An EEG Channel Selection Framework for Driver Drowsiness Detection via Interpretability Guidance

Zhou, Xinliang; Lin, Dan; Jia, Ziyu; Xiao, Jiaping; Liu, Chenyu; Zhai, Liming; Liu, Yang · 2023 · Crossref

DOI: 10.1109/embc40787.2023.10341126

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 improving driver drowsiness detection using electroencephalogram (EEG) signals. While EEG accurately reflects mental fatigue, raw data is high-dimensional, noisy, and contains redundant information. Existing deep neural network (DNN) approaches typically utilize either single-channel data, which lacks sufficient information, or full-head channel data, which introduces noise and bias, resulting in limited detection performance. To resolve this, the authors propose an Interpretability-guided Channel Selection (ICS) framework that selectively identifies and utilizes the most informative EEG channels for drowsiness classification. The ICS framework employs a two-stage training strategy involving a teacher network and a student network. In the first stage, a teacher network is trained on full-head EEG data. The authors then apply Class Activation Mapping (CAM), an interpretability technique adapted for time-series EEG data, to visualize the contribution of each channel to the classification decision. A voting scheme is subsequently used to select the top N contributing channels based on high-confidence predictions from the training set. In the second stage, a student network is trained exclusively on these selected channels to serve as the final detection model. This coarse-to-fine approach aims to refine feature learning by removing irrelevant or noisy channels. Experiments were conducted on a public dataset derived from 27 subjects performing a virtual reality driving task, using a mini-dataset of 11 subjects for evaluation. The study utilized a leave-one-subject-out cross-validation protocol to assess generalizability. The ablation study determined that selecting the top 10 channels yielded optimal performance; fewer channels provided insufficient information, while more channels reintroduced noise. When applied to four baseline CNN models (EEGNet 4,2, EEGNet 8,2, ShallowConvNet, and InterpretableCNN), the ICS framework significantly improved mean detection accuracy. For instance, EEGNet 4,2 accuracy increased from 68.31% to 72.96%, and InterpretableCNN improved from 78.26% to 81.38%. Additionally, the framework reduced the standard deviation of accuracy across subjects, indicating enhanced model stability and robustness. The significance of this work lies in demonstrating that interpretability-guided channel selection can effectively enhance cross-subject drowsiness detection performance. The ICS framework outperformed other typical channel selection methods, such as hybrid optimization algorithms and Pearson Correlation Coefficient-based selection, achieving an average accuracy improvement of approximately 4.0%. The authors conclude that ICS is a universally applicable method that improves model robustness by focusing on informative features. However, they note that the current two-stage process is not end-to-end and is time-consuming, suggesting future work should focus on unified optimization and application to other EEG-based scenarios like epilepsy detection.

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

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: physiological data