Exploring the complexity of African populations variability with Machine Learning

Human skeletal remains are an immense source of data to describe human biodiversity with an intrinsic complexity due to the multifactorial origin of human variability. Evolution and ontogeny produced complex patterns of variation through contingent events and adaptations. Multivariate approaches have been widely adopted in physical anthropology; however, at present, Artificial Intelligence algorithms have scarcely been applied to such datasets. Data analysis techniques based on Artificial Intelligence algorithms have shown to be suitable in many different fields, from engineering and medicine up to cultural heritage and Egyptology. In this work we aim to show how Machine Learning algorithms can be applied in the field of anthropology, using the W.W. Howells dataset of cranial measurements, limited to the analysis of African populations. Principal Component Analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), Spectral Embedding and Uniform Manifold Approximation and Projection (UMAP) were used for dimensionality reduction, along with supervised and unsupervised methods to explore and quantify the differences due to ancestry and sex in the skulls of African populations. Algorithms such as Support Vector Machines and the unsupervised DBSCAN were applied to the data in order to quantify this similarity. This strategy allows a discrimination of sex and ancestry (about 85% of accuracy for both) in human remains, ultimately opening up new routes for anthropological research.