Detection of rEPO administration using deep learning on blood smears

Code: 23D05SV

The aim of this study is to use artificial intelligence (AI) to evaluate blood smears as a potential matrix to detect doping with recombinant EPO (rEPO). Previous publications reported changes in the Red blood Cell Morphology after rEPO administration (e.g. Macrocytes, Stomatocytes). During these times automated haematology analysers and manual microscopy were used for the estimation of these parameters. Recently, digital Morphology is a developing field in Haematology which enables the automated analysis of blood smears by artificial intelligence. With this technology it is possible to evaluate morphological changes with an increased precision, based on a higher number of cells and to discover even minor changes in cellular shapes. AI and deep learning is capable of revealing new insights which conventional approaches were lacking so far, like predicting molecular changes on cytomorphology. Therefore, the goal of this study is to identify relevant changes in cell morphology during rEPO administration which are not addressed using current state of the art techniques. In the long term, a blood smear-based athlete blood passport providing an individual erythrocyte signature might be a prospective application of monitoring athletes by using artificial intelligence based on this postulated deep learning model.

Main findings

A total of thirty healthy athletes participated in a controlled, interventional trial in which they received either rEPO or a placebo over a three-week period. Each participant provided 13 blood samples across an 11-week timeframe, encompassing pre-administration, administration, and post-administration phases. PBS were prepared and digitized at each time point for subsequent analysis.Hematological parameters, including hemoglobin concentration, hematocrit, and reticulocyte indices, showed distinct temporal patterns between the rEPO-treated and placebo groups. Digitized PBS were analysed using a customized deep learning pipeline based on the Haemorasis pipeline, specifically adapted for this study. For the binary classification task comparing PBS samples obtained after iron supplementation but without/before rEPO to those obtained at the peak of reticulocyte response following rEPO administration, the model achieved an area under the receiver operating characteristic curve (AUC) of 0.79 for the entire cohort. Subgroup analyses yielded an AUC of 0.72 for male participants and 0.66 for female participants. These preliminary findings suggest that AI-driven morphological analysis of PBS may serve as a novel and complementary approach for the detection of rEPO doping. However, further validation in larger and more diverse cohorts is necessary to confirm these results and refine the model for potential application in anti-doping efforts.