Efficiency of the Adjusted Binary Classification (ABC) Approach in Osteometric Sex Estimation: A Comparative Study of...

Area of Science:

• Forensic Anthropology
• Bioinformatics
• Machine Learning Applications

Background:

• Osteometric sex classification is crucial in forensic anthropology for identifying individuals from skeletal remains.
• Traditional methods can be limited by accuracy and potential sexing bias.
• Machine learning (ML) offers advanced computational approaches for improving classification accuracy.

Purpose of the Study:

• To evaluate the Adjusted Binary Classification (ABC) approach for enhancing osteometric sex classification accuracy.
• To compare the performance of multiple machine learning techniques (LDA, GLMB, SVM, LR) in sex classification.
• To assess the impact of univariate and multivariate models on classification accuracy and sexing bias.

Main Methods:

• Utilized 13 femoral measurements from 300 individuals in a modern Turkish population sample.
• Employed machine learning classifiers: linear discriminant analysis (LDA), boosted generalized linear model (GLMB), support vector machine (SVM), and logistic regression (LR).
• Trained univariate and multivariate models, selecting the top five performing measurements and using pools of variables, respectively, with and without the ABC approach.

Main Results:

• Unadjusted univariate and multivariate models achieved accuracies of 82-87% and 89-90%, respectively.
• The ABC approach boosted accuracy and predictive values to ≥95% in cross-validation.
• Multivariate models minimized sexing bias, classifying 81-87% of individuals, outperforming univariate models (28-75% classification rate).

Conclusions:

• The Adjusted Binary Classification (ABC) approach significantly enhances the reliability of ML-based osteometric sex classification.
• Multivariate models are superior to univariate models in reducing sexing bias and improving classification rates.
• Logistic Regression (LR) showed sensitivity with small sample sizes, while GLMB demonstrated stability across varying data sizes.