Unlocking forensic potential: machine learning-driven estimation of sex, age, and stature from proximal femur using dual-energy X-ray absorptiometry.

Beyond its clinical applications, Dual-Energy X-ray Absorptiometry (DXA) offers potential for forensic biological profiling, particularly in cases involving fragmented remains. This study establishes proof of concept for the use of DXA-derived variables, including Area, Bone Mineral Density (BMD), and Bone Mineral Content (BMC) across proximal femoral regions for the estimation of sex, age, and stature. Predictive performance was evaluated by directly comparing traditional baseline methods (Discriminant Function Analysis and stepwise regression) against various algorithmic approaches, including classical statistical models and modern Machine Learning (ML) algorithms. Feature importance analysis revealed that Area and BMC were the primary predictors for sex and stature, while BMD was critical for age estimation, consistent with age-related bone loss. For sex estimation, classical Logistic Regression (LR) outperformed the traditional Discriminant Function Analysis (DFA) baseline, increasing accuracy from 90.7% to 93.4% while demonstrating minimal class-specific bias. In stature estimation, Gaussian Process Regression (GPR) improved upon traditional stepwise regression, reducing the estimation error from 4.53 cm to 4.06 cm in the mixed-sex sample. When the sex-specific models were applied, stature estimation errors were 4.28 cm for males and 3.25 cm for females. For age estimation, traditional regression and GPR demonstrated comparable performance in the mixed-sex group (errors of 9.32 and 9.60 years). However, the models exhibited systematic bias at chronological extremes due to regression to the mean. To mitigate this, sex-specific modeling was employed to enhance prediction precision. This study affirms the proximal femur's potential as an effective "three-in-one" estimator when analyzed using DXA and demonstrates that the integration of optimized algorithms enhances overall predictive accuracy. These findings demonstrate the utility of DXA- and ML-based methods for future large-scale, population-specific forensic applications.