Analyzing SHAP values of XGBoost algorithms to understand driving features affecting take-over time from vehicle alert to driver action

• Novel application of XGBoost with SHAP reveals multivariate dependencies shaping TOT. • SAE Level 3 automation prolongs responses, requiring extended time budgets. • Mild deceleration extends TOT, while stronger deceleration triggers faster reactions. • Obstacle effects differ by environment, with highway scenarios prompting quicker takeovers. • Advanced HMI seems to improve awareness but slightly increases reaction time. This study investigates the factors influencing Take-Over Time (TOT) during transitions from automated to manual driving, emphasizing the novelty of applying XGBoost modeling combined with SHAP analysis to uncover non-linear and implicit dependencies between features. Using high-frequency data from a driving simulator, key variables such as automation level, driving measurements, different types of obstacles, and Human-Machine Interface (HMI) conditions were analyzed to understand their effects on TOT. The XGBoost model was optimized using a cross-validation approach, achieving strong predictive performance (R 2 = 0.871 for testing set). Feature importance analysis revealed that Automated Driving (AD) level 2 or 3 was the most influential factor, underscoring how extended time budgets and reduced driver engagement interact in shaping TOT. Higher automation levels resulted in longer TOT, with SHAP values consistently positive for AD Level 3, demonstrating the added value of explainable machine learning in clarifying these patterns. Dynamic driving parameters, such as deceleration and speed variability, were also significant. Strong negative deceleration values were generally associated with shorter TOT, reflecting quicker responses under urgent braking. Speed showed a moderate positive effect on TOT at 80–110 km/h, with drivers taking additional time to assess the environment, but higher speeds (above 110 km/h) resulted in quicker responses. Beyond these established effects, SHAP analysis revealed how automation level, obstacle environment, and HMI design jointly condition driver responses. The HADRIAN HMI, slightly increasing TOT compared to the baseline, simultaneously seems to demonstrate potential safety benefits through tailored guidance and improved situational awareness. By combining methodological innovation with contextual insights, this study contributes to a deeper understanding of takeover behavior and provides actionable evidence for optimizing adaptive HMI design and takeover strategies in AD systems.