Predictive Safety Analytics of Double-Lane Roundabouts: An Explainable AI Approach to Geometric Compatibility and Opportunistic Driving

While two-lane roundabouts increase operational capacity, their excessive lateral space often induces opportunistic driving and safety-critical events. This study presents a data-driven framework to evaluate the non-linear impact of geometric constraints and vehicle kinematics on operational adequacy. Naturalistic driving data (5622 observations) were collected at four urban two-lane roundabouts. To address severe class imbalance without altering empirical distributions, a Cost-Sensitive Random Forest classifier was implemented. The algorithmic logic was decoded using Permutation Feature Importance and SHapley Additive exPlanations (SHAP) to evaluate risk associations across predefined geometric ranges. Results identify the executed maneuver as the dominant predictor, making geometric criticality context-dependent. SHAP dependence analysis revealed that intermediate two-lane exit widths (6–8 m) and exit radii (10–30 m) neutralize traffic calming, significantly increasing inappropriate trajectories and inappropriate speeds. Furthermore, wide two-lane entry widths (>8 m) are positively associated with yielding violations, a pattern hypothesized to stem from attentional dispersion and opportunistic gap acceptance. Overall, the observed behavioral classifications highlight the operational vulnerabilities of uncanalized two-lane roundabouts. By quantifying the geometric incompatibility between permissive layouts and angular maneuvers, these behavioral findings provide empirical motivation to consider physically channelized designs (e.g., turbo-roundabouts) as a design countermeasure.