Sex classification using the human sacrum: Geometric morphometrics versus conventional approaches

Abstract

The human pelvis shows marked sexual dimorphism that stems from the conflicting selective pressures of bipedal locomotion and parturition. The sacrum is thought to reflect this dimorphism as it makes up a significant portion of the pelvic girdle. However, reported sexual classification accuracies vary considerably depending on the method and reference sample (54%-98%). We aim to explore this inconsistency by quantifying sexual dimorphism and sex classification accuracies in a geographically heterogeneous sample by comparing 3D geometric morphometrics with the more commonly employed linear metric and qualitative assessments. Our sample included 164 modern humans from Africa, Europe, Asia, and America. The geometric morphometric analysis was based on 44 landmarks and 56 semilandmarks. Linear dimensions included sacral width, corpus depth and width, and the corresponding indices. The qualitative inspection relied on traditional macroscopic features such as proportions between the corpus of the first sacral vertebrae and the alae, and sagittal and coronal curvature of the sacrum. Classification accuracy was determined using linear discriminant function analysis for the entire sample and for the largest subsamples (i.e., Europeans and Africans). Male and female sacral shapes extensively overlapped in the geometric morphometric investigation, leading to a classification accuracy of 72%. Anteroposterior corpus depth was the most powerful discriminating linear parameter (83%), followed by the corpus-area index (78%). Qualitative inspection yielded lower accuracies (64-76%). Classification accuracy was higher for the Central European subsample and diminished with increasing geographical heterogeneity of the subgroups. Although the sacrum forms an integral part of the birth canal, our results suggest that its sex-related variation is surprisingly low. Morphological variation thus seems to be driven also by other factors, including body size, and sacrum shape is therefore likely under stronger biomechanical rather than obstetric selection.

Fig 1. Landmark configuration representing the entire sacrum.

The landmark configuration consists of 44 fixed landmarks (red dots) and 56 sliding semilandmarks (blue dots) on seven curves (i.e., superior articular surface, dorsal and lateral crests, and auricular surfaces; blue lines).

Fig 2. PCA plots of the Procrustes shape coordinates (PC1 vs. PC2).

PCA plot of (a) the complete sample (n = 107), (b) the European subsample (n = 39), and (c) the African subsample (n = 23). The thin-plate-spline warps represent the real shape variation at the extremes of the ranges of variation. (d) Colour-coded representation of the morphology changes from the female into the male mean shape of the entire sample. Warm colours indicate a positive deviation, cold colours denote a negative deviation. Superimposed male (blue) and female (red) mean shapes in the (e) European and (f) African subsamples.

Fig 3. PCA plot of the sample without central europeans (n = 68).

The Central European specimens from the Weisbach collection were excluded to avoid overrepresentation of the European subgroup. The thin-plate-spline warps show the shapes at the extremes of the axes in (a) shape space (b) form space. (c) Pairwise comparisons between the mean shapes of the different subgroups (Europeans = blue, West Africans = red, Khoesan and Pygmies = orange, South East Asians = green, Native Americans = yellow).

Fig 4. Variation of the sacrum shape depending on size, geographic region, and sex.

(a) Multivariate regression against lnCS. The warped models are based on the regression coefficients and are shown, from left to right, at 50% and 25% smaller than average size, average size, and at 25% and 50% larger than average size. (b) Geographic group mean difference deformations between African and European subsamples (from left to right: exaggerated-African, African mean, general mean, European mean, exaggerated-European); (c) group mean difference deformations between females and males (left to right shapes: exaggerated-female, female mean, general mean, male mean, exaggerated-male). Note that the mean shapes used for group mean differences in (b) and (c) were not corrected for allometry as this is part of the variation in these groups.

Fig 5. Variation of the linear metrics in the European and African subsets.

(a) Scatterplot of sacral width (mm) and corpus width (mm); note the minimal overlap of females (red) and males (blue) in the European subset and the strong overlap in the African subset for corpus width. (b) Scatterplot of corpus depth (mm) and corpus width (mm) showing high correlation. (c) Fequency distribution of the corporo-basal index (CBI) and corpus-area index (CAI) for the complete sample. (d) Frequency distribution of the CBI and CAI for the European and the African subsamples; note the differences in the European (high discriminative power) and the African subsets (no discriminative power for CBI and moderate discriminative power for CAI).

Fig 6. ROC curves based on predicted group membership from linear discriminant analyses for all classification methods reporting the area under the curves (AUCs).

(a) ROC curves for the complete sample (n = 101), (b) the European subset (n = 39) and (c) the African subset (n = 22). Corpus depth was the single most discriminative measurement while sacrum width was the weakest for all samples. Note that dotted lines represent similar thresholds for two parameters (i.e., corpus depth/corpus area in a and b; corpus width/CAI and corpus area/CBI in c).