Smartphone-Based Video Dataset for Acute vertigo Diagnosis in Emergency Settings: The SMART-VERTIGO database.

This dataset contains cropped and pseudo de-identified videos. All personal metadata, or other identifiable information have been removed. While original consent covered general face video recording, the cropped eye images cannot reasonably be linked to any individual and therefore do not constitute personal data under GDPR. Access is provided via Zenodo under restricted permissions and a signed user agreement, ensuring use exclusively for research purposes and the public interest, in accordance with GDPR requirements for scientific research (Articles 4, 5, 6, 9, 32, 89; Recitals 26, 33, 50, 156, 159). Users must not attempt re-identification, redistribute the data, or use it commercially.

Acute Vertigo Eye-Movement Video Database – Summary

This open-source video database captures eye movements from 100 patients presenting with acute vertigo (AV) in the emergency department (ED) of a UK tertiary hospital. It was designed to support machine learning (ML) model development for automated assessment of nystagmus and subtle ocular motor disorders, as well as for educational and research purposes.

Dataset Content

• Total Videos: 1,318 short Eye-movement videos (5–10 seconds each), in MP4 format.

• Frames: Over 400,000 RGB images

• Annotations:

  • Bounding boxes on selected videos (JSON format): Representative bounding box annotations are included as examples.

  • Binary semantic segmentation masks for 54 frames (iris and sclera, PNG format): A subset of semantic segmentation masks is included as an example resource.

• Ground Truth (GT):

  • Expert-verified clinical labels from three specialists (two senior) 

  • Includes test results, type and direction of nystagmus, and clinical outcomes after ≥2 months follow-up

Participant Demographics

• Total Participants: 100 (57 females, 43 males; aged 18–97)

• Age Distribution:

  • 18–30: 7%

  • 31–50: 31%

  • 51–70: 43%

  • 71+: 19%

• Ethnicity Distribution:

  • White: 51% (≈30% White British)

  • Asian: 18%

  • Black: 17%

  • Mixed/Other: 14%

• Gender & Age Coverage: Broadly balanced, though younger patients slightly underrepresented.

Technical Details

• Videos acquired in real-world Adult ED settings using handheld smartphone-based recording

• Pseudo de-identified and cropped eye images to ensure GDPR compliance

• Annotations and segmentations were performed locally using VIA-video annotator (VGG, Oxford) and MATLAB™ Image Segmenter, ensuring data confidentiality

• Binary semantic segmentation masks provide pixel-level ground truth for ML calibration

Database Structure

1. Cropped eye-movement videos (MP4)

2. Binary segmentation masks (PNG) as examples

3. Video annotations and metadata (JSON) as examples

4. Ground truth spreadsheet (Excel)

5. Documentation 

6. GDPR-compliant user agreement

Diversity & Fairness

The dataset reflects real-world biological and situational variability, including differences in eye size, iris color, scleral characteristics, and lighting conditions. By encompassing demographic diversity and environmental variability, it supports the development of robust, generalizable, and equitable ML models.

Access & Compliance

• Pseudo-anonymized

• Available on Zenodo under restricted access with signed user agreements

• Compliant with GDPR for scientific research, including principles of data minimization, restricted access, and public-interest use