Age-at-Death Estimation based on Symbolic Regression Ensemble Learning from Multiple Annotations

The present study addresses the problem of semiautomatic age-at-death estimation from pubic symphysis, a crucial yet complex task in forensic anthropology. Its accuracy directly depends on the quality of the pubic bone trait labeling developed by the forensic practitioners, affected by an inherent uncertainty in their definition. As interpretability is a mandatory requirement, we propose an approach where the model design is based on evolutionary learning, considering genetic programming to frame the problem as a symbolic regression task. Additionally, ensemble learning is considered to address the challenges posed by noise, uncertainty, and conflicting annotations inherent in data collected from multiple subjects. Ensemble learning provides an effective approach to navigate these challenges by facilitating consensus-building through decision making and information fusion. Hence, observer committees are formed, comprising multiple forensic specialists with different skills and expertise which provide alternative annotations. Several ensemble configurations combining different weak learners and aggregation operators are tested to assess their effectiveness in improving accuracy and reliability in age-at-death predictions. Their performance is compared against models trained on single annotations, revealing an improvement in predictive accuracy. The obtained results also highlight the benefits of incorporating diverse perspectives to address the complexities associated with human variability and anatomical assessments.