Personalizing In-Vehicle warnings: A causal machine learning approach to optimizing workload and risk perception

The effectiveness of in-vehicle Connected Information (CI) is often limited by uniform warning strategies that overlook the interaction among warning design, traffic context, and driver state. This study establishes a causal machine learning framework to quantify how CI modality and lead time influence driver workload (WL) and perceived risk (PR) across multiple traffic scenes. A 3 × 3 × 5 factorial driving-simulator experiment collected multimodal physiological, behavioral, and self-reported data from 52 drivers. Double Machine Learning with Causal Forests (DML-CF) was applied to identify both average and heterogeneous causal effects. The results show strong context dependence and clear nonlinear patterns in CI effectiveness. Moderately early warnings in the range of 5.5 to 6.5 s combined with congruent dual-modality cues consistently reduced WL and PR. Delayed or single-modality cues frequently increased cognitive demand and perceived hazard. Conditional treatment effect (CATE) analyses revealed substantial heterogeneity and identified driver subgroups with distinct physiological and behavioral characteristics. For instance, certain individuals exhibited elevated WL due to heightened arousal even when exposed to multimodal early warnings that were beneficial for most drivers. These findings indicate that optimal CI design requires continuous alignment between warning parameters, situational uncertainty, and the driver's state. The evidence provides a quantitative causal basis for developing adaptive CI systems capable of tailoring information delivery to enhance safety and human-machine interaction.