EEG-Based Machine Learning Classification of Shooting Conditions and Performance Levels During the Shooting Preparation Stage

Background: Shooting performance depends on environmental conditions and neural state during preparation. Few studies have applied machine learning (ML) to classify shooting conditions and performance levels using EEG. This study assessed whether pre-shot EEG features could discriminate shooting conditions and performance levels within the same condition. Methods: EEG and behavioral data from 35 shooters were analyzed under normal, noise, and low-light conditions. Band-power, phase-locking value (PLV)-based functional-connectivity, and filter bank common spatial pattern (FBCSP) features were evaluated separately using support vector machine (SVM) and random forest (RF). EEGNet served as an end-to-end benchmark for performance classification. Repeated trial-wise five-fold cross-validation assessed within-dataset trial-level discrimination, while supplementary leave-one-subject-out cross-validation (LOSO-CV) assessed generalization to unseen shooters in the performance task. Results: All three feature types provided discriminative information for shooting conditions and high- versus low-performance trials. FBCSP combined with SVM achieved the best within-dataset performance, reaching 91% for normal versus low-light classification and 81% for high versus low performance under the normal condition. Because trials from the same participants could occur in both training and test folds, these accuracies were not estimates of subject-independent performance. LOSO-CV accuracies ranged from 59% to 70%, indicating limited cross-subject generalization. Beta-band features showed relatively consistent discriminative contributions. Gamma-band and frontal-channel findings require caution because residual muscle activity and attenuation of frontal EEG during virtual-EOG-based correction could not be excluded. Conclusions: Pre-shot EEG features showed preliminary within-dataset discriminability, but inter-individual variability limited generalization. Larger independent datasets, dedicated EOG/EMG recordings, and subject-adaptation strategies are required before practical deployment.