TRIDENT: A multi-task, triple-branch deep learning framework for EEG-based recognition, severity estimation, and future high-anger prediction in an on-road Wizard-of-Oz paradigm.

Driving anger is strongly associated with aggressive driving and elevated crash risk. However, continuous modeling of graded anger-related affective states from physiological signals under controlled on-road exposure remains underexplored. In this study, we investigated driving-anger-related affective states elicited in an on-road Wizard-of-Oz (WoZ) paradigm, in which 24 licensed participants sat in the front passenger seat, imagined themselves as the driver, and were exposed to scripted traffic-conflict maneuvers executed by a trained safety driver. Using synchronized 32-channel electroencephalography (EEG), self-reports, and contextual information, we constructed a four-level anger-state dataset. Building on this dataset, we introduce TRIDENT, a multi-task deep learning model designed to simultaneously recognize four levels of anger states, estimate continuous anger severity (0-100), and predict future high-anger states. TRIDENT integrates multi-scale temporal convolutions, brain-network representations, and sequence modeling to capture complementary spatiotemporal patterns of anger dynamics. Experimental results show that TRIDENT significantly outperforms representative baseline EEG emotion recognition models, achieving up to 85% accuracy in four-class anger-state classification and 87% accuracy in predicting future high-anger states. Scalp topographies and cortical source localization analyses further reveal anger-level-dependent changes in prefrontal, temporal, and limbic brain networks. These findings provide a physiologically grounded perspective on anger-related neural dynamics under controlled on-road conflict exposure and have implications for emotion-aware in-vehicle interfaces and personalized intervention design. Code will be made available upon acceptance at: https://github.com/tianyaz719/TRIDENT.