Estimating distribution shifts for predicting cross-subject generalization in electroencephalography-based mental workload assessment

Albuquerque, Isabela; Monteiro, João; Rosanne, Olivier; Falk, Tiago H. · 2022 · Crossref

DOI: 10.3389/frai.2022.992732

archive: archived pipeline: cataloged

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of cross-subject generalization in electroencephalography (EEG)-based mental workload assessment. While EEG is widely used to monitor cognitive load in high-stakes environments, models often fail to generalize across different individuals due to anatomical and environmental variability. Standard domain adaptation techniques typically assume "covariate shift," where only the input distribution changes while the conditional label distribution remains constant. However, prior research suggests that in EEG applications, the conditional distribution of labels given features also shifts between subjects. This work proposes a strategy to explicitly estimate both marginal and conditional distribution shifts between multiple data distributions to better understand and mitigate these discrepancies. The study utilizes the WAUC dataset, which contains EEG recordings from 18 subjects (9 male, 9 female, average age 27) performing mental tasks while engaging in physical activity (running on a treadmill or pedaling a stationary bike). The authors developed methods to quantify cross-subject mismatch: for conditional shift, they used k-nearest neighbor classifiers to estimate the error rate when applying one subject’s labeling rule to another’s data; for marginal shift, they employed pairwise subject classification error rates to estimate H-divergence. These estimates were aggregated into disparity matrices and normalized Frobenius norms. The experiments compared four feature spaces: raw features, per-subject whitening (z-score), and two baseline normalization strategies. Additionally, the paper evaluates the impact of these normalization schemes on actual mental workload prediction accuracy using a leave-one-subject-out (LOSO) evaluation scheme with a single-source single-target domain adaptation setup. The results demonstrate that different normalization strategies significantly affect the magnitude of estimated statistical shifts. The study finds that common practices in the EEG literature, such as normalizing spectral features against baseline periods, do not uniformly mitigate both conditional and marginal shifts. Instead, the choice of normalization scheme alters the statistical structure of the data, which in turn influences the performance of mental workload prediction on unseen participants. The paper provides a framework for quantitatively assessing the effectiveness of domain adaptation strategies by linking specific distribution shift estimates to model generalization errors. The significance of this work lies in moving beyond the standard assumption of covariate shift in EEG-based brain-computer interfaces. By providing tools to estimate both marginal and conditional shifts, the authors offer a more rigorous method for verifying the underlying assumptions of domain adaptation algorithms. This approach allows researchers to select appropriate feature spaces and adaptation strategies based on the specific statistical properties of their datasets, rather than relying on unverified assumptions. Consequently, this contributes to the development of more robust, calibration-free mental workload monitoring systems that can reliably generalize to new users in real-world applications.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.8-27b-gittensor on 2026-08-23; verification: pending re-verification.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• neuro workload indices