MIMIC-IV-Echo: Pre-Computed Echocardiography Video Embeddings

525,328 echocardiography video clips from the MIMIC-IV-Echo database have been processed into embeddings using Meta's V-JEPA2 self-supervised encoder. The dataset is hosted by MITCriticalData and was last updated on February 25, 2026. Access requires credentialed approval from PhysioNet and a signed Data Use Agreement.

Use Cases

• Develop video classification models for cardiac conditions based on pre-computed embeddings.
• Benchmark self-supervised video encoders on a large-scale medical video dataset.
• Perform similarity search across echocardiography studies using embedding vectors.
• Train downstream predictive models for patient outcomes using video-derived features.

Strengths

• Contains 525,328 echocardiography clips, indicating a substantial scale.
• Embeddings are generated using Meta's state-of-the-art V-JEPA2 encoder.
• Includes metadata linking to clinical notes and patient identifiers for potential multimodal analysis.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Access is restricted and requires a separate credentialed agreement, limiting immediate usability.
• Freshness should be verified as the last update date is in the future (2026).

Provenance

Source

MIMIC-IV-Echo database via PhysioNet

Collection Method

Pre-computed embeddings extracted from video clips using V-JEPA2.

Freshness

Last updated 2026-02-25 10:07:08