ERA flag optimization for longitudinal data and routine implementation

Code: 24T02WM

In this project, tools to support anti-doping decision-making in EPO detection (i.e. the decision to carry out an EPO analysis on an existing sample) shall be studied. Therefore, the ability to identify the presence of EPO in blood samples may be improved by developing a pattern recognition/classification tool that provides a risk score for EPO abuse by sample. Here, the anonymized hematological profiles of athletes shall be studied for finding the most suitable indicators through statistical-based methods, deploying state-of-the-art deep learning algorithms, developing a fully-fledged pipeline model, and performing a decision analysis on the developed model for the detection of suspicious hemtological profiles. The project will focus on further exploring the application of already developed algorithms and the potential of new state-of-the-art algorithms. Here, a set of hematological parameters shall be explored that are consistent with the hematological variables reported in the hematological module of the Althete Biological Passport (ABP) in the Anti-Doping Administration & Management System (ADAMS) database. An assessment of the performance improvement resulting from the addition of variables present in the information set of the previous version of the algorithm but not in ADAMS will be conducted. In the previous phase, due to fewer data samples, each blood sample was considered in isolation, i.e., EPO abuse prediction was made independent of the athletes' history, and the longitudinal behavior of the samples was not explored. In the current project, the longitudinal aspect of the individual's sample series will be explicitly integrated (i.e., to exploit the biological signature by minimizing the impact of between subject marker variance) using state-of-the-art algorithms for processing sequential data. The tool will be optimized for both scenarios. i.e., it will be able to search within already collected samples (updatin the associated probabilities for existing samples each time a new sample is added), and it will be able to compare new hematological data from a new sample to the already existing data in the profile directly after the upload. Within the development phase, different levels of sensitivity and specificity will be tested to figure out reasonable parameters for the approach.

In order to support the deign, development, and future embedding of the machine learning tool into existing procedures, interviews with stakeholders involved in the decision process of EPO detection will be conducted, specifically with Athlete Passport Management Units (APMUs), who are tasked with recommending EPO analysis on specific samples. In addition, we will evaluate the tool together with the APMUs to capture the potential impact of machine learning supported tools on decision-making in EPO detection.