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

Full Dataset Files (not just embeddings)

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

File	Rows	Description
`train.csv`	223,414	Full CheXpert training set — Path, Sex, Age, AP/PA, 14 diagnosis labels
`valid.csv`	234	Full CheXpert validation set — same schema
`mimic_cxr/mimic-cxr-2.0.0-chexpert.csv`	227,827	Full MIMIC-CXR diagnosis labels — 14 CheXpert-style labels per study
`mimic_cxr/mimic-cxr-2.0.0-metadata.csv`	377,110	Full MIMIC-CXR DICOM metadata — view position, rows, cols, study date
`chexpert_multibackbone/race_mapping.csv`	—	CheXpert patient-to-race mapping (from CHEXPERT DEMO)

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

Embedding Subsets (for benchmark reproduction)

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

Metadata CSV Format

All metadata CSVs share a common schema:

Column	Type	Description
`task_label`	int	Binary task label (0/1)
`sex`	str	Male / Female
`age`	float	Patient age
`age_bin`	str	Age group: <40, 40-60, 60-80, 80+
`race`	str	WHITE, BLACK, ASIAN, OTHER (MIMIC-CXR only)

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

Column	Values	Description
`Atelectasis`	1.0 / 0.0 / NaN	Positive / Negative / Unlabeled
`Cardiomegaly`	1.0 / 0.0 / NaN
`Consolidation`	1.0 / 0.0 / NaN
`Edema`	1.0 / 0.0 / NaN
`Enlarged Cardiomediastinum`	1.0 / 0.0 / NaN
`Fracture`	1.0 / 0.0 / NaN
`Lung Lesion`	1.0 / 0.0 / NaN
`Lung Opacity`	1.0 / 0.0 / NaN
`No Finding`	1.0 / 0.0 / NaN
`Pleural Effusion`	1.0 / 0.0 / NaN
`Pleural Other`	1.0 / 0.0 / NaN
`Pneumonia`	1.0 / 0.0 / NaN
`Pneumothorax`	1.0 / 0.0 / NaN
`Support Devices`	1.0 / 0.0 / NaN

Reproduction Scripts

Dataset	Extraction Script	Prerequisites
CheXpert (MedCLIP)	`scripts/setup_chexpert_data.py`	Existing `data/chest_*.npy`
CheXpert (multi-backbone)	`scripts/extract_chexpert_hf_multibackbone.py --device mps --parallel`	`pip install datasets`, network access
MIMIC-CXR (embeddings)	`scripts/setup_mimic_cxr_data.py`	qml-mimic-cxr-embeddings repo
MIMIC-CXR (diagnosis labels)	`scripts/join_mimic_diagnosis_labels.py`	PhysioNet `mimic-cxr-2.0.0-chexpert.csv`
CelebA	`scripts/extract_celeba_embeddings.py`	`pip install datasets`, network access

Data Provenance

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.

Notes

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).

Downloads last month: 6