The detection in urine of recombinant human erythropoietin (rHuEPO), a hormone misused by endurance athletes as a doping agent, is based on the differentiation of its isoelectric pattern from that of the corresponding natural hormone. Different empirical criteria have been proposed for discriminating the images of the patterns but none of them have been elaborated from a rational statistical approach. Discriminant analysis was applied to a dataset of profiles defined as positive (116 profiles from 26 subjects) (presence of rHuEPO and possibly residual natural endogenous hormone) and negative (131 profiles from 131 subjects) (presence of natural endogenous hormone only). The different bands were numbered according to a template of 16 possible positions and their relative intensities constituted the 16 variables of the statistical analysis. This method was then tested with data from an administration trial of low doses (6.7–10 IU/kg) following high-dose (265 IU/kg) injections (71 profiles from one subject). The analysis of the dataset clearly separated the negative and positive profiles. A cross-validation procedure confirmed that the analysis was extremely stable: with ten-fold cross-validation, no false positives were observed even with 100 000 simulations. Furthermore, the detection of rHuEPO in the profiles from the low-dose trial was greatly improved in comparison with a previously validated empirical criterion.