Abstract
The steroidal module of the Athlete Biological Passport (ABP) longitudinally monitors five ratios between urinary concentrations of endogenous anabolic and androgenic steroids. Even though it has improved detection of testosterone doping, the interpretation of data from multiple discrete biomarkers is complex. This study sought to create a single score to identify doping rather than relying on the interpretation of each parameter alone. A Bayesian model was used to define an ABP sequence probability for each biomarker to assess the extremity of a measurement relative to the expected levels from ABP. This was used to discriminate between doped and presumed clean individuals based upon pattern classification of biomarkers using classification algorithms. Data were obtained from laboratory-controlled experimental studies as well as routine doping control tests. A laboratory model (where classifier is trained using the laboratory-controlled data only) and a mixed model (where classifier is trained on combined laboratory-controlled and doping control data) were developed and tested on the doping control data. Logistical regression was seen to have the best classification performance across the methods used, with the Abnormal Steroid Profile Score (ASPS) representing the estimated probability from the logistical regression model. Classifier performance produced an AUC of 0.67 and 0.75 when trained on the laboratory model and the mixed model, respectively, with T/E and 5α-Diol/5β-Diol representing the main biomarkers driving the ASPS. These findings demonstrate that the ASPS can discriminate between the doping status of individuals, even if a mixture of steroids, administration methods and doses are used.