Skeletal Sex Estimation for Human Remains From Archaeological Contexts: Machine Learning Models Based on Ancient Dion, Greece

The estimation of sex in the analysis of human remains from archaeological contexts is an essential tool for reconstructing the demographic profile of past populations and their lifestyles. Methods for skeletal sex estimation are commonly based on visual assessment of the pelvis and cranium, but their application is often limited by the poor preservation of these elements in archaeological collections. Several standards have been developed to predict skeletal sex from metric methods, but interpopulation differences and secular change make the applicability of these methods in archaeological contexts problematic. In this paper, we propose population-specific standards for sex estimation using metric data from the postcranial skeletons of 48 individuals (18 males and 30 females) excavated at ancient Dion, Greece. We applied different imputation methods for missing data and different models for sex prediction (Logistic Regression, XGBoost, LightGBM, and Random Forest) and compared their performance using a range of metrics. The results show that classification performance varies depending on the skeletal measurements used, the amount of missing data, and whether variables are analyzed individually or in groups. Nonetheless, the accuracies achieved are very high (around or above 90%), both for most univariate and almost all multivariate models. Despite the limitations imposed by the small size of the sample, more such initiatives in the future will improve population-specific sex prediction models by including additional archaeological assemblages from other regions and periods and assemblages with larger sample sizes.