A Deep Learning Model for Ancestry Estimation with Craniometric Measurements

Ancestry estimation from human skeletal remains is a significant component in forensic anthropology studies. Although some tools have been developed, their performances are not very good in practice. In this paper, we evaluated the utility of the deep learning method in cranial ancestry estimation based on Howells craniometric data. Specifically, 2,524 cranial individuals from the Howells main datasets and 468 from the Howells test datasets were analyzed in the paper. The individuals of 82 craniometric measurements in the Howells datasets were clustered into six ancestry groups: African, Austro-Melanesian, Polynesian-Micronesia, East Asian, Native American, and European. After the data engineering process, the Howells datasets with 30 craniometric measurements were fed into a feedforward neural network (FNN) model. The FNN model was trained on 80% of the Howells main datasets, validated on 10% of the Howells main datasets, and tested on another 10% of the Howells main datasets. The model's prediction accuracy for all six ancestry groups reached 80.6%. Compared with popular ancestry estimation programs AncesTrees and Fordisc 3.1 using the Howells test dataset, the performance of the FNN deep learning method is on par overall and better for some ancestry groups.