GraTED: Graph--Tensor EEG Decoding for Passenger Hazard Perception in Autonomous Vehicles

Yuxuan Gu, Yingkai Yang, Qiyu Rao, Tania Stathaki, Danilo MandicarXiv 2026 · arXiv

TL;DR

GraTED is a novel deep learning framework that decodes passenger brain activity (EEG) to predict road hazards in autonomous vehicles. By combining graph neural networks to respect the brain's physical geometry with low-rank tensor classification to preserve multidimensional data structure, it achieves state-of-the-art accuracy while using up to 87% fewer parameters than previous models.

What Problem Does This Paper Address?

As autonomous vehicles (AVs) become more common, a critical gap has emerged: a vehicle’s objective performance doesn't always align with a passenger's perceived risk. If a passenger feels unsafe, even if the car is technically fine, trust in the technology erodes.

Electroencephalography (EEG) offers a window into this "gut feeling," capturing anticipatory neural responses before a hazard occurs. However, decoding these signals is difficult. Traditional methods often force the irregular, 3D arrangement of scalp electrodes into artificial 2D grids (like a flat image) or flatten complex time-feature data into simple matrices. This "information destruction" obscures the natural spatial relationships of the brain and ignores higher-order interactions between time and neural features, leading to models that are either inaccurate or overly complex and prone to overfitting.

Methods

The authors propose GraTED (Graph-Tensor EEG Decoding Model), which treats EEG data with the structural respect it deserves through two key innovations:

  1. Electrode Graph Construction: Instead of a grid, GraTED models electrodes as nodes in an undirected graph based on their actual 3D scalp coordinates. It uses Chebyshev Graph Convolutions to propagate information between neighboring electrodes, capturing spatial dependencies without the computational burden of explicit eigendecomposition.
  2. Tucker-Parameterized Tensor Classification: Rather than flattening the learned features into a long vector for a final decision, GraTED keeps them in their multidimensional "tensor" form. It uses Tucker decomposition—a method for low-rank tensor approximation—to model interactions across time, features, and classes simultaneously. This acts as a powerful "inductive bias," forcing the model to learn a compact, structured representation that is highly resistant to overfitting, even with limited data.
Fig. 1. The GraTED architecture, consisting of risk prediction (RP) and danger identification (DI) branches.

The model features two branches: Risk Prediction (RP), which looks at the window before an event to assess potential risk, and Danger Identification (DI), which detects actual danger around the moment an event occurs.

Datasets

The model was evaluated on the PassengerEEG dataset. This benchmark includes:

  • 11 participants recorded over 45 hours.
  • EEG data aligned to 15 different simulated autonomous driving traffic scenarios.
  • Trials categorized into three levels: safe, low-risk, and high-risk.

Key Insights and Ideas

  • Geometry is a Feature: Respecting the physical 10–20 scalp geometry via graphs is significantly more effective than forcing data into a rectangular grid. It allows the model to learn more physiologically plausible spatial filters.
  • The Power of "Not Flattening": By avoiding the "flattening" of tensors, the model natively captures cross-mode dependencies—for example, how a specific neural frequency at a specific millisecond relates to a "risky" label.
  • Efficiency as Regularization: The use of Tucker decomposition isn't just about saving memory; it’s a form of structural regularization. By drastically reducing the number of trainable parameters, the model is forced to find the most essential patterns in the noisy EEG signal.

Key Results

GraTED consistently outperformed seven state-of-the-art baselines (including EEGNet and 3D-CRNN) across all evaluation settings.

  • Accuracy: It achieved the highest Balanced Accuracy (BA) in Risk Prediction (0.946), Danger Identification (0.827), and DI with Risk-aware Sequential Labeling (0.848).
  • Parameter Efficiency: This is the most striking result. GraTED achieved these scores while using significantly fewer parameters. Compared to the previous best-performing model (3D-CRNN), GraTED reduced the parameter count by 80.9% for RP and 87.3% for DI.
Table 1. Performance and model-size comparison (mean ± standard deviation across 11 participants). Params. denotes the trainable parameters for each setting. The best results in each metric are bold.

Limitations and Future Work

While GraTED shows remarkable efficiency and accuracy, the authors note a few areas for growth:

  • Subject-Independence: The current results are averaged across participants, but the model's ability to generalize to a completely new user (zero-shot) wasn't the primary focus.
  • Real-World Noise: The data comes from simulated driving. Real-world environments introduce motion artifacts and electrical noise that may require more robust preprocessing.
  • Online Deployment: Future work will involve testing GraTED in real-time, closed-loop systems to see if its predictions can actively adjust AV behavior to improve passenger comfort.

Summary updated

Abstract

Passenger electroencephalography (EEG) can reveal road risk through anticipatory responses before hazardous events and neural reactions around their onset. However, existing EEG decoding methods often map physiologically meaningful arrangement of electrodes onto artificial 2D grids and thus flatten the higher-order time-feature representations into matrices, obscuring meaningful structure in the data. To preserve and exploit the spatial and time-feature structure of EEG recordings, we propose the Graph-Tensor EEG Decoding Model (GraTED) for passenger risk prediction (RP) and danger identification (DI) in autonomous-vehicle scenarios. GraTED models the electrode arrangement as an undirected graph based on their scalp locations, using graph convolution to capture spatial dependencies and low-rank tensor classification to preserve the multidimensional time-feature structure, while controlling model complexity. Evaluated on the PassengerEEG dataset, comprising 11 participants and 45 hours of recordings, GraTED achieves the highest balanced accuracy, compared to the state-of-the-art, across all three evaluation settings, while requiring the fewest learnable parameters among seven baseline methods.
BibTeX
@misc{gu2026grated,
  author = {Yuxuan Gu and Yingkai Yang and Qiyu Rao and Tania Stathaki and Danilo Mandic},
  title = {{GraTED: Graph--Tensor EEG Decoding for Passenger Hazard Perception in Autonomous Vehicles}},
  year = {2026},
  eprint = {2611.02827},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.02827},
}
Figure