Statistical and Spectral Analysis of EEG Biomarkers for Reliable Stress Detection During Yoga-Based Relaxation Practices

Abstract

Stress detection using electroencephalography (EEG) has gained significant interest due to its potential for objective neurophysiological assessment. However, existing models often suffer from subject dependency, limited statistical validation, and lack of biomarker interpretability. This study presents a statistically validated EEG-based stress detection framework evaluated during yoga-based relaxation practices. The proposed framework was evaluated using Leave-One-Subject-Out (LOSO) cross-validation, in which each participant was held out for testing while the remaining participants were used for model development. The resulting accuracy, balanced accuracy, AUC-ROC, and F1-score were 1.000 ±0.000 (95% CI: [1.000, 1.000]) across the evaluated folds. All data-dependent preprocessing, feature selection, and model tuning were restricted to the training participants within each fold. The reported performance reflects the evaluated dataset and validation protocol and requires independent external validation before broader generalization can be established. High-dimensional statistical evaluation using paired t-tests with False Discovery Rate (FDR) control identified significant neurophysiological differences across states. The initial analysis identified a difference in the alpha-to-beta ratio at Fz between the evaluated conditions (t = −29.267, pFDR = 6.188 × 10⁻⁵⁴). This result requires verification of the statistical procedure and multiple-comparison correction before a definitive interpretation. Post-hoc permutation feature importance analysis initially produced values of 0.000 ± 0.000 across the evaluated features. The initial permutation feature-importance analysis yielded 0.000 ± 0.000 across the evaluated features. These values require verification and do not, by themselves, establish feature redundancy or distributed feature contributions. A linear mixed-effects model was considered to examine variation associated with participant identity. The initial mixed-effects analysis reported a participant-level random-intercept variance of 0.000. This boundary estimate does not establish the absence of participant effects or subject-related confounding. The potential contribution of spectral and network-derived EEG features to subject-independent stress classification requires confirmation through verified evaluation and independent external validation.