A comparison between neural network and other metric methods to determine sex from the upper femur in a modern French population.

Abstract Forensic anthropologists are frequently asked to assess partial or badly damaged skeletal remains. One such request led us to compare the predictive accuracy of different mathematical methods using four non-standard measurements of the proximal femur (trochanter–diaphysis distance (TD), greater–lesser trochanter distance (TT), greater trochanter width (TW) and trochanter–head distance (TH)). These measurements were taken on 76 femurs (38 males and 38 females) of French individuals. Intra- and inter-observer trials did not reveal any significant statistical differences. The predictive accuracy of three models built using linear and non-linear modelling techniques was compared: discriminant analysis, logistic regression and neural network. The neural network outperformed discriminant analysis and, to a lesser extent, logistic regression. Indeed, the best results were obtained with a neural network that correctly classified 93.4% of femurs, with similar results in males (92.1%) and females (94.7%). Univariate functions were less accurate (68–88%). Discriminant analysis and logistic regression, both using all four variables, led to slightly better results (88.2% and 89.5%, respectively). In addition, all the models, save the neural network, led to unbalanced results between males and females. In conclusion, the artificial neural network is a powerful classification technique that may improve the accuracy rate of sex determination models for skeletal remains.