Detection of sample swapping in anti-doping investigations using machine learning.

Area of Science:

• Sports Science
• Analytical Chemistry
• Machine Learning

Background:

• Sample swapping, the substitution of a urine sample with a clean one, is prohibited by the World Anti-Doping Agency (WADA).
• Detecting an athlete reusing their own clean sample is difficult using conventional analytical methods.
• The Athlete Biological Passport relies on robust methods to ensure sample integrity.

Purpose of the Study:

• To propose a novel similarity detection framework for identifying reused or identical urine samples.
• To enhance the detection of sample swapping in anti-doping efforts.
• To leverage machine learning for improved analysis of urinary steroid profiles.

Main Methods:

• Development of a similarity detection framework utilizing convolutional neural networks.
• Analysis of complex patterns and subtle variations in urinary steroid profiles.
• Evaluation using a large dataset of 67,651 steroid profiles and both synthetic and real-world similar samples.

Main Results:

• The proposed convolutional network framework demonstrated superior performance compared to baseline models.
• The framework achieved higher accuracy in detecting similar urine samples, accounting for realistic variability.
• Machine learning offers a powerful tool for automated detection in large-scale anti-doping sample collection.

Conclusions:

• The developed framework shows significant potential for improving the accuracy and efficiency of anti-doping detection methods.
• Machine learning, specifically convolutional networks, can effectively address the challenge of sample swapping.
• Automated detection of sample swapping will strengthen the integrity of the Athlete Biological Passport program.