This dataset contains metadata derived from MIMIC-IV-Echo, which is a credentialed dataset on PhysioNet. To access this dataset, you must have an active PhysioNet credentialed account with signed Data Use Agreement (DUA) for MIMIC-IV-Echo.

MIMIC-IV-Echo V-JEPA & EchoJEPA Embeddings

Pre-computed video embeddings for all ~525K MIMIC-IV-Echo echocardiography clips using V-JEPA2 / V-JEPA2.1 and EchoJEPA fine-tuned variants.

Access requirement: You must have PhysioNet credentialed access and a signed DUA for MIMIC-IV-Echo before requesting access.

Models

Folder names map directly to source checkpoints. V-JEPA 2 and V-JEPA 2.1 are distinct — see the Description column. EchoJEPA models are all V-JEPA 2.1 fine-tuned on echocardiography.

Status	Folder	Arch	Dim	Description	Checkpoint
✅	vjepa2-vitl	ViT-L 300M	1024	V-JEPA 2 base model, pretrained on natural video — no echo fine-tuning	`vitl.pt`
✅	vjepa2-vitl-scratch-pt-210-c25	ViT-L 300M	1024	V-JEPA 2 architecture, trained from random init on MIMIC-IV-Echo. pt-210 = 210 pretraining epochs; c25 = checkpoint 25	`vitl-scratch-pt-210-c25.pt`
✅	vjepa2.1-vitl-mimic-pt-117	ViT-L 300M	1024	EchoJEPA — ViT-L fine-tuned from V-JEPA 2.1 on MIMIC-IV-Echo. pt-117 = 117 fine-tuning epochs	`vjepa21_vitl_mimic_pt117.pt`
✅	vjepa2.1-vitb-mimic-pt-169-c60	ViT-B 86M	768	EchoJEPA — ViT-B fine-tuned from V-JEPA 2.1 on MIMIC-IV-Echo. pt-169 = 169 epochs; c60 = checkpoint 60	`vjepa2_1_vitb_mimic_pt169_c60.pt`
✅	vjepa2.1-vitl-mimic-pt-100	ViT-L 300M	1024	EchoJEPA — ViT-L fine-tuned from V-JEPA 2.1 on MIMIC-IV-Echo. pt-100 = 100 fine-tuning epochs	`vjepa21_vitl_mimic_pt100.pt`
✅	vjepa2-vitl-vmix22m-pt-220-c55	ViT-L 300M	1024	V-JEPA 2 ViT-L fine-tuned on a 22M-clip video mix. pt-220 = 220 epochs; c55 = checkpoint 55	`vitl-vmix22m-pt220-c55.pt`

Dataset


Videos	525,328 echocardiography clips
Subjects	~4,800 patients
Studies	~7,200 echo studies
Embedding dim	1024 (ViT-L) / 768 (ViT-B) float32
Format	Sharded Parquet (10 shards per model, ~2 GB per model)

Structure

mimic-iv-echo-jepa-embeddings/
├── vjepa2-vitl/                          ✅ V-JEPA2 base, pretrained on natural video
│   ├── train-00000-of-00010.parquet
│   └── ... (10 shards)
├── vjepa2-vitl-scratch-pt-210-c25/       ✅ V-JEPA2 arch, trained from scratch on echo
│   ├── train-00000-of-00010.parquet
│   └── ... (10 shards)
├── vjepa2.1-vitl-mimic-pt-117/           ✅ fine-tuned 117 epochs
│   ├── train-00000-of-00010.parquet
│   └── ... (10 shards)
├── vjepa2.1-vitb-mimic-pt-169-c60/       ✅ fine-tuned 169 epochs, ViT-B (768-dim)
│   ├── train-00000-of-00010.parquet
│   └── ... (10 shards)
├── vjepa2.1-vitl-mimic-pt-100/           ✅ EchoJEPA, fine-tuned 100 epochs
│   ├── train-00000-of-00010.parquet
│   └── ... (10 shards)
└── vjepa2-vitl-vmix22m-pt-220-c55/       ✅ V-JEPA2, video mix fine-tune (22M-clip video mix)
    ├── train-00000-of-00010.parquet
    └── ... (10 shards)

Each shard corresponds to one MIMIC-IV patient folder (p10–p19). Shard index is consistent across all models.

Columns

Column	Type	Source	Description
`subject_id`	int64	echo-record-list.csv	MIMIC patient ID
`study_id`	int64	echo-record-list.csv	Echo study ID
`dicom_id`	str	filename	Original DICOM identifier (e.g. `94106955_0001`)
`file_path`	str	embedding key	Relative path to source MP4
`acquisition_datetime`	str	echo-record-list.csv	Per-video acquisition timestamp
`study_datetime`	str	echo-study-list.csv	Per-study timestamp
`note_id`	str	echo-study-list.csv	Clinical note reference (nullable)
`note_seq`	str	echo-study-list.csv	Note sequence number (nullable)
`note_charttime`	str	echo-study-list.csv	Note chart time (nullable)
`embedding`	list[float32]	encoder	1024-dim (ViT-L) or 768-dim (ViT-B) video embedding

Metadata joined from MIMIC-IV-Echo CSVs so each row is self-contained:

echo-record-list.csv (525K rows) — per-video: subject, study, acquisition time
echo-study-list.csv (7K rows) — per-study: study time, clinical notes

Embedding representation

Each embedding column stores a mean-pooled video-level descriptor: the encoder outputs a sequence of spatiotemporal patch tokens ([num_patches, embed_dim]) and we average across all tokens (features.mean(dim=1)) to produce a single fixed-size vector per video.

This keeps storage tractable (~2 GB per model for 525K videos) and works well for classification, retrieval, and most clinical outcome prediction tasks.

Note on full-token embeddings: Retaining the full token sequence (e.g. ~1500 tokens × 1024 dim) instead of mean-pooling would give better performance on tasks that require fine-grained spatiotemporal structure — in particular, right ventricular (RV) function tasks, where temporal dynamics across frames carry discriminative information that mean-pooling discards. The trade-off is storage: full-token embeddings for 525K videos would require on the order of several terabytes per model, and dimensionality reduction (PCA or token selection) would be needed before training. Extracting and publishing full-token variants is an open contribution opportunity.

Usage

from datasets import load_dataset

# V-JEPA2 base model (pretrained on natural video)
ds = load_dataset("MITCriticalData/mimic-iv-echo-jepa-embeddings",
                  data_dir="vjepa2-vitl")

# EchoJEPA — V-JEPA2.1 fine-tuned on MIMIC-IV-Echo
ds = load_dataset("MITCriticalData/mimic-iv-echo-jepa-embeddings",
                  data_dir="vjepa2.1-vitl-mimic-pt-100")

print(ds["train"][0])  # {'subject_id': 10002221, 'embedding': [...], ...}

Extraction Pipeline

From EchoJEPA-VE:

Step	Command
Extract (SLURM)	`sbatch scripts/extract-embeddings/extract_echo_slurm.sh <model>`
Merge .pt files	`python scripts/extract-embeddings/merge_embeddings.py --model <model>`
Convert to Parquet	`python scripts/extract-embeddings/to_parquet.py --model <model> --subfolder <vjepa2-or-vjepa2.1-variant>`

Config: L40S GPU, batch=256, 8 DataLoader workers, ~465 vids/min per GPU.

Citation

@article{mimic-iv-echo,
  title={MIMIC-IV-Echo: A Large-Scale Echocardiography Dataset},
  note={PhysioNet, https://physionet.org/content/mimic-iv-echo/}
}

@article{bardes2025vjepa2,
  title={Revisiting Feature Prediction for Learning Visual Representations from Video},
  author={Bardes, Adrien and Garrido, Quentin and Ponce, Jean and Chen, Xinlei and Rabbat, Michael and LeCun, Yann and Assran, Mahmoud and Ballas, Nicolas},
  year={2025}
}

Downloads last month: 37

Total file size:

12.5 GB