Sex estimation from patellar measurements using machine learning in a Turkish population

Forensic investigators strive to determine the four main characteristics of biological identity: ancestry, sex, age, and stature. In forensic contexts where primary indicators such as the pelvis, skull, or long bones are missing or highly fragmented, alternative skeletal elements become essential for reliable sex estimation. Addressing this need, the present study aims to develop a robust sex estimation model using patella measurements derived from computed tomography scans of a Turkish population from the Central Anatolia region. The study sample comprised a total of 794 patellae from the Karaman Training and Research Hospital archives. Sex was predicted using eight different machine learning algorithms from patellar height, width, and thickness measurements. In order to enhance the reliability of the study, 5-fold cross-validation was applied, and the dataset was split into 80–20% training-test data. The accuracy rates of predictive models were found to be between 81.86 and 88.54%. The AUC values of these models were found to be between 81.78 and 88.66%. Patellar width was determined to be the most effective parameter for sex estimation. Overall, the logistic regression algorithm produced the best results, with an accuracy of 88.54% and an AUC of 88.66%. For sex estimation, machine learning models generated from patellar measurements have the potential to achieve high accuracy in the Turkish population.