Deng H., Lin S., Fan C., Wu X., Wang K., Peng Y., Yannis G., “Comprehensive Assessment of Driver Takeover Readiness via a Multimodal Fusion Multi-Task Learning Model”, Advanced Engineering Informatics, 76, (2026) 105094.

Effective assessment of driver takeover readiness is essential for improving takeover performance and ensuring takeover safety. However, existing takeover readiness assessment models have the following limitations. First, they generally produce only a single-dimensional or overall readiness score, lacking the multi-dimensional and interpretable outputs required for practical application. Second, their input data sources are relatively limited, and they fail to effectively utilise the relationships among different readiness dimensions, which ultimately undermines model accuracy. To address these issues, this study constructs a multimodal takeover readiness dataset that incorporates visual, physiological, gaze, and demographic features. The dataset was collected from 32 participants using a within-subject design (7 non-driving related tasks × 3 takeover scenarios), yielding a total of 20,160 s of synchronized multimodal time-series data. Post-trial self-reported questionnaire scores were used as the ground-truth labels for takeover readiness. Based on this dataset, a multimodal fusion multi-task learning model is proposed to jointly assess 5 dimensions of takeover readiness, including cognitive readiness, postural readiness, takeover request perception readiness, situation awareness readiness, and overall readiness. The experimental results demonstrate that the proposed model effectively fuses multimodal features and captures task-specific representations, achieving an average Accuracy of 79.0%, Balanced Accuracy of 78.5%, and Macro￾F1 score of 78.2% across five 5-class classification tasks, outperforming the selected baselines. Ablation studies on both model architecture and input feature categories confirmed the rationale of the model design and emphasise the benefits of using multimodal data fusion and multi-task learning to model takeover readiness states. These findings provide insights for the development of more precise and adaptive intervention strategies in future automated driving systems.