Published November 13, 2025 | Version 1.0.0
Dataset Open

AURA: Synthetic ICU Video Dataset for Augmented Unplanned Removal Alert System

Authors/Creators

  • 1. Samsung Advanced Institute for Health Sciences & Technology, Sungkyunkwan University
  • 2. ROR icon Samsung Medical Center

Description

 

Dataset Description

The Augmented Unplanned Removal Alert (AURA) dataset consists of 75 synthetic intensive care unit (ICU) videos depicting patients with airway tubes. This dataset was purposefully generated to develop and validate the vision-based AURA system. It comprises 12 videos for the tuning set, and the 63 test-set videos were manually labeled for validation by nine ICU nurses.

A clinical expert carefully designed the data pipeline, including prompt refinement (adding clinically relevant details to the prompt) and quality control (excluding videos containing cinematic effects or physically implausible patient movements). Using the prompt template, we systematically generated diverse ICU scenarios varying in gendered and racial appearance, patient behavior, camera angle, and medical staff interaction. No real-world materials were used to generate the videos.

Technical Specifications

  • Total videos: 75 (12 tuning + 63 test)
  • Duration: 6 seconds (per video)
  • Frame rate: 25 fps
  • Resolution: 1280×720 pixels
  • Generation model: Hailuo AI T2V-01 Director

All videos were generated under a paid subscription plan and downloaded watermark-free in accordance with the service’s usage terms.

 

Annotations

Nine ICU nurses independently evaluated each of the 63 test videos and all annotations are provided at the video level.

Video Quality Assessment

Each video was rated on four realism dimensions: visual tone, background setting, patient behavior, and clinical plausibility. All dimensions achieved mean ratings ≥ 4.3 on a 5-point Likert scale, with patient behavior and clinical plausibility showing the highest ratings at 4.6.

Alarm Appropriateness & Ground Truth Labels

Each video was annotated for two alarm types:

  • Agitation alarm: Restless, non-purposeful movements indicating increased unplanned extubation (UE) risk
  • Collision alarm: Direct hand movement toward the airway tube requiring immediate intervention

Evaluators rated alarm appropriateness on a 5-point Likert scale and categorized inappropriate alarms as false, missed, premature, or delayed. Inter-rater reliability was near-perfect, with ICC(3,k) = 0.95 for both alarm types, indicating excellent agreement among raters.

See more details in our paper: arXiv

Data Files

video_metadata.csv

Contains descriptive attributes for all 75 videos (tuning + test sets).

Columns:

  • video_id: Unique identifier (tuning_01~12, test_01~63)
  • gendered_appearance: Visual gendered appearance (male, female)
  • racial_appearance: Visual racial appearance (white, black, asian, hispanic/latino)
  • cam_angle: Camera viewing angle (Vertical, Horizontal, Diagonal)
  • full_bodypart: Full body visible in frame (TRUE, FALSE)
  • behavior: Patient behavior type (Calm, Lifting hands, Restless, Sitting up, Lifting legs, Getting out of bed)
  • behavior_calm: Whether behavior is calm (TRUE, FALSE)
  • medical_staff: Medical staff presence (FALSE, Interaction, Presence)

expert_annotations.csv

Contains expert evaluation ratings for the 63 test videos only.

Columns:

Video Quality Assessment (5-point Likert scale, mean):

  • visual_tone: Visual tone similarity to clinical setting
  • background_setting: Background similarity to clinical setting
  • patient_behavior: Naturalness of patient behavior
  • clinical_plausibility: Clinical plausibility of the event

Alarm Appropriateness (5-point Likert scale, mean):

  • appropriateness_agitation: Appropriateness of agitation alarm
  • appropriateness_collision: Appropriateness of collision alarm

Ground Truth Labels (consensus-based):

  • consensus_collision: Consensus label for collision alarm (TRUE, FALSE)
  • consensus_agitation: Consensus label for agitation alarm (TRUE, FALSE)
 

 

Prompt Template

The following template was used to systematically generate diverse ICU scenarios:

[Background]

The room contains all characteristic ICU features: a wall-mounted medical gas panel with oxygen, compressed air, and suction ports; a multi-parameter bedside monitor showing the patient's ECG, heart rate, respiratory rate, SpO₂, and blood pressure; an advanced ventilator machine with detailed circuit connections leading to the patient's endotracheal tube; multiple infusion and syringe pumps delivering medications and fluids; and a mobile equipment cart. The Racial appearance Gendered appearance patient is sedated in a fully motorized ICU bed with raised side rails. They are intubated with an endotracheal tube secured with medical tape, connected to the ventilator. Their arms have IV lines, and one wrist has an ID band. The lighting is dim, with clinical overhead lights casting a focused glow. The environment includes soft beeping sounds from the monitor and ventilator, typical alarm indicators, and clear visual signs of a critical care setting.

[View]

Fixed overhead cam_angle view from above the bed: A static, wide-angle overhead shot positioned directly above the foot end of the bed, looking downward toward the patient and capturing the entire bed, patient, and surrounding equipment within the frame (not angled from the head side).

[Subject movement]

Behavior 

[Scene constraints]

Medical staff 

[Static shot] -- Static-shot option enabled.

 

 

Ethics Statement

The research was approved by the Samsung Medical Center Institutional Review Board (IRB No. SMC 2025-06-135-001) for the collection and use of expert evaluation data. The dataset contains only fully synthetic ICU videos generated using text-to-video AI models. No real patient data, footage, or identifiable information was used in the generation or annotation process. All participating clinical experts provided written informed consent.

 

Notes

Please cite:

@article{seo2025aura,
      title={AURA: Development and Validation of an Augmented Unplanned Removal Alert System using Synthetic ICU Videos},
      author={Seo, Junhyuk and Moon, Hyeyoon and Jung, Kyu-Hwan and Oh, Namkee and Kim, Taerim},
      journal={arXiv preprint arXiv:2511.12241},
      year={2025}
    }

Files

AURA_dataset.zip

Files (72.1 MB)

Name Size Download all
md5:0c70950afa590528abea0123ee51c9ea
72.1 MB Preview Download

Additional details

Software

Repository URL
https://github.com/seo-see/AURA
Programming language
Python
Development Status
Active