Machine Learning-Based Classification Models for Brain Doping Detection from Urinary Neurochemical Alterations

Abstract

Brain stimulation techniques that enhance cognitive and behavioral performance through electrical stimulation of specific brain regions are increasingly misused in sports. This practice, termed brain doping, enables performance enhancement without effort, undermines the spirit of sport, and poses health risks due to uncontrolled stimulation. Despite these concerns, no established monitoring method exists for brain doping detection. In our previous study, LC–MS–based urinary neurochemical analysis identified markers reflecting stimulation effects beyond individual variability and exercise-induced changes. However, conventional linear statistical approaches were insufficient to clearly distinguish stimulated from sham conditions. To address this limitation, we applied machine learning methods capable of capturing nonlinear relationships and built classification models. We employed three classification models — random forest, XGBoost, and logistic regression — optimized with a hybrid feature selection strategy combining filter and wrapper methods using 34 neurochemicals. Synthetic data were used for training in time-specific models, whereas only real samples were used in the integrated model. Model performance was evaluated by 5-fold cross-validation using the F1 score and the area under the receiver operating characteristic curve (AUC). In the time-specific models, the optimized model using individual neurochemical features achieved an AUC of 0.9089 at 2 hours post-stimulation, and the model including ratio-based features achieved perfect classification at 8 hours post-stimulation. In the integrated model, the logistic regression model showed superior classification performance with an AUC of 0.9213 and an F1 score of 0.8254, demonstrating the strongest discriminative power for positive samples. Although based on a limited dataset, we demonstrated the feasibility of constructing classification models for brain doping detection. We will expand clinical samples, validate the physiological relevance of selected features, and further advance the models. Importantly, this study proposes the first machine learning–based strategy for brain doping detection, providing valuable scientific evidence to support future monitoring frameworks by the World Anti-Doping Agency.