Machine learning algorithms in the estimation of sex from 3DCT-generated cranial and pelvic measurements.
Authors
Affiliations (3)
Affiliations (3)
- Department of Anatomical Sciences, School of Biomedical Sciences, Faculty of Health Sciences, University of the Witwatersrand, Johannesburg, South Africa.
- Department of Electrical Engineering, College of Engineering, Qatar University, Doha, Qatar.
- Department of Basic Medical Sciences, College of Medicine, QU Health, Qatar University, Doha, Qatar. [email protected].
Abstract
Sex estimation from skeletal remains is a key component of forensic anthropology, with the skull and pelvis being the most sexually dimorphic elements in terms of morphology. Traditional morphometric approaches, such as discriminant function and logistic regression analyses, have achieved high accuracy in estimating sex across various populations, including South Africans. However, the use of machine learning (ML) for sex estimation based on cranial and pelvic measurements has not yet been explored in any South African population. This study assessed the potential of ML algorithms to estimate sex from cranial and pelvic measurements derived from computed tomography (CT) scans of contemporary Black South Africans. The sample included 680 skeletal elements (400 crania and 280 pelvic bones) with an equal distribution of males and females. CT scans archived at the Department of Radiology, Charlotte Maxeke Johannesburg Academic Hospital, were reconstructed into 3D models using Xiris and IntelliSpace software, from which eight cranial and eleven pelvic measurements were collected. Seven classical ML algorithms were applied and feature ranking techniques were used to determine the most informative variables for sex estimation. A stacking ML model was then developed, incorporating the top three classifiers as base learners. Their outputs were combined and passed to different ML algorithms acting as meta-learners for final classification. The findings confirmed strong sexual dimorphism in cranial and pelvic bones, and the stacking models achieved superior accuracy (cranium: 80.3-94.3%; pelvis: 86.1-96.1%) compared to traditional multivariate methods, demonstrating the potential of ML in forensic sex estimation.