Enhancing the accuracy of intelligent doping diagnosis through non-labeling machine learning optimization

Abstract

Unintentional positive doping tests due to anti-doping substances in foods and dietary supplements are on the rise. This trend highlights the need for analytical methods that can rapidly and accurately screen for a wide range of doping substances in foods and dietary supplements. This study presents a novel high-throughput analytical method designed to screen for a wide range of anti-doping substances in these matrices. QuEChERS (Quick, Easy, Cheap, Effective, Rugged, and Safe) sample preparation techniques and liquid chromatography-tandem mass spectrometry (LC-MS/MS) were chosen as the main analytical techniques due to their excellent sensitivity, selectivity and ability to handle complex mixtures. The method has been optimized to efficiently separate the target compounds in 10 minutes and its high throughput allows rapid screening of large sample volumes, making it suitable for routine analysis in doping control laboratories. To validate the applicability, the developed method was applied to 10 food matrices with different physical properties and successfully identified and quantified the target compounds. In conclusion, the developed LC-MS/MS method provides a reliable, sensitive and high-throughput approach for the screening of more than 400 compounds in various foods and dietary supplements. This is expected to strengthen doping control measures in sports, helping to reduce unintentional doping and maintain fair and clean competitioAccurate and prompt identification of positive and negative doping cases is paramount to ensure fairness in assessments. Our previous research has demonstrated the feasibility of AI-based doping diagnosis through machine learning applied to datasets from doping control center. However, the inherent class imbalance, where positive doping cases represent less than 1% of all results, poses a considerable challenge during model training. While previous works have employed deliberate grouping strategies to mitigate this imbalance, there is concern that this may oversimplify the complexity of the dataset. In present study, we adopted a non-labeling approach that avoids grouping and prioritizes hyperparameter optimization through a grid search technique. This methodological shift allowed us to fully exploit the potential of machine learning in doping diagnosis, effectively addressing class imbalances without compromising the enriched nature of the dataset. This approach demonstrates the capability to maintain consistently high accuracy, even with the inclusion of more substances and data. These advancements are anticipated to substantially improve anti-doping efforts in sports.