Abstract

We present the Trauma THOMPSON dataset and benchmarks designed to advance artificial intelligence research for real-time decision support in emergency and austere medical environments. The dataset contains 220 unscripted egocentric videos of five emergency procedures, including a diverse collection of "just-in-time" (JIT) life-saving interventions performed under resource-constrained conditions. These JIT scenarios more closely reflect the realities of humanitarian and field-based operational medicine, where standard protocols must often be adapted or creatively executed. To support deeper visual understanding, we introduce two new layers of fine-grained annotations: object detection labels for critical medical instruments and supplies and hand annotations to facilitate hand tracking and surgical skill assessment. These additions enable new research directions in spatiotemporal reasoning, interaction modeling, and AI copilots that interpret and guide complex procedures in real time. The Trauma THOMPSON dataset includes benchmark tasks in action recognition, action anticipation, visual question answering (VQA), object detection, and hand localization. We evaluate state-of-the-art models across these tasks, identifying current strengths and open challenges in developing robust AI for field-deployable decision-making. The dataset is available at https://github.com/zhuoyp/TTD, and it can serve as a foundation for building intelligent systems that assist frontline caregivers.