EgoEMS

EgoEMS 由弗吉尼亚大学等机构联合急救医学专家构建，是首个面向院前急救场景的端到端高保真多模态第一视角（egocentric）数据集，用于支持智能认知助手在紧急医疗救治中的感知与决策研究。数据集覆盖 233 个标准化模拟急救场景、超过 20 小时的程序化操作视频，涉及 62 名参与者，其中包含 46 名专业急救人员，累计标注 2694 个关键步骤实例。EgoEMS 采用低成本、可复现的开源采集系统，从多名施救者的头戴视角同步采集视频、语音及传感器信号，并在符合国家规范的流程设计下，由专家完成关键步骤标注、含说话人分离的时间戳语音转录、操作质量指标以及关键急救物体的框选与分割掩膜标注，全面刻画真实急救互动过程。该数据集主要服务于急救场景下的多模态关键步骤识别、步骤分割与操作质量评估等任务，可用于训练和评测面向 EMS 的 AI 认知助手、可穿戴辅助手段与教学评估系统，为缓解一线急救人员的认知负荷、提升决策准确性与患者预后提供数据基础。

EgoEMS, developed in collaboration with emergency medicine experts by multiple institutions including the University of Virginia, is the first end-to-end high-fidelity multimodal egocentric dataset tailored for pre-hospital emergency care scenarios, designed to support research on perception and decision-making for intelligent cognitive assistants in emergency medical treatment. The dataset encompasses 233 standardized simulated emergency care scenarios, over 20 hours of procedural operation videos, and involves 62 participants, among whom 46 are professional emergency medical personnel, with a total of 2694 annotated key step instances. Adopting a low-cost, reproducible open-source data acquisition system, EgoEMS synchronously collects video, audio, and sensor signals from the first-person head-mounted perspectives of multiple rescuers. Under workflow designs compliant with national regulations, experts perform comprehensive annotations including key step labeling, timestamped audio transcription with speaker diarization, operational quality metrics, as well as bounding box and segmentation mask annotations for key emergency medical objects, which fully characterizes real-world emergency care interaction processes. This dataset primarily serves tasks such as multimodal key step recognition, step segmentation and operational quality assessment in emergency care scenarios. It can be utilized to train and evaluate AI cognitive assistants, wearable assistance tools and teaching evaluation systems for EMS, providing a foundational data resource for alleviating the cognitive load of frontline emergency personnel, enhancing decision-making accuracy and improving patient prognosis.

原始信息汇总

EgoEMS: 高保真多模态第一人称紧急医疗服务认知辅助数据集

数据集概述

高保真第一人称多模态EMS数据集，带有丰富标注，旨在推进认知辅助研究。

核心数据统计

• 时长：20+小时同步多模态录制
• 场景：233个模拟紧急场景
• 参与者：62名参与者（包括46名EMS专业人员）

数据模态与标注

• 模态：第一人称视频、音频、IMU
• 标注内容：关键步骤、分段转录文本、CPR指标、边界框和分割掩码

应用基准

• 关键步骤分类：从多模态输入分类程序步骤
• 关键步骤分割：检测步骤间的时间转换
• CPR质量评估：从IMU和视频估计速率和深度

数据采集系统

开源、低成本、可复现的同步第一人称多模态数据采集系统

数据获取

完整数据集将公开托管，发布时链接将在此处显示

相关资源

• 论文：AAAI 2026
• 预印本：arXiv
• 代码：可用
• 视频：可用

搜集汇总

背景与挑战

背景概述

EgoEMS是一个高保真多模态数据集，包含20+小时的同步EMS场景记录和丰富注释，旨在支持AI认知辅助研究。数据集涵盖233个模拟紧急场景和62名参与者(含46名EMS专业人员)，提供关键步骤识别和CPR质量评估等基准任务。