Predictive modeling of sex from proximal femoral morphometry using machine learning algorithms: a radiological study in Egyptian adults

Sex estimation is a foundational step in forensic identification, and the proximal femur, whether assessed through direct skeletal measurement or radiographic imaging, has shown considerable promise as a durable, sexually dimorphic anatomical region for this purpose. Machine learning (ML) classifiers offer a flexible alternative to traditional discriminant function analysis, yet Egyptian data combining radiographic proximal femoral morphometry with supervised ML remain limited. Hence, this study aimed to develop and compare ML models for sex estimation from proximal femoral radiographic measurements in an adult Egyptian population. Digital pelvic radiographs of 306 Egyptian adults, 135 males and 171 females, aged ≥ 30 years, were retrospectively analyzed. Four proximal femoral measurements, femoral head diameter, femoral neck axis length, femoral neck width, and intertrochanteric distance, were obtained after confirming bilateral symmetry. Nine supervised algorithms were trained and evaluated using cross-validation together with an independent hold-out test set, with performance assessed through accuracy, F1-score, and receiver operating characteristic analysis. All four measurements differed significantly between sexes, with femoral head diameter showing the largest effect size and emerging as the dominant predictor across most models. The decision tree classifier achieved the highest test-set accuracy at 86% (F1-score = 0.83), followed by the multilayer perceptron at 83% (F1-score = 0.82), although sex-specific accuracy revealed nontrivial classification bias for most models, with only gradient boosting and AdaBoost remaining within the 5% bias threshold generally considered acceptable in forensic practice. Area under the curve values ranged between approximately 0.85 and 0.91 on the test set. These findings suggest that proximal femoral radiographic measurements, and femoral head diameter in particular, may offer a reliable and population-specific basis for ML-assisted sex estimation in Egyptian adults. However, sex-specific performance, not accuracy alone, should guide the choice of any model intended for forensic application.