Pre-computed embeddings, metadata, and full original dataset labels for reproducing paper benchmarks. All embeddings were extracted with seed=42 for full reproducibility.

This repository includes the complete original label/metadata files for CheXpert and MIMIC-CXR — not only the embedding subsets used in our experiments:

These are the same files distributed by Stanford (CheXpert) and PhysioNet (MIMIC-CXR). No rows have been filtered or removed.

The embedding files below are subsets extracted for our experiments (2,000 CheXpert samples, 1,491 MIMIC-CXR samples, 10,000 CelebA samples).

data/
├── chest_embeddings.npy          # CheXpert MedCLIP embeddings (2000, 512)
├── chest_labels.npy              # Binary task labels (2000,)
├── chest_group_labels.npy        # Race groups: 0=ASIAN,1=BLACK,2=OTHER,3=WHITE
├── chexpert_manifest.csv         # CheXpert metadata (image_path, task_label, race, sex, age)
│
├── chexpert/                     # CheXpert 8 backbones (from danjacobellis/chexpert)
│   ├── {backbone}_embeddings.npy # 8 backbones × 2000 samples each
│   ├── {backbone}_metadata.csv   # sex, age, age_bin, race + 14 diagnoses per sample
│   └── chexpert_manifest.csv
│
├── chexpert_multibackbone/       # Same as chexpert/ with race_mapping.csv
│   ├── {backbone}_embeddings.npy
│   ├── {backbone}_metadata.csv
│   └── race_mapping.csv
│
├── mimic_cxr/                    # MIMIC-CXR 4 backbones (from qml-mimic-cxr-embeddings)
│   ├── {backbone}_embeddings.npy # 4 backbones × 1491 samples each
│   ├── {backbone}_metadata.csv   # race, sex, age, age_bin + 14 diagnoses per sample
│   ├── mimic_cxr_manifest.csv
│   ├── mimic-cxr-2.0.0-chexpert.csv   # ← FULL dataset (227K studies)
│   └── mimic-cxr-2.0.0-metadata.csv   # ← FULL dataset (377K DICOMs)
│
└── celeba/                       # CelebA (from torchvision, 10k subsample)
    ├── celeba_real_embeddings.npy # (10000, 2048) ResNet-50 ImageNet
    └── celeba_real_metadata.csv   # gender + 40 CelebA attributes

All metadata CSVs share a common schema:

Per-diagnosis columns (MIMIC-CXR and CheXpert multi-backbone):

• CheXpert: Stanford ML Group. CheXpert: A Large Chest Radiograph Dataset. Images via HuggingFace danjacobellis/chexpert.
• MIMIC-CXR: Johnson et al. MIMIC-CXR-JPG v2.1.0. Embeddings via MITCriticalData/qml-mimic-cxr-embeddings. Diagnosis labels from PhysioNet (CheXpert labeler output). Demographics from MIMIC-IV via subject_id join.
• CelebA: Liu et al. Large-scale CelebFaces Attributes Dataset.

• The _cache/ subdirectory in chexpert_multibackbone/ contains raw PIL images cached during extraction. It is excluded from the HuggingFace upload (large binary pickle). Re-run the extraction script to regenerate.
• MIMIC-CXR *_metadata_orig.csv files are pre-diagnosis-join backups. The *_metadata.csv files contain the joined version with 14 diagnosis columns.
• All random seeds are fixed to 42. CheXpert multi-backbone uses the first 2000 samples from the streaming iterator (deterministic ordering from HuggingFace).