Abstract 4370142: Machine Learning–Driven Optimization of EEG Channel Selection for Cognitive State Monitoring

Abstract

Background: Wearable electroencephalography (EEG) devices enable noninvasive, real-time monitoring of cognitive states such as attention and stress. However, their practical deployment is limited by high data dimensionality, interchannel multicollinearity and hardware constraints making efficient channel selection critical for balancing performance and usability. Hypothesis: A sparsity-promoting method that combines Elastic Net–regularized Common Spatial Patterns (EN-CSP) with Automated Machine Learning (AutoML) can isolate minimal yet highly informative EEG channel configuration without sacrificing classification fidelity. Methods: Electroencephalography recordings from the Mental Attention States (MAS) cohort (14 channels) and the Stress&Anxiety Monitoring (SAM) cohort (32 channels) were band-pass filtered (0.5–45 Hz) and cleaned of eye-blink artifacts. From each epoch, we derived spectral and time-domain features. Electrode utility was quantified with EN-CSP that applied joint L1/L2 regularization and 1,000 stratified bootstrap runs produced stable importance scores for every channel. Highest-ranked channels were screened with Minimum Redundancy Maximum Relevance (MRMR) and Bagged Classification Ensemble (BCE) algorithms to group into candidate configurations containing 4, 8, 12, or 16 electrodes. The AutoML pipeline selected the optimal classifier for each configuration and summarized performance by overall accuracy and F1 score. Results: Across the two cohorts, the EN-CSP pipeline reduces the number of required electrodes by roughly one-half without degrading performance. In the MAS cohort, the BCE reached a macro-F1 of 0.889 and 91.4 % accuracy when restricted to the eight highest-ranked channels; MRMR and EN-CSP followed closely with F1 scores of 0.874 and 0.867. In the SAM cohort, EN-CSP is most optimized using only the four most informative channels, achieving an F1 of 0.752 and 81.3 % accuracy. Electrode rankings generated by the AutoML validation step were highly concordant with EN-CSP scores (Spearman ρ = 0.78, p < 0.01). Frontal (F3, F4), occipital (O1, O2), and parietal (P7, P8) leads were repeatedly identified as the most informative across both tasks. Conclusions: The novel machine learning framework consistently identified compact channel sets that preserved predictive accuracy while halving electrode count, supporting its value for lightweight, wearable EEG platforms aimed at continuous monitoring of attention and stress.