Adelaide University CENTRE FOR AUTOMOTIVE SAFETY RESEARCH

Adelaide University   /   Centre For Automotive Safety Research   /   Publications   /   List   /   Details

Publication Details

TitleMMDL-Net: A Multimodal Deep Learning Network for Post-Collision Injury Prediction and Triage Support
AuthorsZou Z, Elsegood ME, Ponte G, Doecke SD, Majumdar A, Baker C
Year2026
TypeJournal Article
AbstractRoad traffic collisions can cause severe and time-critical injuries, making rapid and reliable post-crash triage essential for improving patient outcomes. With advances in vehicle technology, data capture systems, and policy frameworks, increasingly rich crash-related data are becoming available to support early injury assessment. However, current dispatch and triage decisions still often rely on limited real-time information. This study proposes a compact multimodal deep learning framework for predicting occupant-level, vehicle-level, and body-region-specific injury risk, integrating static crash descriptors, irregularly sampled event data recorder timelines, and post-collision images. Modality-specific encoders capture complementary information from each data source, and a lightweight fusion module then combines them. The approach is designed for small, information-rich datasets and near-real-time inference. The framework jointly predicts occupant injury severity, vehicle severity, and body-region-specific risk of serious injury. This approach is evaluated using the CASR-EDR dataset, a real-world Australian crash dataset maintained by the Centre for Automotive Safety Research (CASR) at Adelaide University. The dataset comprises over 1,060 collisions involving more than 3,100 occupants collected since 2017, linking detailed vehicle telemetry with police reports and verified hospital injury outcomes. The multimodal model demonstrates consistent improvements over representative classical and learning-based baselines in terms of both discrimination and calibration, while maintaining computational efficiency compatible with dispatch-time deployment. These results indicate that combining vehicle sensor data with visual evidence can enhance early estimates of injury risk and support automated post-crash triage.
Journal Title(May 09, 2026). Available at SSRN: https://ssrn.com/abstract=6739618 or
Journal Volume (Issue)http://dx.doi.org/10.2139/ssrn.6739618
NotesAvailable at SSRN https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6739618

Reference
Zou Z, Elsegood ME, Ponte G, Doecke SD, Majumdar A, Baker C (2026). MMDL-Net: A Multimodal Deep Learning Network for Post-Collision Injury Prediction and Triage Support. (May 09, 2026). Available at SSRN: https://ssrn.com/abstract=6739618 or, http://dx.doi.org/10.2139/ssrn.6739618.