Robust analysis and spectral-based deep learning to detect driving fatigue from EEG signals

Quintero-Rincón, Antonio; Chaari, Lotfi; Batatia, Hadj · 2022 · Crossref

DOI: 10.1109/ictih57289.2022.10111943

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 critical safety issue of driver fatigue, a major contributor to traffic accidents, by proposing a novel, real-time detection method using electroencephalogram (EEG) signals. The research is motivated by the need for simple, fast, and reliable systems that can monitor drivers in real-time, particularly during monotonous driving conditions. The study aims to identify the single most relevant EEG channel for fatigue monitoring and develop a deep learning model capable of distinguishing between alert and fatigue states using data from that specific channel. The methodology employs a four-stage data science process: data acquisition, dimension reduction, data transformation, and data modeling. The dataset consists of 32-channel EEG signals from 12 male university students, recorded during both alert and fatigue states in a driving simulator. To identify the most significant electrode, the authors applied Maximum Covariance Determinant (MCD) robust analysis, a statistical technique resistant to outliers. This analysis identified the TP7 channel, located in the left tempo-parietal region associated with spatial awareness and visual-spatial navigation, as the most relevant for fatigue detection. For the predictive model, the time-series data from the TP7 channel was transformed into the spectral domain. Two spectral features were extracted for each signal segment: instantaneous frequency and instantaneous spectral entropy. These features formed a sequence input into a Bidirectional Long Short-Term Memory (biLSTM) neural network, which was designed to classify the EEG segments into alert or fatigue states. The experimental results demonstrated that the proposed biLSTM model achieved a classification accuracy of 75%, with a sensitivity of 67.7% and a specificity of 86%. A key finding was the model's prediction time-delay of only 1.4 seconds. When compared to classical machine learning techniques such as Decision Trees, Naive Bayes, Support Vector Machines, k-Nearest Neighbors, and Ensemble Boosted Trees, the biLSTM model exhibited lower accuracy metrics. However, the traditional methods suffered from significantly higher prediction delays, ranging from 5.82 to 19.80 seconds. The statistical significance of the results was confirmed with an ANOVA p-value of less than 0.01. The significance of this work lies in its demonstration that a single-channel, spectral-based deep learning approach can provide sufficiently fast detection for real-time applications. While the accuracy is lower than some multi-channel or complex feature-based methods, the minimal latency of the biLSTM model makes it more suitable for immediate driver warning systems. The identification of the TP7 channel as a robust indicator for fatigue provides valuable insight for the design of simplified, cost-effective wearable devices for monitoring driver state.

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