Label-Based Alignment Multi-Source Domain Adaptation for Cross-Subject EEG Fatigue Mental State Evaluation

Zhao, Yue; Dai, Guojun; Borghini, Gianluca; Zhang, Jiaming; Li, Xiufeng; Zhang, Zhenyan; Aricò, Pietro; Di Flumeri, Gianluca; Babiloni, Fabio; Zeng, Hong · 2021 · Crossref

DOI: 10.3389/fnhum.2021.706270

archive: archived pipeline: cataloged verified

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

Summary

This study addresses the challenge of cross-subject electroencephalogram (EEG) fatigue mental state evaluation, a critical task for reducing traffic accidents caused by driver fatigue. While EEG-based methods are effective for detecting mental fatigue, significant individual differences in EEG signals—caused by biological and physical factors—hinder the generalization of models across different subjects. Traditional unsupervised domain adaptation (UDA) methods often fail to extract robust domain-invariant features, particularly for samples with inconspicuous features near decision boundaries, leading to classifier confusion and poor performance in cross-subject scenarios. To overcome these limitations, the authors propose a Label-based Alignment Multi-Source Domain Adaptation (LA-MSDA) framework. The method operates in three stages. First, it extracts common domain-invariant features using a shared EEGNet-based network (C-EEGNet) and domain-specific features using multiple unshared CNN subnets (S-CNNs). Second, it employs a Local Label-based Maximum Mean Discrepancy (LLMMD) strategy to align the local feature distributions of relevant labels between source and target domains. This label-based alignment, rather than global feature alignment, helps eliminate the negative impact of individual differences by focusing on label-specific distributions. Third, the model utilizes a global optimization strategy that aligns the prediction probability distributions of target samples across all source domain classifiers. This addresses classifier confusion by setting similarity weight constraints based on prediction results, thereby improving generalization. The experimental validation involved 15 healthy subjects performing simulated driving tasks. EEG data was recorded from 61 channels and preprocessed using bandpass filtering and Independent Component Analysis. Features were extracted using Power Spectral Density (PSD) in the theta, alpha, and beta frequency bands. The study focused on two mental states: an awake state (TAV3 task, high workload) and a fatigue state (DROWS task, low workload/monotonous). The LA-MSDA model was trained on data from multiple source subjects and tested on target subjects to evaluate cross-subject generalization. The results demonstrate that LA-MSDA achieves remarkable performance in cross-subject EEG fatigue evaluation, outperforming existing methods by effectively handling individual differences and inconspicuous features. The label-based alignment and global optimization strategies significantly enhance the model's ability to generalize across subjects. These findings suggest that LA-MSDA has strong potential for practical applications in brain-computer interfaces, such as online monitoring of driver fatigue and the development of on-board safety systems.

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 qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

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

• drowsiness detection algorithms