Abstract
OBJECTIVE: Predicting driver fatigue degree is crucial for traffic safety. This study proposes a deep learning model utilizing electroencephalography (EEG) signals and multi-step temporal data to predict the next time-step fatigue degree indicator percentage of eyelid closure (PERCLOS) while exploring the impact of different EEG features on prediction performance. APPROACH: A CTL-ResFNet model integrating CNN, Transformer Encoder, LSTM, and residual connections is proposed. Its effectiveness is validated through two experimental paradigms, Leave-One-Out Cross-Validation (LOOCV) and pretraining-finetuning, with comparisons against baseline models. Additionally, the performance of four EEG features-differential entropy, α/β band power ratio, wavelet entropy, and Hurst exponent-is evaluated, using RMSE and MAE as metrics. MAIN RESULTS: The combined input of EEG and PERCLOS significantly outperforms using PERCLOS alone validated by LSTM, and CTL-ResFNet surpasses baseline models under both experimental paradigms. In LOOCV experiments, the α/β band power ratio performs best, whereas differential entropy excels in pretraining-finetuning. SIGNIFICANCE: This study presents a high-performance hybrid deep learning framework for predicting driver fatigue degree and reveals the applicability differences in EEG features across experimental paradigms, offering guidance for feature selection and model deployment in practical applications.