Mechanism-oriented distributed crash reasoning for interpretable multiclass motorized rickshaw injury severity...

Motorized three-wheeler rickshaws are a major urban transport mode across South Asia, yet their injury-severity patterns remain insufficiently understood. Existing rickshaw safety studies rely largely on structured crash attributes, while narrative-based approaches often treat textual information as a single undifferentiated source. As a result, the distinct roles of contextual conditions, administratively reported crash causes, and crash-event descriptions are rarely preserved within a unified and interpretable prediction framework. To address this gap, this study proposes a mechanism-oriented Distributed Crash Reasoning (DCR) framework for interpretable multiclass injury-severity prediction. DCR preserves three functionally distinct evidence streams: structured contextual conditions, administratively reported crash-cause information, and crash-event descriptions. DCR processes these evidence streams through separate reported-cause, event-interaction, and structured-context modules and integrates them through attention-based multimodal fusion. The framework was evaluated using 2,742 motorized rickshaw crash records from Rawalpindi, Pakistan, covering 2017-2019. The proposed DCR achieved the highest recall, F1-score, and PR-AUC among the evaluated models. Controlled experiments further showed that multimodal structured-text configurations generally outperformed unimodal alternatives. Evidence-stream ablation and faithfulness analyses further identified structured context and crash-event information as the principal predictive sources, while reported-cause information provided a complementary contribution, particularly for Fatal-class discrimination. The multimodal advantage also persisted when models trained on 2017-2018 records were evaluated on 2019 data, indicating later-year predictive consistency within the available dataset. The statistical analysis further identified significant temporal (i.e., weekend, peak period, and nighttime), roadway (i.e., posted speed limit), vehicle (i.e., number of vehicles involved), and human (i.e., driver age and gender) factors, with their associations differing between Severe and Fatal injury outcomes. These findings demonstrate the value of preserving the semantic roles of heterogeneous crash information for class-imbalanced injury-severity prediction while providing diagnostically traceable evidence on the relative roles of different information streams in model predictions.