PLAX EF Prediction Model

This repository hosts pretrained r2plus1d_18 models for estimating left ventricular ejection fraction (EF%) from parasternal long axis (PLAX) echocardiography clips. The models were developed as part of our research on learning EF from scarce data in MIMIC-IV Echo.

Citation

If you use this dataset, please cite the following works:

Primary manuscript
Gao et al., Learning from Scarce Labels: Multi-View Echocardiography for Ejection Fraction Prediction.
Preprint, 2025.

Prior conference paper
Gao, Z., Yurk, D., & Abu-Mostafa, Y. S. (2025). Machine Learning with Scarce Data: Ejection Fraction Prediction Using PLAX View.
In Medical Imaging with Deep Learning (MIDL).
https://openreview.net/forum?id=JEN5FzeFZj

BibTeX

@inproceedings{gao2025machine,
  title     = {Machine Learning with Scarce Data: Ejection Fraction Prediction Using {PLAX} View},
  author    = {Gao, Zhiyuan and Yurk, Dominic and Abu-Mostafa, Yaser S.},
  booktitle = {Medical Imaging with Deep Learning},
  year      = {2025},
  url       = {https://openreview.net/forum?id=JEN5FzeFZj}
}

---

For labels and dataset preparation details, see the companion GitHub repo:
👉 Jeffrey4899/PLAX_EF_Labels_202509

Model Details

Architecture: r2plus1d_18 (video-based CNN)
Input: PLAX echo clips (MP4, H.264, ~64 frames, resized 112×112)
Output: Scalar EF estimate (0–100%)
Performance: ~7% MAE on the held-out test set (see publication for R² and full results).
Dataset: Labels derived from the MIMIC-IV Echo dataset.

⚠️ Two representative model checkpoints are provided here for reproducibility and simplicity:

0_0_r21d.pth
0_2_r21d.pth

In practice, EF prediction performance is obtained by aggregating predictions from both models (50%–50% averaging).

Intended Use & Limitations

Research and education purposes only.
Not for clinical deployment.
Trained solely on PLAX view — does not generalize to A4C or other views.
Assumes reasonable video quality and clip length.

Disclaimer

⚠️ This model is not a medical device and must not be used for clinical diagnosis or treatment.

How to Use

from huggingface_hub import hf_hub_download
import torch, torchvision

ckpt = hf_hub_download("Jeff4899/PLAX_EF", "0_2_r21d.pth")
model = torchvision.models.video.r2plus1d_18(weights=None)
model.fc = torch.nn.Linear(model.fc.in_features, 1)
state = torch.load(ckpt, map_location="cpu")
model.load_state_dict(state, strict=False)
model.eval()

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support