Article begins

In 2023, we witnessed a tragic surge in migrant deaths along migratory routes worldwide, marking it as one of the deadliest years on record according to the statistics from the Missing Migrant Project, with an estimated more than 8,000 cases. However, this figure probably underestimates the true magnitude of the problem, since many disappearances are not reported in official databases.

The movement of people across borders is driven by a complex web of political, social, and cultural factors, making migration pathways a pressing challenge that affects both Global North and South countries. Examples include the dangerous Mediterranean route to Europe and the Darien Gap on the border between Panama and Colombia. The causes of these deaths and disappearances vary, from accidents to drowning and lack of access to basic needs such as food and medical care throughout the journey. These reasons reveal that behind every statistic is a human story, highlighting the harsh realities of a system sustained by structural inequalities.

As we face the immense challenges surrounding the deaths of undocumented migrants, introspection becomes crucial. In forensic anthropology (FA), the human identification process is a pivotal and exigent task for forensic personnel, as this process is essential for providing closure to families, aiding in legal proceedings, and ensuring justice. Of the four parameters that compose the biological profile—age, sex, stature, and affinity population—the first of them poses a particularly intricate challenge, especially in adult individuals, giving rise to various questions and unresolved technical and theoretical problems. 

This problem highlights the need for interdisciplinary methodologies and technologies that fuse technical competence with sociocultural insight. In this crossroads where the data and statistics intersect with the social reality, we have a challenge to translate all these insights into tangible outcomes that return a human history and identity. This opens a lot of technical and theoretical questions that force us to ask ourselves: How might the social sciences, particularly FA, harness state-of-the-art artificial intelligence (AI) and machine learning (ML) to transform these statistics into narratives of human identity? And most specifically, how do we ensure that the ML-based models developed for the human identification process can offer effective, accurate, and equitable responses, especially to the families affected by these tragedies?

Based on these considerations, we need to explore the main problems of age-at-death estimation, not only to develop ML-based models, but also to understand how these can be applicable to complex forensic scenarios.

Firstly, to estimate age in adult individuals, forensic anthropologists evaluate a series of degenerative changes observed in specific areas of the human skeleton, known as osseous indicators of age. Traditional methods used by forensic anthropologists have been developed from analyzing these degenerative changes in the bones that compose the pelvic girdle, specifically the hip bones. These bones exhibit several traits that change throughout life, providing valuable indicators of an  individual’s age and are widely used in forensic anthropology. These methods, developed during the 1990s and early 2000s, are known as traditional or standard methods for age estimation.

Although these traditional methods are widely used, they are often characterized by a lack of precision, and many of them do not meet the 95% accuracy requirement for human identification process standards. Furthermore, the tedious application of these methods can hinder the speed of results, which is essential in humanitarian contexts such as migration crises. Obviously, these problems make it difficult to accurately estimate age and, consequently, can hinder the human identification process, as well as making decisions about the admissibility of evidence in medico-legal instances. However, these problems are only part of all the difficulties facing age estimation.

A second aspect of complexities is related to two critical factors: the reference samples and validation studies, two aspects that are not always mentioned or explained in detail. Firstly, the traditional methods used for age estimation are principally developed from osteological samples derived from the United States or European countries. For instance, in the next map we can observe the provenance of the datasets for developing some of the most common traditional methods for age estimation.

Credit: Noemí Aedo-Noa
Geographic origins of osteological collections used for the development of some of the traditional age estimation methods such as the pubic symphysis, auricular surface of the ilium, and acetabulum.
Geographic origins of osteological collections used for the development of some of the traditional age estimation methods such as the pubic symphysis, auricular surface of the ilium, and acetabulum.

If we look at the map, a large part of the planet lacks representative data. This is particularly problematic, because not only is there a bias toward data from the Global North, but, for instance,  data derived from osteological collections in South American countries are particularly underrepresented in statistical terms. Consequently, this leads to inherent biases and, although it is always expected to find a certain level of bias in age estimation, the lack of representation can limit the applicability of the methods across other regions.

Secondly, validation studies in FA are pivotal for ensuring the effective application of a method in different populations. Unfortunately, the traditional methods are often validated on osteological samples from the Global North, or generalized as a global standard, without a comprehensive analysis of their adaptability in non-Western populations and, although there are currently important efforts to validate age estimation methods in populations from the Global South, these remain scarce. Such built-in biases not only severely compromise the validity of the results, but result in statistical bias, which in FA is called “age mimicry,” where the estimate of age in a target sample tends to reflect the distributions of the reference sample from which the method was developed.

Clearly, age estimation is a challenge for researchers and forensic personnel. This task demands all their ingenuity, particularly in demanding contexts such as disaster victim identification (DVI).

Today, many fields of investigation can benefit from the potential of AI and ML systems. In simple terms, we can define AI as the simulation of human intelligence processes by machines, especially computer systems. ML is a subfield of AI that uses advanced computational algorithms to analyze large data sets and extract information and significant patterns. 

In FA, the use of ML-based models can guide us in the analysis of large data sets to recognize patterns of interest, automate methods so that their application is faster, improve the effectiveness and precision of traditional methods, compare the efficiency between algorithms and select the most accurate for a specific taskevaluate the validity of new methodologies, or generate robust validation studies, among others. Nevertheless, the use of these tools and approaches, in a field such as FA,  requires more than just the application of algorithms for a specific task.

The challenges associated with age estimation show us that not everything comes down to accuracy. It is also crucial to consider aspects such as reproducibility and reliability. Anthropologists must ask themselves: How can we guarantee these cutting-edge tools are applicable and effective in the human identification process, especially in challenging context such as the identification of the undocumented migrants? And, how can we avoid providing methodological approaches that, while they may be very accurate in some contexts, may also continue to exacerbate statistical bias?

In 1979, statistician George E. P. Box used the phrase “all models are wrong, but some are useful” to explain that all statistical models are wrong, because they are a simplification of reality, and the important question should be: is the model revealing and useful? If we focus this sentence on the use of ML-based models in FA, researchers should ask themselves, how can we develop ML-based methodologies that are effective in humanitarian forensic contexts? And, how can we ensure that we do not fall into the so-called black box models that, although they tend to have great precision, fail to be applicable and understandable to forensic personnel or search teams, nor do they meet the admissibility criteria in the courts? In such cases, the need has been raised to work from four criteria that allow creating a glass box, which means that systems are explainable, interpretable, transparent and easily generalizable to all stakeholders involved in the human identification process.

Credit: Noemí Aedo-Noa
Four key considerations for developing effective ML-based models to ensure the admissibility of scientific evidence in real forensic scenarios.
Four key considerations for developing effective ML-based models to ensure the admissibility of scientific evidence in real forensic scenarios.

Glass box models allow FA experts to understand how a given age estimate or any other biological profile parameter is arrived at; that is, they provide explainability and interpretability. This is essential to verify the validity of the predictions and provide clear explanations both in court and at the forensic investigation. On the other hand, transparency in the decision-making process is crucial to generating confidence in the results. In this sense, glass box models allow the reasoning behind each prediction to be followed step by step, facilitating the identification of possible errors or biases in the analysis.

By adopting these criteria, FA promotes more ethical standards within Humanitarian Forensic Action, since peer monitoring and evaluation is facilitated, ensuring that the estimates are rigorous and all interested parties can clearly understand the limitations or biases. In turn, these criteria improve the generalization of the models, allowing their effective application in various forensic scenarios. This aspect is particularly significant in FA, where traditional age estimation methods often lack the ability to accurately adapt to diverse geographic regions. 

However, in addition to these technical considerations, there is an essential part that transcends the simple fact of using these benefits of these cutting-edge tools. We must recognize and understand that forensic anthropologists are primarily social scientists tasked with addressing complex issues intertwined with social, cultural, and political aspects. While technical expertise is crucial, humanitarian forensic efforts transcend mere technical competence. 

The complexity of the human identification process requires understanding how to effectively use analytical approaches to provide meaningful answers to pressing questions that are not limited solely to a certain geographic region. In the case of age estimation, it is essential to understand that the best method is not always the one with the lowest bias rate, but rather the one that demonstrates a strongly reliability and reproducibility, for instance, through validation based on studies from various geographic regions, since data have quality when they reflect the variability of the process that generates them. In this sense, the ML-based model is a powerful tool, but is not a final solution, especially without several considerations to apply these techniques in social sciences.

The path to developing effective approaches in complex scenarios, such as the identification of undocumented immigrants, is a considerable challenge. Without understanding all the real implications of our work, we will not offer truly effective solutions. In fields such as FA, to transform data into human stories and restore identities, it is necessary to consider all aspects of our methodological approaches, even those that force us to remember the true social objective of our discipline. This is essential to avoid falling into a simplistic reductionism that leads us to believe that systems based on AI and ML are a magic solution for the human identification process. In the midst of these challenges, George Box’s reflections become important, because “to find out what happens to a system when you interfere with it you have to interfere with it (not just passively observe it).”


Noemí Aedo-Noa

Noemí Aedo-Noa is a Physical Anthropologist and PhD student in Biomedicine at the University of Granada. Her research focuses on the use of Machine Learning tools for estimating age-at-death. She is interested in the biological profile reconstruction methodologies, the sociocultural role of Forensic Anthropology, and the potential use of programming languages in forensic investigation.

Cite as

Aedo-Noa, Noemí. 2024. “Transforming Data into Human Stories: Machine-Learning-Based Models to Estimate Age-at-Death of Undocumented Migrants.” Anthropology News website, June 27, 2024.