EEG-based Prediction of Driver Takeover Performance

Huang, Jinhui; Zhang, Tingru · 2024 · Crossref

DOI: 10.54941/ahfe1005233

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the critical safety challenge of driver takeover performance in conditional autonomous driving (SAE Level 3). While existing research has examined factors influencing takeover quality, there is a lack of predictive models capable of assessing takeover readiness before a Takeover Request (TOR) is issued. Specifically, the authors aim to determine whether Electroencephalogram (EEG) signals can predict driver performance and to identify the most effective machine learning algorithms for this task. The motivation stems from the need to enhance safety by anticipating poor takeover outcomes when drivers are engaged in non-driving related tasks (NDRTs) and may have reduced attention. To investigate this, the researchers conducted a driving simulation experiment with 72 licensed participants (36 male, 36 female, aged 18–28). Participants performed a typing task on a tablet while the vehicle operated in autonomous mode at 100 km/h. Eight TOR scenarios were triggered by hazards such as broken-down vehicles or pedestrians, requiring drivers to resume manual control. EEG data were recorded using a 32-channel system at 500Hz. After excluding three participants due to technical issues, data from 69 subjects were analyzed. The EEG signals were preprocessed using band-pass filtering, artifact removal via Independent Component Analysis, and downsampling to 100Hz. Features were extracted from a 20-second window preceding the TOR, including power spectral density and amplitude for δ, θ, α, and β frequency bands across 30 channels, resulting in 480 features after clustering binning. Takeover performance was labeled as high, medium, or low based on video ratings using the Take-Over Controllability (TOC) scale. The study employed five machine learning models—Decision Trees (DT), Support Vector Machine (SVM), Random Forest (RF), Light Gradient Boosting Machine (LightGBM), and Multi-layer Perceptron (MLP)—trained with five-fold cross-validation. Feature selection using Logistic Regression and Recursive Feature Elimination reduced the dataset to 130 key features. The LightGBM model demonstrated superior performance, achieving an accuracy of 84.2% and an F1 score of 83.0%, significantly outperforming the other models, particularly DT, which had the lowest accuracy (59.4%). Feature importance analysis revealed that β and α band features were most predictive, with the β band signals from the T8 channel (right temporal lobe) being the most significant indicators, likely reflecting cognitive processing speed. The findings confirm that EEG-based machine learning models can effectively predict driver takeover performance prior to a TOR. The high accuracy of the LightGBM model suggests its potential for integration into autonomous driving systems to monitor driver readiness and enhance safety. The identification of specific EEG bands and channels provides insights into the neural correlates of takeover readiness, offering a foundation for future developments in driver assistance systems and human-machine interaction in automated vehicles.

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
• takeover transitions

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, behavioral performance data
• Theoretical Contribution: computational model