A Multi-Scale Adaptive EEG Feature Selection and Weighting Method for Mental Workload Estimation in Rapid Serial Visual Presentation.

Accurate mental workload estimation is crucial for enhancing performance and reliability in Rapid Serial Visual Presentation- based Brain-Computer Interfaces (RSVP-BCIs). However, existing (Electroencephalography)EEG-based methods suffer from high-dimensional data, feature redundancy, and inflexible feature representations, limiting their accuracy and generalization. In this work, we designed an RSVP-based aircraft target detection task with varying presentation rates to induce workload, collecting behavioral, subjective, and high-density EEG data. And, we propose a Multi-scale Adaptive Feature selection and Augmented Weighting (MAFA) framework, combining adaptive channel selection with multi-scale feature compression and weighting for mental workload classification. First, an adaptive channel selection model is developed to assess channel importance through Ridge Regression and automatically determine the optimal subset of electrodes through Kneedle-based knee point detection for reducing spatial redundancy. Second, to highlight discriminative features, multi-scale features extracted from EEG are reweighted with averaged coefficients vector derived from L1-regularized multinomial logistic regression. Experimental results indicate that our multi-rate RSVP paradigm can elicit distinct workload levels with significant differences in EEG features across these levels. Our method achieves higher classification accuracy than comparison approaches. The visualization of EEG features weights reveals that the proposed MAFA framework can adaptively pay attention to features in posterior regions dominated by theta and alpha bands, which is consistent with the analysis of brain patterns in workload. These results demonstrate the feasibility and interpretability of our proposed MAFA for mental workload estimation, highlighting its potential application in RSVP-BCI systems.