EEG-Based Cross-Subject Driver Drowsiness Recognition With an Interpretable Convolutional Neural Network

Cui, Jian; Lan, Zirui; Sourina, Olga; Müller-Wittig, Wolfgang · 2023 · Crossref

DOI: 10.1109/tnnls.2022.3147208

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 developing calibration-free, cross-subject driver drowsiness recognition systems using electroencephalogram (EEG) signals. While EEG offers high temporal resolution for monitoring brain dynamics, significant variability in signals across different subjects—due to factors like electrode displacement, head shape, and noise—hinders the creation of robust models. Furthermore, existing deep learning approaches often function as "black boxes," lacking interpretability regarding which specific signal features drive classification decisions. The authors propose an "InterpretableCNN," a novel convolutional neural network designed to recognize drowsiness across subjects while providing sample-wise explanations of its predictions. The study utilizes a public EEG dataset collected from 27 subjects performing a sustained-driving task in a virtual reality simulator. After preprocessing and filtering, the authors selected 11 subjects, extracting 3-second EEG samples prior to lane-departure events. Samples were labeled as "alert" or "drowsy" based on reaction times to these events. The proposed model employs a compact architecture using depthwise separable convolutions to process spatial-temporal sequences. Specifically, it uses pointwise convolutions to demix signals and reduce redundancy, followed by depthwise convolutions to extract temporal features. To enable interpretability, the authors developed a technique inspired by Class Activation Maps that traces discriminative activations back to the input signal, generating heatmaps that highlight regions contributing most to the classification. Experimental results demonstrate that the InterpretableCNN achieves an average accuracy of 78.35% on leave-one-out cross-subject validation. This performance surpasses conventional baseline methods, which ranged from 53.40% to 72.68%, and outperforms state-of-the-art deep learning models, which achieved between 71.75% and 75.19%. The interpretation analysis revealed that the model successfully learned biologically meaningful features, such as Alpha spindles, as strong indicators of drowsiness. By visualizing the heatmaps, the authors confirmed that the network focuses on relevant neurophysiological patterns rather than noise, and they used these insights to analyze misclassified samples. The significance of this work lies in its dual contribution to accuracy and interpretability in EEG-based mental state recognition. By integrating an interpretation technique directly into the model evaluation, the study moves beyond treating deep learning as a black box, offering a method to validate that models rely on valid physiological markers. This approach not only improves recognition accuracy across diverse subjects but also provides a framework for discovering common drowsiness-related patterns in complex EEG data, paving the way for more reliable and transparent driver monitoring 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
• drowsiness

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