Sex estimation from the greater sciatic notch: a comparison of classical statistical models and machine learning algorithms

The greater sciatic notch (GSN) is a useful element for sex estimation because it is quite resistant to damage, and thus it can often be assessed even in poorly preserved skeletons. This study aimed to develop statistical models for sex estimation based on visual and metric analyses of the GSN, and additional variables linked to the GSN. A total of 60 left coxal bones (30 males and 30 females) were analysed. Fifteen variables were measured, and one was a morphologic variable. These 16 variables were used for the comparison of six statistical models: linear discriminant analysis (LDA), regularized discriminant analysis (RDA), penalized logistic regression (PLR) and flexible discriminant analysis (FDA), and two machine learning algorithms, support vector machine (SVM) and artificial neural network (ANN). The statistical models were built in two steps: firstly, only with the GSN variables (group 1), and secondly, with the whole variables (group 2), in order to see if the models including all the variables performed better. The overall accuracy of the models was very close, ranging from 0.92 to 0.97 using specific GSN variables. When additional variables starting from the deepest point of GSN are available, it is worth to use them, because the accuracy increases. PLR (after optimization of parameters) stands out from other statistical models. The position of the deepest point of GSN (Fig. 2) probably plays a crucial role for the sexual dimorphism, as stated by the good performance of the visual assessment of this point and the fact that the A2 angle (posterior angle with the deepest point of the GSN as the apex) is included in all models.

Automated summary

This study developed statistical models for sex estimation based on visual and metric analyses of the greater sciatic notch (GSN) and additional variables linked to the GSN. The models showed overall accuracy ranging from 0.92 to 0.97, with penalized logistic regression (PLR) standing out after parameter optimization. Additionally, the inclusion of additional variables starting from the deepest point of the GSN improved accuracy.