Case-specific accuracy in sex estimation from long bones in forensic anthropology: An "accuracy x-factors" approach

Sex estimation from human skeletal remains is a cornerstone of forensic anthropological analysis. Long bones, despite exhibiting less pronounced dimorphism than pelvis, serve as invaluable substitutes. However, traditional statistical approaches for sex estimation from long bone measurements often lack the precision and case-specific reliability demanded by stringent legal standards. This study addresses these critical limitations by rigorously exploring the potential of machine learning (ML) to significantly enhance sex estimation from long bones.

We analyzed 16 osteometric measurements from the humerus, radius, femur, and tibia of 2969 individuals (1207 females, 1762 males) across eight skeletal collections. Eleven ML algorithms were trained and cross-validated, then validated on an independent South African sample. To address the common issue of incomplete remains, we developed an "accuracy x-factors" approach. This method simulates missing data scenarios and selects tailored training subsets, yielding individualized reliability assessments adapted to specific measurement availability.

Linear Discriminant Analysis (LDA) consistently achieved the highest performance, with accuracies up to 93 %. The "accuracy x-factors" approach proved effective in providing per-individual confidence measures, highlighting that prediction reliability varies with data completeness. Adjusting thresholds to higher confidence levels (e.g., >0.7) substantially reduced error rates, allowing a conservative yet legally robust classification of a smaller but more reliable subset of cases.

ML offers a powerful framework for sex estimation from long bones. The proposed "accuracy x-factors" approach introduces a significant methodological advance by delivering transparent, case-specific confidence levels. This strengthens both the forensic applicability and the legal admissibility of long bone-based sex estimation.