An EEG-Based Transfer Learning Method for Cross-Subject Fatigue Mental State Prediction

Zeng, Hong; Li, Xiufeng; Borghini, Gianluca; Zhao, Yue; Aricò, Pietro; Di Flumeri, Gianluca; Sciaraffa, Nicolina; Zakaria, Wael; Kong, Wanzeng; Babiloni, Fabio · 2021 · Crossref

DOI: 10.3390/s21072369

archive: archived pipeline: cataloged verified

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 fatigue detection using electroencephalogram (EEG) signals, a critical issue for reducing traffic accidents caused by fatigued driving. While EEG is an effective objective measure of mental state, significant inter-subject variability in signal distribution hinders the transfer of classification models from one subject to another. Existing Domain-Adversarial Neural Network (DANN) approaches, successful in other fields, struggle with EEG data due to severe class imbalance between source and target domains and the risk of "negative transfer" when source data distributions differ significantly from the target. To overcome these limitations, the authors propose Generative-DANN (GDANN), a novel model that integrates Generative Adversarial Networks (GAN) with DANN to align data distributions and balance datasets across subjects. The study involved 13 healthy volunteers participating in a driving simulation experiment designed to induce varying levels of mental fatigue. Participants performed a primary driving task alongside a secondary Attentional and Vigilance Task (TAV) with five difficulty levels, followed by a monotonous driving phase to induce drowsiness. EEG data were recorded from 61 channels at 200 Hz, preprocessed using band-pass filtering and Independent Component Analysis to remove artifacts, and segmented into 1-second windows. Power Spectral Density (PSD) features were extracted from the 0.1–30 Hz frequency range, resulting in 1830-dimensional feature vectors per sample. The experimental design utilized a leave-one-subject-out cross-validation strategy, where data from 12 subjects formed the source domain for training, and data from the remaining subject served as the unlabeled target domain for testing. The GDANN architecture improves upon standard DANN by incorporating a GAN component to generate synthetic target-domain data, thereby balancing the dataset and mitigating distribution mismatches. The model employs LeakyReLU activation functions, Tanh output layers, and adaptive optimization algorithms (Adam for GAN components, SGD for DANN components) to enhance convergence and performance. The training process involves a dual adversarial game: one between the GAN generator and discriminator to produce realistic target-domain samples, and another between the DANN feature extractor and domain classifier to learn domain-invariant features. This approach allows the model to select source subjects with distributions closest to the target, reducing negative transfer. Experimental results demonstrate that GDANN significantly outperforms traditional classification models in cross-subject fatigue detection. The proposed method achieved an average accuracy of 91.63% in distinguishing between fatigue and alert states across subjects. This performance improvement highlights the effectiveness of combining GANs with domain adaptation techniques to handle the inherent variability and imbalance in EEG data. The findings suggest that GDANN offers a robust solution for practical brain-computer interaction applications, particularly in real-time driver monitoring systems where individual calibration is impractical.

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