Biological identification from proximal femur using artificial intelligence in a Thai population

Abstract

The objective of this dissertation is to utilize proximal femur in the estimation of biological profiles. The proximal femur is a skeletal structure that articulates with the acetabulum of the pelvic bone, forming the hip joint, and constitutes approximately one-fourth of the femoral length. The identifiable features of the proximal femur, including the femoral head, fovea capitis, femoral neck, greater and lesser trochanters, and proximal shafts, can be used to estimate crucial biological parameters, including stature, sex, age, and ancestry. Therefore, in situations where only the proximal femur is available for examination, a comprehensive biological profile of an unknown individual can be obtained through osteometric approaches applied to the proximal femora. The study involved radiographic and manual measurements of 354 left femora, with comparison of virtual measurements obtained through an image processing program to manual measurements taken from dry femora. The virtual measurements showed good agreement with the dry femur data for all variables except for femoral neck diameter. However, the level of agreement between the two measurement methods did not fall within the acceptable error range. Thus, based on these findings, it cannot be recommended that the virtual method proposed in this study be used as a substitute for dry-bone measurement. The proximal femur may also be used to determine population affinity and estimate ancestry when skulls are not available for analysis. The study investigated the ancestral differences between Thai femora in a local skeletal collection and femora from different ethnicities in an international collection. The measurements obtained from the two collections were found to differ significantly, with the exception of femoral neck diameter and the distance between the apex of the greater trochanter and the lateral margin of the articular surface of the femoral head. The kernel logistic regression model was found to outperform the other algorithms tested for ancestry estimation, with an accuracy of 80.88%. Proximal femur dimensions can also be used to determine sex from fragmented remains. The Naïve Bayes algorithm achieved an accuracy of 91.2% and outperformed the other models tested. These findings reinforce the utility of proximal femur dimensions in sex estimation from fragmented remains, and may prove valuable in forensic cases where critical skeletal elements are absent. Furthermore, this study also aimed to determine the role of the proximal femur in estimating the stature-at-death of skeletal remains when complete long bones are not available. We used various machine learning algorithms to measure the proximal femur from radiographic images. The results showed that Gaussian process regression was the most effective method, with a mean error of 4.68 cm and a standard deviation of 3.93 cm. Finally, the study evaluated the usefulness of the proximal femur in the estimation of age by analyzing changes in bone density through radiographs. Femoral head, femoral neck, Ward’s triangle, and greater trochanter showed negative correlations with age-at-death, with females displaying stronger correlations than males. Additionally, machine learning models were evaluated for age estimation from radiographic images of proximal femora. The support vector machine performed the best for both sexes, with a root mean square error of 12.56 years and a correlation coefficient of 0.53. For females, the best-performing model was linear regression with femoral neck and Ward’s triangle as selected attributes, while for males, the best-performing model was linear regression. These results suggest that the proximal femur is useful in estimating biological profiles and can be applied in forensic contexts where critical skeletal elements are missing.