care-dataset

care-dataset is the official Python package for accessing and managing the CARE multimodal clinical dataset.

The package provides a high-level interface for loading CARE from Hugging Face, filtering participants using metadata, selecting modalities, creating profile-level train/test and cross-validation splits, and preventing profile leakage in machine learning experiments.

Installation

The package requires Python 3.11 or newer.

Install the package either directly from PyPI:

pip install care-dataset

or by cloning the source repository:

git clone https://gitlab.inesc-id.pt/hlt/public/datasets/multimodal_care.git
cd multimodal_care
pip install .

Quick start

from care_dataset import CAREDataset

care = CAREDataset()

care.info()

Filter the dataset

dataset = care.filter(
    label_id__in=["CONTROL", "PARKINSON"],
)

Select one or more modalities

dataset = dataset.select_modalities(
    "egemaps_func",
)

Create a profile-level cross-validation split

cv_iters, report = dataset.cross_validation(
    k=5,
    keep_duplicate_profiles=False,
    report=True,
)

report.info()

CARE dataset

The CARE dataset, documentation, feature archives, and usage examples are available from the official Hugging Face Dataset repository:

https://huggingface.co/datasets/inesc-id/multimodal_care

License

This package is distributed under the MIT License.

The CARE dataset itself is distributed separately under the CARE Data Use Agreement (DUA).