Real-Time Driving Risk Assessment Using LSTM-attention and XGBoost with Physiological and Driving Behavior Data

Early identification of driving risks is crucial for driving safety. This study presents a deep learning method based on Long Short-term Memory (LSTM) with attention mechanism and XGBoost to recognize risk levels during driving simulation and to provide early warnings to drivers. In this study, we collected multimodal data including driving data, eyes status, and basic physiological data from driving simulators, eye trackers, and smartwatches for comprehensive analysis. By combining driving behavior and physiological features in each time window, the ability of LSTM networks was applied to analyze temporal features for prediction of the risk levels (Level 0, Level 1, and Level 2) in the next 20 seconds. Meanwhile, an attention mechanism was introduced to understand the importance of multiple features as well. Then the output of LSTM-attention was further refined using XGBoost, which demonstrated excellent performance in classification tasks. The integrated model showed that the overall accuracy achieved is above 89.97%, with an F1-score of 89.15%.