Adaptive Multi-feature Attention Neural Tree for EEG-based Fatigue Detection

Driver fatigue assessment is important to traffic safety, and the gold standard for fatigue detection is widely regarded as the EEG signals. Traditional machine learning methods fall short in accurately identifying EEG signals, while deep learning methods, though highly accurate, often lack model interpretability. This paper introduces an Adaptive Multi-feature Attention Neural Tree (AMANT) for fatigue signal recognition, leveraging the spatial structure of EEG signals. Initially, the EEG time series signal is converted into a two-dimensional matrix that preserves the information on channel positions and spatial details. Subsequently, a channel attention mechanism is applied, allowing the module to adaptively adjust the significance of each frequency, thereby enhancing the model's expressive capability. The AMANT integrates neural network modeling with decision tree technology, offering improved model interpretability. Experiments conducted on the SEED-VIG dataset achieved an average accuracy of 96.38%(DE) and 96.85%(PSD), with fewer model parameters and statistically significant results, thereby verifying the efficacy of the proposed approach. AMANT facilitates the automatic exploration of suitable network architectures, feature extraction, and end-to-end training while providing model interpretability. It effectively combines the strengths of traditional machine learning with deep learning, offering a robust solution for fatigue recognition.