Machine Learning-Based Sex Estimation from Computed Tomography-Derived Pelvic Morphometric Measurements in a Contemporary Turkish Population.

Background/

Objectives

Sex estimation is a fundamental component of the forensic biological profile, and the pelvis is widely regarded as the most sexually dimorphic skeletal element. This study aimed to quantify pelvic sexual dimorphism in a contemporary Turkish sample using computed tomography (CT)-derived morphometric measurements, and to evaluate the performance of several machine learning (ML) algorithms for sex classification.

Methods

Fourteen pelvic measurements were obtained from CT reconstructions of 201 individuals (101 males, 100 females) by a single observer; a stratified subset of 30 cases (15 males, 15 females) was remeasured by the same observer and by a second observer to assess intra- and inter-observer reliability. Four supervised ML classifiers-logistic regression, linear discriminant analysis (LDA), random forest, and support vector machine (SVM)-were trained and evaluated using stratified 10-fold cross-validation and an independent 30% hold-out test set.

Results

Thirteen of the fourteen measurements differed significantly between sexes (p < 0.05). Intra-observer reliability was excellent for all 14 measurements (ICC = 0.943-0.998), and inter-observer reliability was good to excellent (ICC = 0.834-0.996), with the lowest value observed for pubic length. Cross-validated classification accuracy ranged from 98.0% (random forest) to 99.0% (logistic regression, LDA, and SVM), with corresponding cross-validated AUC values of 0.991-0.994, with logistic regression, LDA, and SVM each achieving 100% accuracy (AUC = 1.000) on the independent hold-out set (random forest: 96.7%, AUC = 0.999). Acetabular width was identified as the single most informative predictor (univariate area under the curve = 0.972), followed by acetabular height, the subpubic angle, ischial length, the transverse pelvic outlet, and the angle of the greater sciatic notch.

Conclusions

These findings demonstrate that CT-derived pelvic morphometrics, combined with machine learning classification, provide a highly accurate and reproducible method for sex estimation in a Turkish forensic context, and support the development of population-specific standards for forensic anthropological casework, pending independent external validation; the classification and effect-size estimates reported here were, however, stable under repeated cross-validation and hyperparameter optimisation.