InstanceEasyTL: An Improved Transfer-Learning Method for EEG-Based Cross-Subject Fatigue Detection

Zeng, Hong; Zhang, Jiaming; Zakaria, Wael; Babiloni, Fabio; Gianluca, Borghini; Li, Xiufeng; Kong, Wanzeng · 2020 · Crossref

DOI: 10.3390/s20247251

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 driver fatigue. While EEG is an effective indicator of mental state, significant inter-subject variability in signals and the difficulty of collecting sufficient labeled data for each individual hinder the deployment of robust detection systems. The authors propose InstanceEasyTL, an improved transfer-learning method derived from the EasyTL model, which was previously successful in image recognition but struggled with the distributional differences inherent in cross-subject EEG data. The goal is to create a classifier that requires less subject-specific training data while maintaining high accuracy and robustness across different drivers. The study utilized EEG data from 15 subjects participating in a driving simulation experiment approved by the University of Rome “La Sapienza.” The protocol involved a two-hour session with distinct stages: a baseline warm-up, performance tasks, alert/vigilance tasks with varying stimulus frequencies, and a drowsiness stage. EEG signals were recorded using a 61-channel system at 200 Hz, preprocessed with band-pass filtering (1–30 Hz) and independent component analysis to remove artifacts. Power spectrum density features were extracted from theta, alpha, and beta frequency bands, resulting in 1647-dimensional feature vectors. The InstanceEasyTL method modifies the original EasyTL approach by incorporating a weighted alignment strategy. It splits the target domain into two parts: one is merged with the source domain to form an expanded training set, while the other serves as the test set. This allows the model to adapt to the specific distribution of the target subject’s EEG signals through iterative weight updates based on classification errors. Experimental results demonstrate that InstanceEasyTL outperforms several baseline methods, including Support Vector Machines (SVM), Transfer Component Analysis (TCA), Geodesic Flow Kernel (GFK), Domain-adversarial Neural Networks (DANN), and the original EasyTL. In a leave-one-subject-out cross-validation scheme, InstanceEasyTL achieved the highest classification accuracy in distinguishing between alert (TAV3) and drowsy (DROW) states across subjects. The method proved more robust and required less labeled data from the target subject compared to existing models. The findings indicate that by strategically borrowing and weighting samples from the target domain during training, InstanceEasyTL effectively mitigates the domain shift problem caused by inter-subject variability. This approach offers a promising solution for developing efficient, accurate, and scalable EEG-based fatigue detection systems that can be deployed with minimal calibration for new users.

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