<p>Traditional fatigue monitoring methods often suffer from in-sufficient real-time performance and low accuracy. To address these limitations, this study proposes a novel facial feature-based approach for driver fatigue monitoring and prediction, aiming to enhance both detection accuracy and real-time responsiveness. Eye features are one of the key indicators for assessing driver fatigue. This research first introduces a new quantitative metric based on eye features to improve the stability and robustness of fatigue characterization. Subsequently, a pre-trained facial keypoint detection model is employed to extract dynamic eye and mouth features, constructing a novel fatigue index that dy-namically reflects real-time changes in facial behavior for more accurate fatigue state assessment. To further improve prediction accuracy, a lightweight CNN-BiLSTM-Attention hybrid model is designed. This model integrates spatial features extracted by convolutional neural networks (CNN) with temporal dependencies captured by a bidirectional long short-term memory (BiLSTM) network, while an attention mechanism is introduced to optimize fatigue-level prediction capabilities. To validate the proposed method, experiments were conducted with six drivers, each lasting no less than 20 min. The results demonstrate that the new fatigue index and prediction model can accurately capture dynamic changes in driver fatigue levels with low prediction error rates. Finally, we also verified the effectiveness of the fatigue index through validation in a public dataset.</p>