At major sporting events such as the Olympic Games, the blood and urine samples of thousands of athletes are examined to find possible evidence of doping. However, the capacity for the analysis is limited, so the controls remain patchy. A solution could now be provided by software that uses artificial intelligence to automatically search for abnormalities in the samples. Since it uses previous urine samples from the same people as a reference, the researchers say it could also track down cases in which samples were exchanged.
So far, very few cases of doping have been reported during the Olympic Games in Paris. But does that mean that the competition is actually largely “clean”? Or are the controls inadequate? At previous Olympic Games, it was often only discovered afterwards that many medal winners had taken banned performance-enhancing substances. In addition, not all urine samples are tested for all conceivable substances. For many doping substances, the official protocol of the World Anti-Doping Agency (WADA) only prescribes random checks due to the relatively high effort involved. At the Olympic Games in Paris in 2024, too, there is criticism that the capacity to evaluate the tests is insufficient to reliably and quickly identify black sheep.
Capacity for human review too low
A team led by Maxx Richard Rahman from Saarland University has now presented software at the International Joint Conference on AI in South Korea that uses artificial intelligence to evaluate which urine tests are not suspected of doping and which should be subjected to closer analysis by humans. “So far, all samples have been evaluated manually,” explains Rahman’s colleague Wolfgang Maaß. As a result, the already incomplete evaluation usually takes weeks – and is only completed long after the end of the respective sporting event.
Another problem: When doped athletes exchange their own urine samples for “clean” samples from another person, the laboratory results are unremarkable. This is how the doping of the Russian team at the 2014 Winter Olympics in Sochi initially went undetected. So far, only DNA analyses can prove whether a sample actually comes from the person who claims to have provided it. “But that is expensive and time-consuming,” says Maaß. The procedure is therefore not suitable as a standard procedure for all samples.
Evaluation with Artificial Intelligence
Rahman and Maaß, together with their team, have therefore developed a different approach. “The whole problem practically cries out for machine analysis,” says Maaß. The researchers have developed software that can analyze data from urine samples automatically, quickly and inexpensively. “During doping tests, the concentrations and ratios of various steroids are measured and checked for conclusiveness,” explains Maaß. The AI software compares this data with previous test results from the same person. The program only needs three samples to learn which concentrations of individual substances are typical for the respective athlete.
The software then compares the new measurements with previous results. If everything matches, it classifies the sample as normal. Only samples that show deviations from the usual pattern – for example because the person has been doping or because the urine sample was replaced by that of another person – are then sent for human examination. “This small remainder can then be examined more closely by the biochemists in the laboratory using DNA analysis,” says Maaß.
Experiments with real data sets have shown that the software does indeed reliably sort out innocents and can thus help to focus on cases where there is reasonable suspicion. “Someone who has increased their performance with a substance that can be detected in urine can be found with relative certainty using our software,” says Maaß. The researchers hope that their system can help to track down doping offenders better and more quickly in the future.
Source: Maxx Richard Rahman (Saarland University) et al., International Joint Conference on Artificial Intelligence, IJCAI, 2024