Driver distraction caused by the use of in-vehicle intelligent devices poses a serious threat to road safety. However, existing near-collision prediction methods have not fully accounted for drivers’ distraction states, resulting in limited predictive accuracy. Using the VTTI 100-Car naturalistic driving database, this study develops a risk identification framework that integrates vehicle kinematic indicators with drivers’ visual attention (AttenD) features. Kinematic and AttenD metrics were extracted separately, and both XGBoost and gated recurrent unit (GRU) models were trained to evaluate system performance within a 2–4 s warning horizon. In addition, the influence of including distraction-related variables in model inputs was examined. The results show that incorporating AttenD features substantially improves model performance—particularly in short-term warning scenarios (≤ 2 s), where the GRU model achieved an accuracy exceeding 92% and a recall above 95.24%, outperforming traditional approaches. The findings also indicate that visual attention features continue to enhance discriminative power even at a 4 s warning lead time, although overall model performance declines as the warning horizon increases. This study provides an effective methodology and empirical support for enhancing active safety system warning capabilities through the integration of drivers’ physiological and behavioral data.