AI-Driven Doping Detection Using LOOCV to Address Data Imbalance in Mass Spectrometry Analysis

Abstract

Doping testing in sports presents significant analytical challenges due to the vast array of banned substances and the large volume of samples requiring analysis. Identifying a positivity rate as low as 1% from extensive datasets is particularly demanding. In this study, we developed and applied an artificial intelligence (AI)-based diagnostic system to automate the analysis of mass spectrometry (MS) data for detecting doping substances. The AI algorithm classifies samples by analyzing peak areas corresponding to the retention times of detected compounds. A total of 684 athlete urine samples―643 negative and 41 positive―were analyzed using gas chromatography coupled with tandem mass spectrometry (GC-MS/MS) to detect 153 banned substances. We trained six different machine learning models―logistic regression, K-nearest neighbor (KNN), support vector machine (SVM), Gaussian Naive Bayes, random forest (RF), and extreme gradient boosting―using both K-fold cross-validation and leave-one-out cross-validation (LOOCV). The RF and KNN models trained using LOOCV achieved precision, recall, accuracy, and F1 scores of 100%. LOOCV demonstrated superior capability in addressing sample imbalance issues, significantly enhancing the system’s ability to detect rare positive cases, as compared to K-fold cross-validation. This AI-based approach substantially reduces manual effort, accelerates the speed and consistency of doping testing, and minimizes human errors. The study underscores the potential of LOOCV to improve analytical methods, particularly in the context of complex and imbalanced datasets common in anti-doping analyses.