Artificial intelligence could help keep sporting competitions clean and fair in the future: Professor Wolfgang Maaß and his team at Saarland University and the German Research Centre for Artificial Intelligence are researching how to expose doping violations faster and easier with self-learning computer systems. In projects with the World Anti-Doping Agency Wada, the business informatics specialist is working on training data systems that he developed for Industry 4.0 with data from doping controls in order to efficiently uncover sporting cheating.
Unequal opportunities, unfair competition, unclean sport - when doping takes place, it is not only justice that falls by the wayside. Athletes also put their health at risk. The fight against doping is a difficult undertaking. Exposing sports cheating is complex and costly. Artificial intelligence could support doping controllers in this task, Professor Wolfgang Maaß is convinced. "Artificial intelligence methods can speed up doping control procedures and make them more efficient," explains the business information scientist. Normally, Maaß and his team ensure transparency in Industry 4.0 with their intelligent data systems at Saarland University and the German Research Centre for Artificial Intelligence (DFKI): they predict early and reliably whether, for example, disruptions to plants or supply chains are imminent and find suitable solutions. Now Maaß also wants to use the smart computer algorithms against doping. "Our data systems are also informative beyond the economic sector," he explains.
The results of three projects so far, in which Maaß is cooperating with the World Anti-Doping Agency Wada, provide clear indications that the Saarbrücken AI systems, which have been pre-trained with economic data, also work in the case of doping and can quickly and reliably detect sports fraud of various kinds. "Our research work so far has been very promising. The AI-based analyses of biochemical and other data from doping controls are delivering very good results," says Maaß. His vision is to support the work of doping laboratories in the future through analyses in the virtual laboratory.
Doping controls generate a lot of data. For example, blood and urine samples are taken during training and competitions, sometimes over long periods of time, analysed in laboratories and tested for banned substances and methods. Other information about the athletes is also collected. Since the controls can take place anytime and anywhere, the athletes also report their current whereabouts, for example. The AI data systems can be fed with all such data in order to track down manipulations.
With the help of machine learning methods and deep learning, the researchers teach the system to recognise doping accurately from typical patterns. It learns to identify the tiniest but characteristic doping features like pieces of a puzzle. For this purpose, Maaß' team trains the computer system with data from doping controls of many athletes. The system is able to look through all possible links in the data, i.e. it understands how the individual pieces of the multi-dimensional puzzle of a doping case are connected.
The AI system records and weights blood markers or steroid profile data in urine, for example, and takes into account causal and temporal processes in several samples as well as the chemical transformations of the substances in the body or even the effects of the doping substance. With a little training, it finds patterns and the smallest nuances in the digital data and columns of numbers that indicate deviations and are typical for doping cases. In this way, it predicts how likely it is, given a certain puzzle constellation, that someone has used banned substances or otherwise falsified the tests.
"To be able to train complex models, we need a lot of data material, especially on positive doping cases. With so-called generative machine learning methods - we speak of 'GAN' for short - we can enrich our data pool. Only a deep understanding of the data and the underlying biological correlations allows us to develop and test such AI models," explains Wolfgang Maaß.
In 2015, Maaß tested for the first time in cooperation with Wada whether his data systems are generally capable of detecting doping with the drug erythropoietin, or "EPO" for short, on the basis of an anonymised selection of biochemical analysis data. This method of blood doping involves increasing the number of red blood cells in order to transport more oxygen and thus increase performance. "We used several blood indicators and a series of questionnaires about athletes' activities as a data set for training," Maaß explains. When compared with the actual results of doping controls, the data system already showed a very good hit rate in this first, not yet specifically further developed stage.
Another project with Wada confirmed the very good prediction quality: this time Maaß and his research team worked with the Department of Nutrition, Exercise and Sport at the University of Copenhagen in a study also on EPO doping. Without any further background information, the Saarbrücken researchers used their data system to evaluate certain blood markers and blood values from different training groups with a total of 50 test persons: The computer algorithms recognised the EPO doping group and filtered it out of the comparison groups - including a group of athletes whose blood cell count was naturally elevated due to training at high altitudes.
In the most recent project with Wada and the Institute of Biochemistry at the Cologne Sports University, the results of which the team presented in December, the computer algorithms also detected other manipulation beyond the classic doping analyses through special training: "Our system also tracks down the exchange of urine samples," says Maaß. This type of cheating, in which the urine sample of a doped athlete is swapped with a "clean" urine sample, is extremely difficult for doping controllers to detect and involves highly complex analyses. With deep learning training, the Saarbrücken data system also recognises the typical patterns of such manipulations. "The sensitivity and specificity of our methods already reach high levels; however, we can also improve this even further. This suggests that the use of AI methods in doping can effectively support the controls," says Maaß, who now wants to further intensify the cooperation with Wada and the Cologne Sports University.
Currently, Wolfgang Maaß is working on expanding research activities against doping across borders: He is leading the establishment of a German-French network that is to intensify the use of artificial intelligence in the fight against doping: The French anti-doping laboratory in Paris (Laboratoire AntiDopage Français, LADF, Université Paris-Saclay), the French National Research Institute for Informatics and Automation INRIA (Institut national de recherche en informatique et en automatique), the Sport University Cologne as well as Maaß' chair at Saar University and his research group "Smart Service Engineering" at the German Research Centre for Artificial Intelligence, DFKI, are involved in this.
www.uni-saarland.de
