil cases. The intra-observer reproducibility varies between 79.2 and 89.2 % with soft evidence, whilst it decreases from values between 0.589 and 0.763, when only hard evidence is allowed to experts showing the usefulness of the MBN approach. In civil cases, imposing the constraint of classifying at least 95 % of the individuals, the method achieved a rate of correct classification in the range 80-83 % depending on the expert. In criminal cases, we tuned the ZOI size to achieve 85 % of individuals correctly classified and the model succeeded in classifying 66-81 % of the sample, the variability still being dependent on the expert's ability. After a review of several studies concerning the age classification of young individuals by using dental evidence, we must conclude that it is almost impossible to make a comparison among them. To rank the effectiveness of different methods is to challenge them with the same problem and data, looking at the results measured by the same accepted scoring rule. It could also be interesting to repeat the experiment in different conditions varying the reference population and considering if some important covariates, like sex and health status, influence the model performances.